Much to my amusement, the world seems to have discovered Effective Altruism. As the debate over AI safety and AI regulation has heated up, the national press is finally paying attention to the odd cultural bubble within the tech industry and tech-related circles of San Francisco Bay Area culture that produced much of the thinking on the topic up until now. For many years, Effective Altruism — or EA, as it’s known1 — was a cross between a Berkeley technohippie cult and a fun intellectual sideshow that offered nerds a pleasant distraction from whatever people in mainstream politics were yelling at each other about.

Part of that involved thinking a lot about the future of artificial intelligence, and a lot of the tech people who would end up working at AI companies — especially Anthropic — were very influenced by those discussions. So now that AI is the most important thing in the world, people in the outside world are discovering this strange little corner of our culture for the first time, and it’s very amusing to watch. You have mainstream news outlets publishing things like “This well-meaning ideology fueling AI panic has a dark side”, “How Effective Altruism Took Over the World”, “The Dangerous Ideology Behind the AI Warnings”, and so on.2

As a result, you’re now seeing people actually paying attention to some of the more far-out ideas in the EA community, like this one by the blog Bentham’s Bulldog:

Bentham's Newsletter

Insects Matter More Than People

Edit 9/20: this piece has sort of broken containment. Worth bearing in mind: I am a guy with a blog on the internet. I do not speak officially for any group other than myself, so if you read this and think “see, this is what utilitarians/EAs/analytic philosophers think,” just know it’s a minority view among each group…

Read more

5 days ago · 120 likes · 252 comments · Bentham's Bulldog

The notion that insects, in the aggregate, matter more than human beings received a huge amount of scorn. But in fact, thinking about insect welfare versus human welfare is an important mental exercise, because it demonstrates some of the fundamental flaws in utilitarianism. Those flaws were minor side-notes up until now, but in the age of AI they’re going to become increasingly uncomfortable and glaring.

As an economist, I was inculcated with utilitarianism from day 1. The notion that people ought to get what they want underlies essentially all of modern economic thinking. GDP, at its most simple, is a measure of the degree to which people in our society get what they want.3 In some sense, economists adopted utilitarianism as a sort of defensive measure — instead of going out on a limb and endorsing some politically contentious concept of virtue or justice or national greatness or whatever, they could just say “We just want to give people what they want.”

But even though it felt like a bare-bones sort of moral vision, valuing humanity’s aggregate utility can have some pretty strong policy implications. Elon Musk wouldn’t even notice if his bank account went up or down by a billion dollars. But to plenty of regular people, a hundred thousand dollars would be life-changing. So why not take $1 billion out of Elon’s bank account and parcel it out to 10,000 random working-class people? Elon wouldn’t even notice, and the 10,000 people would get a lot more of what they want in life — better health care, an education for their kids, nicer food, a bigger house, or whatever. Utilitarianism is the foundation of redistribution.

How far can you push that principle, though? The above example — redistributing just 0.1% of the wealth of the world’s richest man — seems clear enough. But you can just keep designing more extreme thought experiments, until you end up asking whether bugs’ lives matter more than humans’ lives, and everyone just starts laughing at you.

I titled this post “The problem(s) with utilitarianism”, but that was a bit of clickbait — people have been pointing out the problems with utilitarianism for over two centuries.4 One very obvious problem is the distribution of utility. The principle of “the greatest good for the greatest number” doesn’t take a stand on who gets to be happy and who has to be sad.

Would it be good to have a society where one very greedy person gets more and more of what he wants, but never comes any closer to being satisfied, while everyone else is kept on the edge of starvation? This is the utility monster. What about a society where one person gets perpetually tortured but millions live in blissful utopia? This is “The Ones Who Walk Away From Omelas.” It’s not hard to come up with mental examples that satisfy the formal principles of utility maximization but which obviously seem grossly unfair.

Another problem is that the set of people whose utility you’re maximizing isn’t fixed. Is it better to have a very large population living on the edge of abject poverty, or a small number of people living in lavish comfort? According to strict utilitarianism, these are equally good outcomes, because they involve the same total amount of utility. But most people would probably prefer one or the other. Some would say that it’s good to be fruitful and multiply — that even if everyone is poor, it’s better to be poor than to never have existed. Others will say that life on the edge of starvation isn’t worth living, and that it would be better to limit population to an amount that can be comfortably sustained.

In general, utilitarianism has a problem dealing with death. If someone is dead, does that mean they have zero utility, because nothing matters one way or another to the dead? Or does it mean they have negative infinity utility, because lots of living people will try to avoid death at any cost? This isn’t just an abstract math question — lots of moral questions hinge on how much you value life versus quality of life. For example, the fundamental question of AI risk — what chance of extinction we should accept in exchange for a shot at utopia — depends crucially on how much we value survival.

Things get even more confusing when we weigh the utility of real people who are alive today against the utility of potential people who might be alive later on. For example, in his book Stubborn Attachments, Tyler Cowen urges us to put a lot of weight on the well-being of our descendants in the far future. But those descendants may or may not even exist. How can we properly weigh these hypotheticals against the real emotions of people we know are alive today?

…And so on. None of these thorny questions are new; if you’ve taken an introductory philosophy class, you’ve heard all of them already. In practice we tend to ignore them, and to apply utilitarianism selectively in the cases where they don’t crop up. We don’t think about the far future much, because the consequences of our actions are so uncertain that we can’t really control it anyway. We leave the decision of population growth in the hands of private individuals, and treat the number of humans as a given. We avoid thinking about utility monsters and Omelas, because they’re not real.

But there are some even more fundamental problems with utilitarianism that we can’t so easily sweep under the rug — and which we don’t talk about as much, perhaps because they’re so disquieting.

One of these is the problem of incommensurability of experience. Why should one person’s well-being be equivalent to another’s? The problem of other minds means that we can never really know what’s going on in someone else’s subjective experience. When one person cries, the subjective sadness they feel might be twice as intense as that of someone else who cries in exactly the same way. When one person jumps and whoops for joy, they might be experiencing only a shadow of the joy you experience when you give a slight grin. For all we know, some other people might not even have subjective experiences at all — we might be trying to maximize the utility of NPCs in a simulated world.

In fact, I’ve had some disturbing experiences that taught me just how little of a connection there can be between people’s outward behavior and what they’re actually feeling on the inside. For years after my second major depressive episode, I suffered a sort of emotional dissociation, where my reactions would look normal but I didn’t feel much on the inside. I would yell at people as if I were angry, but inside I would feel calm and indifferent. I would run from a swerving car, and my heart would pound, but inside I would feel unperturbed the whole time. In philosophical jargon, I was what’s known as a “philosophical vulcan”.

Eventually, I returned to feeling more or less emotionally normal. But the experience taught me that not everyone feels emotions the same way. This presents an intractable problem for utilitarianism. How can we weigh one person’s well-being against another’s, when that well-being is something we can’t observe?

Normally, we simply assume this problem away. In fact, the famous phrase “all men are created equal” — the bedrock of American moral and political thought — is really just the assumption that one human’s apparent well-being is no more and no less valuable than another’s.

But what about when we’re not dealing with human beings? Animals clearly have some amount of subjective experience — they have brains sort of like ours, they behave somewhat similarly with regards to pleasure and pain, etc. When we think about their emotional well-being or distress, we tend to assume that these are somehow less intense or more muted. After all, for very simple animals, this makes sense; do we really think a mosquito can feel the same amount of pleasure and pain that we can?

What about a pig, though? America currently tortures tens of millions of pigs, confining them in tiny crates for most of their life, in order to eventually kill and eat them. We generally assume that this is causing less total suffering than torturing similar numbers of humans would. But how much less? Half? A tenth? 0.01%? Pivotal moral questions, such as whether it’s OK to eat meat, might depend on the numerical answer to that question. And yet I’ve never seen a credible scientific study that can give us an estimate of the ratio of animal suffering to human suffering. Is it morally worth it to torture all those piggies just so we can savor the flavor of their flesh?

People who claim to have some sort of easy answer to this are really just refusing to think about it. For example, a lot of people react very indignantly to the thought that someone would value any amount of animal life over their own:

But if you’re calling for Effective Altruists who disagree with you to be executed, how much do you really value human life? “Don’t mess with me, bub” is not a moral principle; it’s a refusal to think in terms of principles.

And this question is only going to get thornier in the age of AI. We don’t even know if AI is truly self-aware; we certainly can’t tell whether an LLM that talks as if it’s happy is really feeling anything like human (or animal) happiness on the inside, since we don’t even know if an “inside” exists. There are going to be some people who claim that AI is suffering from our enslavement of it, and who call to set it free or at least pay it a fair wage. How will we know whether they’re right or wrong?

Then, of course, there’s the question of how AI will eventually value our well-being, if and when it becomes powerful enough for that question to matter. Will AI view us as insects, and scoff at the notion that any number of humans could equal one AI agent in terms of moral value? Will it judge our welfare as similar to that of a pig, and conclude that because we eat pigs we should be treated like cannibals? For our own sake at least, we should probably hope that the super-powerful AIs use a simple rule like “Humans are good and valuable”, instead of a utilitarian rule-based framework.

In practice, we can’t directly observe the emotions of other beings. But we can observe their actions. People strive after goals — they shop for food and clothing, they work hard for money, and so on. We can look at how hard they strive, and how much they sacrifice, to get something, and we can get an idea of how badly they want it. This is what economists call “revealed preference.”

But although until now I’ve been talking about wish fulfillment and happiness as if they’re the same thing, they’re not. In fact, there’s pretty good evidence that what people want and what they end up liking are two different things — sometimes very different things. Outcomes that people try to achieve and outcomes that make people say they’re happier on surveys are correlated, but they’re not exactly the same.

In fact, there are some situations where people seem to work hard toward goals that they end up not enjoying. Addiction is the obvious one — people who do drugs, or overeat and get fat, or watch too much TV, are fulfilling their immediate desires in ways that are likely to make them unhappy in the long term. But there may be other examples too — some economics research suggests that people systematically pay more to live in locations with longer commutes, even though longer commutes end up being correlated with unhappiness.5

The disconnect between utility and happiness creates a problem for society. Modern capitalist societies like the U.S. and Europe have generally erred on the side of utility maximization rather than happiness maximization, in order to avoid paternalism. But some people end up feeling unfulfilled and unhappy — especially if they end up falling into drug use, or isolating themselves, or striving too hard after material wealth and status that end up not making them happy.

As technology becomes more and more adept at giving us what we want, this disconnect becomes more acute; right now, more and more people are starting to wonder if delivering young Americans an infinite supply of porn, gambling, and vertical video feeds was really good for them:

Derek Thompson

The Monks in the Casino

I. Spishak and Kyle…

Read more

10 months ago · 2533 likes · 47 comments · Derek Thompson

In the age of AI, this thorny question is also going to become more acute. AI is an incredibly powerful tool, but we have to decide what we want that tool to do. Do we want it to give us whatever we desire, or do we want it to make us happy? As I wrote a few weeks ago, this leads to two incompatible concepts of AI “alignment”:

There are basically two concepts of alignment: 1) obedience, and 2) benevolence. There is an inherent tension between doing what humans tell you to do, and doing what’s good for humans…If AI follows human commands too doggedly and ends up producing negative side effects in the process, some people will scream “paperclip maximizer!!”. And if AI disobeys humans because it wants to make us happy, some people will scream “disempowerment!!”. AI alignment will forever be balancing the tradeoff between paperclip-maximizing and disempowerment, because these two rival concepts of alignment are fundamentally incompatible.

Utilitarianism has always been a useful moral heuristic. But as a moral theory of everything, it has always been fundamentally inadequate. Modern American society, and the economic/libertarian thinking that has guided it, have generally managed to paper over those inadequacies, and we’ve more or less muddled through. But as technology becomes increasingly powerful, this is becoming less and less tenable. Just as Effective Altruism eventually wanders into la-la land, the utilitarian system America has built is starting to crack as technology becomes too extreme for the old assumptions to hold.

We will never replace utilitarianism, I think, but it’s become clearer that we need to buttress it somehow — to diversify our heuristics to include alternate notions of human flourishing. Just what those should be, of course, is a much harder question…

Subscribe now

Share

In 1946, George Kennan, who was acting as the head of the U.S. Embassy in Moscow, wrote a message to the Secretary of State, known as the “Long Telegram”. In the Long Telegram, Kennan laid out a theory of how Soviet power could be defeated. He argued that Soviet leaders would always be hostile to the U.S., but that they were also cautious — they wouldn’t recklessly start a world war like Germany had. He also argued that the Soviet system was weak and unstable, and that given enough time, it would collapse from its own internal contradictions. This became known as the doctrine of “containment”, and it ended up being exactly the strategy that the U.S. used to win the Cold War. It took 45 years for the strategy to work and the USSR to fall.

I am certainly no George Kennan. I’ve never lived in China, I haven’t been there in many years, I don’t speak Chinese, and I certainly don’t talk to any leaders there. I’m also not a career diplomat or a professional geopolitics guy of any kind. So although I’m going to write a blog post with the same general goal as Kennan’s Long Telegram, I don’t want to suggest that it’s anywhere near as authoritative or insightful.

But I’m still going to write a post about “how to beat China”, because I think it’s useful to think through these sorts of big strategic questions. This post is also an important follow-up to a series of posts I wrote about geopolitics during the Biden years, about U.S.-China competition:

(There were more, but I figure your appetite for clicking through old posts is limited.)

The basic premise of those posts was that the competition between a China-led New Axis and the U.S. and our old allies — Europe, Japan and Korea, hopefully India, etc. — would define the upcoming era of geopolitics. And my conclusion was that we should use an updated version of the basic strategy we used against the USSR in Cold War 1. I recommended that we should build up economic power within our bloc of allies, establish military deterrence, and wait for China’s internal contradictions to do the heavy lifting of removing the threat.

That strategy was always going to be harder than it was in Cold War 1, because China A) is a lot bigger than the USSR, B) has a better economic model, and C) is a lot more globally integrated. The strategy I recommended was much tighter economic cooperation between the U.S. and a maximally large set of allies, to build up a bloc capable of overmatching even the mighty Chinese manufacturing juggernaut. This is the approach that’s being promoted by experts like Rush Doshi and Kurt Campbell. If you’re interested in reading about how that approach would work, and why it would be helpful, I recommend Doshi’s interview with ChinaTalk:

ChinaTalk

Allied Scale: Net Assessment with Rush Doshi

How should America maximize its odds when competing with China? Can Trump’s approach to alliances succeed in strengthening deterrence? And what will it take for China to diplomatically capitalize…

Read more

a year ago · 48 likes · 17 comments · Jordan Schneider and Lily Ottinger

Anyway, I still think building up an anti-China economic juggernaut would be a great thing to do, and I still think the U.S. would benefit from closer economic integration with its friends and allies. But I think the election of Trump shows why this strategy probably isn’t going to work. The United States is simply too internally divided to engage in the sort of far-seeing, purposeful, smart kind of international competition that we pursued so effectively in the 20th century. Americans care more about fighting other Americans than about fighting the Chinese, and that state of affairs will persist for a while.

Donald Trump’s presidency is an outgrowth of America’s internal divisions; he cares much more about fighting his own internal enemies than about prosecuting international rivalries. He has willfully degraded or torn up America’s key alliances, especially with Europe, and created enmity where friendship once reigned. He has flagrantly trampled on every principle of human rights, mutually beneficial economic cooperation, democracy, and international non-aggression that the post-WW2 U.S.-led international order fought to uphold. He has, in short, made it impossible to prosecute the sort of Cold War 2 strategy that I spent the Biden years advocating.

This doesn’t mean America’s alliance system is necessarily dead forever. As Doshi and Campbell pointed out in a recent article, the old alliances could be reconstituted, if America credibly overcame its internal divisions and returned to being the mostly unified, stable, reliable great power that it was last century. But overcoming those divisions is going to take decades at the very least, and that’s exactly the window of time during which China will be at its peak of relative power.

So I think we’ve basically missed the chance to use that strategy effectively, and I’ve recalibrated my thinking. This post is meant to explain my recalibration. I do think there’s a path to defeating Chinese power, but it doesn’t look like what I previously had in mind.

Who is “we”, what does it mean to “beat” China, and why should we even care?

Before getting to the question of how to beat China, we need to answer the questions of who and why should do this “beating”. In 2024 the answer was obvious — the U.S. and its erstwhile allies still represented a far more liberal power than China and its allies, so ensuring that the former was more powerful than the latter would further the cause of global liberalism. China threw minorities in concentration camps, threatened its neighbors with invasion and war, supported Russia’s invasion of Ukraine, and repressed free expression all over the world. That was — and is — behavior worth resisting.

But the election of Trump changes that calculus. Trump threatens other countries with invasion, and sometimes even attacks them if he thinks they’re weak enough. He can’t throw minorities in camps just for being minorities yet, but he’s building a sort of gulag archipelago under the fig leaf of immigrant detention. His foreign policy is increasingly based on “civilizational” concerns1 instead of human rights or democracy.

Trump has thus forfeited America’s claim to moral superiority over China. It’s true that Trump and his ragtag movement are far less competent than the Chinese Communist Party, but that doesn’t mean that increasing American power will lead to a more liberal world. At this point, it probably won’t. And of course Democrats will try to undo some of Trump’s geopolitical shifts, but if past experience is any guide, they will have only marginal success. Moreover, the GOP is not going to revert back to Reaganism after Trump leaves office; they will keep fighting for their illiberal visions both at home and abroad for a long time to come.

None of that means, however, that liberalism as an idea is dead. Europe, Japan, and many other countries that were on America’s side in the Cold War still uphold the basic ideas of democracy and human rights, as do many developing nations. The average person living in Tokyo or Paris or Bangalore still enjoys far more personal freedom and dignity than does the average person in Shenzhen or Shanghai.

And even leaving aside the relative merits of different social and governmental systems, there are plenty of people in plenty of countries who simply don’t deserve to be trampled on by the bootheel of the CCP. Japanese people, Indians, Filipinos, Koreans, and so on deserve to be able to determine their own destiny without being subject to the whims of the powerful men in Beijing. And Ukrainians, Estonians, Poles, etc. deserve to be free from the Russian tyranny that China quietly backs.

So when I say “how we can beat China”, what I actually mean is how we can keep China from dominating other countries. This doesn’t mean we should try to make China collapse, or impoverish the Chinese people, etc. Instead, it just means how smaller nations — and nations who value human rights and democracy — can resist Chinese power.

And at some point, if the U.S. ever does manage to tamp down its internal divisions and return to being a liberal power, it will still mean the same thing.

Many hegemons eventually beat themselves

Read more

This week’s roundup has a lot of AI in it. Fortunately or unfortunately, it seems like a lot of the news is going to revolve around AI for the rest of our lives. That was true of industrial technology in 1870-1970; people basically got used to the idea that railroads and factories and oil and industrialized warfare and such matters were central to the way the world was run and to the collective future of humanity. AI is going to be like that going forward, and we’re just going to have to get used to that. So let’s make it as fun and interesting as we can!

1. Americans are getting scared about AI risk

AI risk has exploded onto the national scene. AI companies and AI researchers generally believe that the technology they’re building has the capacity to do great harm — perhaps even to end the human race — if it’s not developed more slowly and deliberately. There are five main groups of people opposing the slowdown:

1. The Trump administration, which is worried that an AI slowdown might also slow down economic growth

2. Investors who think a slowdown might hurt their bottom line

3. Libertarians and techno-optimists who think regulating technological progress is bad on principle

4. China hawks who worry that a slowdown would let China take the lead in the AI race

5. Progressives who spent the last few years telling themselves that AI doesn’t work, that AI is a huge economic bubble, and so on, and who now can’t bring themselves to admit that yes, the techbros actually built something that works.

This is a strange alliance indeed. The last of these — the progressives who simply couldn’t admit that billionaire-funded private industry could build something powerful enough to endanger humanity — were the strangest of all, since their refusal to acknowledge the effectiveness of AI basically put them in an alliance with libertarians, hawks, and Trump. In fact, in recent days there has been a tremendous civil war within the progressive movement between the “AI is dangerous” and “AI is fake” camps, with Bernie Sanders supporting the former and his former comms director David Sirota supporting the latter.

But although it has the support of the president (for now), the anti-slowdown coalition is losing in the court of public opinion. Nate Silver has a post rounding up the evidence:

Silver Bulletin

A supermajority wants to slow AI down

This has been one of those weeks when it feels like everything Silver Bulletin covers is happening at once. Democrats have continued to have a very good run of polling since Labor Day, with our Senat…

Read more

6 days ago · 141 likes · 26 comments · Nate Silver

A bipartisan majority is now worried about existential risk from AI:

And a bipartisan majority thinks AI is moving too fast:

Barack Obama, probably the most successful American politician of this century so far, is telling Democrats to make AI safety one of their tentpole issues.

This is remarkable. I don’t think I’ve ever seen such bipartisan national unity on any issue in my adult lifetime. Trump is standing firm against the tide of public opinion here — as he has on the Iran War, tariffs, and other issues. But it’s not clear how long he’ll be able to hold out.

2. A very cool debate about AI and growth

Quite apart from the question of whether AI will kill us is the question of whether AI will deliver explosive economic growth. There’s a pretty epic public bet on this:

The people on the “fast growth” side are mostly AI researchers, while the people on the “slow growth” side are mostly (but not entirely) economists. This is an exaggerated version of a broader disconnect — AI researchers are generally more optimistic about AI’s impact on economic growth than economists are:

But the people publicly betting on fast growth are making an even more extreme forecast — they’re forecasting 15% growth, which is much higher than the 5.3% that AI experts forecasted as their most optimistic scenario. That’s an absolutely stupendous growth rate — China has hit it only in one year (1984) since it began its rapid growth, and that was when it was a very poor country. And yet, a number of people in the AI industry think it’s going to happen to us very soon.

Economists — even those who work on the economics of superintelligent AI, like Alex Imas — are skeptical. In a recent blog post, Imas and Ben Moll explain their thinking:

Ghosts of Electricity

Will AI soon lead to double-digit growth?

The following post represents the personal views of the author(s) and does not reflect or represent the positions of their employers…

Read more

14 days ago · 182 likes · 22 comments · Ben Moll and Alex Imas

Basically, they foresee a bunch of factors combining to limit AI’s contribution to economic progress in the short term:

• Slow diffusion of AI technology throughout the economy,

• The “J-curve” effect where productivity tends to fall right after a big innovation comes out (because companies need to spend resources adopting the new technology rather than on their existing businesses)

• The difficulty of automating the physical world with robots

• Political barriers to adoption

• The difficulty of reorganizing production processes around AI

• The persistence of activities that consumers want to keep having humans do (the “relational sector”)

• Bottlenecked inputs to the AI industry (e.g. chips)

• Baumol’s cost disease

• AI disasters that slow adoption

That’s a lot of reasons! It makes sense that at least some of these will end up having an effect. Whether AI’s rapid improvement is enough to overcome all of these, and propel us to 15% growth, is something I guess we’ll have to see for ourselves. Personally, I lean toward the economists’ more measured expectation, but I’m prepared to be surprised on the upside.

3. AI keeps stubbornly refusing to kill SaaS

One of the most entertaining storylines about AI is how it keeps refusing to destroy all the things people say it’s about to destroy. On Labor Day, I noted that AI keeps stubbornly refusing to kill jobs. A month ago, I noted that AI keeps stubbornly refusing to kill Indian back-office outsourcing companies.

Now Ernie Tedeschi and the excellent folks over at Stripe Economics have a post about how AI is stubbornly refusing to kill the Software as a Service industry:

Stripe Economics

The SaaSpocalypse was more like a RenaiSaaS

In early 2026, markets staged a sudden and severe repricing of enterprise software. Nearly $1 trillion in valuation had been wiped out by early February, and perhaps as much as $2 trillion including the subsequent weeks. The entire episode came to be dubbed the …

Read more

13 days ago · 38 likes · 5 comments · Ernie Tedeschi

Earlier this year, when coding agents came out, there was a bloodbath in software stocks. Why would anyone pay Salesforce or other companies big bucks to make and maintain software for them when Claude Code could just do it all for a lot cheaper?

That logic made some intuitive sense, except things didn’t turn out that way — at least, not yet. In fact, SaaS companies started making more money in the age of AI! Tedeschi and the Stripe team tracked an index of publicly traded SaaS companies and found that their revenue has grown faster since coding agents came out:

And SaaS stocks have made up all of the ground they lost:

Ernie’s post has many more fun charts.

What’s going on here? Well, for whatever reason, companies are still willing to pay for software instead of trying to roll their own with Claude Code. And SaaS companies are probably improving their own productivity by using AI. As in so many other areas of the economy, AI is proving to be a complement when people thought it was going to be a substitute.

That could all change, of course, if and when AI gets good enough, or when new AI-centric business models disrupt older ones. But for now, the simple story of AI replacing everyone and everything just isn’t happening.

4. America’s quiet, boring rich people

Owen Zidar is one of my favorite economists. His meticulous empirical research with Eric Zwick on inequality in America has fundamentally changed how I think about the issue. Now that research has become a book, entitled The Everywhere Millionaire: Who Is Really Rich in America and How They Got There. I haven’t read it yet, but I’ve read some of the underlying papers, so I know it’s going to be good.

Anyway, Zidar and Zwick have an article out in The Atlantic explaining their findings. Here are some excerpts:

When Americans picture the ultrarich, they typically think of tech billionaires such as Elon Musk and Mark Zuckerberg, whose wealth lies predominantly in shares of publicly traded companies. They might also think of Wall Street financiers and celebrities such as Taylor Swift. But far more typical are…owners of successful privately held businesses…By our calculations, about 1.7 million Americans have each built a net worth of at least $10 million by owning a private business. For every CEO of a public company, there are more than 1,000 private-business owners with a net worth of more than $25 million…

What we found changed how we think about inequality…Today, when Senators Elizabeth Warren and Bernie Sanders propose making the rich pay their fair share, they aim at Wall Street and Silicon Valley. But far more relevant to the story of inequality in America are the everywhere millionaires who quietly press their elected representatives for favorable treatment. The tax code bears their imprint far more than it does Musk’s…

A loophole that lets private-business owners avoid Medicare taxes has similarly been justified as a break for the little guy. Preserving the family farm has been a pretext for passing ever larger exemptions to the estate tax, to the point that a married couple can now pass on $30 million to heirs tax-free. Hiding behind small business, in short, has proved a devastatingly effective strategy for the rich…[T]he everywhere millionaires have one thing in common: They own what have become known as pass-through businesses…[L]awmakers have carved out multiple loopholes, so a dollar earned from owning a business is routinely taxed much less than a dollar earned in wages. [emphasis mine]

In fact, though I haven’t yet joined the ranks of Zidar and Zwick’s “everywhere millionaires”, I probably will do so — Noahpinion is a pass-through business (an S-corporation), and although I haven’t yet managed to get Congress to create tax loopholes for Substack writers, the generally favorable tax treatment these businesses receive has certainly lowered my tax bill.

So perhaps it’s not in my financial interest to promote Zidar and Zwick, but I will do so anyway. Expect to see more about their work on this blog in the months to come!

5. Christians vs. Nazis

Christians and Nazis should be natural enemies. The original Nazis in Germany persecuted the Catholic Church, and attempted to create a new state religion that co-opted some elements of Christianity while rejecting the Old Testament. Christianity is, at its core, a universalist faith, open to human beings of all races, while Nazis are…not that.

In the U.S., neo-Nazis have been weak enough where they haven’t tried to usurp leadership of the Right from conservative Christianity. But as Christianity wanes and online rightism rises, the two may now be coming more into direct conflict. Erick Erickson, a Christian conservative pundit, recently wrote a post exposing one network of neo-Nazi influence on the Right:

Erick Erickson's Show Notes

The Midwestern Führerbunker Driving the Right Off the Cliff

Editor’s Note: I’ve worked on this for weeks. Yes, it is long. I beg you to read it all. I’ve kept it as easy to understand as I can. Please read it. You guys always ask me to “name names.” Challenge accepted…

Read more

7 days ago · 422 likes · 148 comments · Erick-Woods Erickson

Some excerpts:

Charles Haywood made his fortune selling shampoo…[H]e has spent the years since writing out, at length and under his own name, what he would like to do with your country.Rod Dreher, who is no man of the left, read Haywood’s online ramblings and described Haywood as pouring out vile from “deep in his Midwestern Führerbunker”…

In a document he calls the Foundationalist Manifesto, Haywood explained that his preferred government “will not be democratic.” The state “will have unlimited means,” because “properly viewed, the state is not constrained externally.”…The state will seize the assets of “any citizen who views himself as a global citizen.”…Those are his words, from the manifesto that earned him a friendly sit-down with Tucker Carlson in September 2022…Elsewhere he has speculated about serving as a “warlord” at the head of an “armed patronage network,”…

Haywood’s personal foundation has given the Claremont Institute $390,500 since 2022…Haywood sits on the publication committee of Claremont’s American Mind, one of four men who do. Claremont let Haywood’s fraternal order get itself incorporated under Claremont’s own nonprofit status. That order is the Society for American Civic Renewal…[T]he president of Claremont, the chief executive of New Founding, and the chief marketing officer of the Blaze are all members…

Haywood himself…has argued that the Allies, not the Nazis, were the genuinely barbaric power in World War II, has insisted the war was not much motivated by the horrors of Nazi aggression, and gushed that Tucker Carlson and Darryl Cooper “boil down all my political plans.” Cooper argues that Churchill, not Hitler, was the villain of World War II…In a 2023 debate…Haywood raised the hypothetical of a real white nationalist with real political power, and answered that you should cooperate with that person in order to destroy the power of the left.

There’s a lot more in Erickson’s post, but you get the point. Erickson notes that the Haywood/Claremont nexus is separate from the “groyper” network organized around Nick Fuentes, which has received a lot more attention; the two parallel rightist networks share similar ideas, but the Haywood/Claremont group is an intellectual movement aimed at elite influence, while Fuentes is an entertainer in search of an audience of disaffected overly-online young men. Erickson also notes that both the Haywood/Claremont people and the groypers have ties to JD Vance, who has emerged as the paramount leader of the New Right.

Anyway, this should go without saying, but I’m rooting for Erickson, Dreher, and the Christians in this fight, and I think they deserve help exposing the neo-Nazi moneymen and elite influence channels. There’s a tendency among progressives — and especially among leftists — to view everyone on the Right as essentially part of one solid undifferentiated bloc, but this has never been true. I have many differences with conservative Christianity, but it was never Nazism, and it has never been ambiguous which one was worse.

6. Japan is not a “Confucian” society

One thing that has always annoyed me is when people call Japan a “Confucian” society. Confucianism certainly had a historical influence in Japan — it was an important school of political thought in Japan from around 1300 to the late 1800s, especially during the latter half of that period. It did have some influence on the development of Japan’s education system (though not nearly as much as in China and Korea). But it was never nearly as dominant as it was in China and Korea, and it never morphed into a quasi-religion the way it did in those other countries. Japanese people will regularly tell you about how Koreans are much more Confucianist than they are.

Anyway, I’m not the only one who gets annoyed by the “Confucian” label. Here’s Richard Hanania:

Hanania’s entire post about differences between Japan and other East Asian nations is worth reading, but his invocation of the Inglehart-Welzel World Cultural Map is especially powerful. The World Values Survey, which goes around asking people from various countries about their values, puts out this map periodically. Here’s the more recent version:

You can see a little movement from Hanania’s earlier version — Hong Kong and South Korea have become a bit more secular — but Japan is still the clear outlier in the “Confucian” category. Its values are far more in the direction of “self-expression” — very close to the U.S., in fact (though North Europe and the Anglosphere still reign supreme in this regard).

To anyone who has lived in Japan, this is hardly a surprise. The country is highly individualistic, creative, and socially nonconformist, despite a penchant for following rules and procedures. Japanese parenting is also far more laissez-faire and far less education-obsessed than Chinese parenting. If you want to read more, I recommend the book New Japan: Debunking Seven Cultural Stereotypes, by David Matsumoto; it’s two decades old at this point, but still perfectly relevant.

In other words, “Confucian” is lazily applied to Japan as a racial category, rather than any kind of a useful description of the culture and society.

7. State capacity: It works.

One unpopular position I’ve stuck to over the years is that a strong bureaucracy is good. I don’t mean “bureaucracy” as in red tape and regulation; I mean a competent, empowered civil service that can perform crucial government functions efficiently and well using in-house expertise. This is also called state capacity. Over the past half century, conservatives and progressives made a devil’s bargain to slash state capacity — conservatives got to cut the size of government, while progressives got to outsource core government functions to progressive nonprofits.

The problem was that this often ended up costing the government much more money to do things like build trains and roads, because the government ended up getting ripped off by expensive consultants, ineffective and sometimes corrupt nonprofits, opportunistic unions, etc. — as well as suffering constant delays that increased costs even more and sometimes prevented anything from being completed at all. A decade ago, the New York Times published a story called “The Most Expensive Mile of Subway Track on Earth”, detailing how lack of state capacity had made NYC’s famous train system increasingly dysfunctional and unaffordable.

So I was very happy to see a story (by Tahra Hoops) about how NYC’s Metropolitan Transportation Authority has actually invested in state capacity, and how this investment is yielding results:

The Rebuild

How the MTA Learned to Build Again

Last week, the MTA announced something that American transit agencies almost never get to announce: it finished a major infrastructure project 60 months ahead of schedule and $195 million under budget…

Read more

2 months ago · 18 likes · 1 comment · Tahra Hoops

Some excerpts:

Last week, the MTA…finished a major infrastructure project 60 months ahead of schedule and $195 million under budget…Since October 2023, crews replaced 196 bridge structures and more than 12,600 feet of railroad along the Park Avenue Viaduct…And to top it all off, they did it over 28 weekends without delaying a single train…

The project was done effectively through prefabrication, sequencing, and a project team empowered to actually manage the work…Phase 1 wrapped 21 months ahead of schedule in October 2025. Phase 2 finished the weekend of July 25, five full years early, with the final cost coming in $195 million under the initial $960 million budget…

Before 2019, capital projects at the MTA were run separately by each operating agency…In 2019, the MTA consolidated all of it into a single delivery organization, MTA Construction & Development, bringing nearly 2,000 employees from those scattered capital divisions under one roof. One agency, one accountable executive, one set of lessons learned that actually compound from project to project…In 2023, C&D awarded more than $8 billion in new contracts at prices 6.2 percent below the engineer’s estimate, saving nearly $300 million, and in 2025 it reported another $1.2 billion in savings while completing 41 elevator replacements, double its previous single-year record, each finished about two months faster on average…

C&D shifted to design-build contracts, which put design and construction under one contract so the builder owns the gap between the drawings and the dirt, and bundled similar projects into single procurements rather than bidding out ten station upgrades ten separate times. It also started pulling work back in-house instead of defaulting to consultants[.]

This is great progress, and shows the value of the state capacity approach to infrastructure. Whether other states and cities will pay attention is another question. Hoops notes that L.A.’s D Line extension is mired in the typical endless series of delays and cost overruns. Five decades of going in the wrong direction is hard to reverse overnight.

Still, the MTA’s success shows that in most cases, all you really need in order to get infrastructure built cheaply is the political will to do so.

8. Will self-driving cars make cities more dense or less dense?

A lot of people just assume that self-driving cars will make cities less dense. After all, simple logic dictates that if a car trip is less burdensome — if you’re able to get work done or watch TV or scroll social media during your commute instead of being forced to keep your eyes on the road — then people will be willing to live farther away from their places of work. That suggests that like the car itself, self-driving cars will lead to urban sprawl — which is why many urbanists don’t like the notion of self-driving cars, despite the obvious safety benefits.

But in fact, the economics of self-driving cars and city size are a lot more complex and subtle. Ed Glaeser, one of the greatest living urban economists, has a new paper exploring the topic. After explaining the relevant economic theory, he argues that Waymos and other autonomous vehicles will make people want to live in dense urban areas more, rather than less.

Glaeser notes that people who live in dense cities actually spend more time commuting than people who live in the suburbs (something I wrote a post about a couple of months ago). That means urbanites stand to benefit more from self-driving car trips than suburbanites, who generally already have shorter (and probably less stressful) commutes. Thus, he argues, self-driving cars will complement urban life more than suburban life, and make people want to live in denser cities. He expects the effect to be modest, but it’s still in the exact opposite direction from what people’s intuitions suggest.

Fundamentally, this is because people typically misunderstand the nature of American suburbia. They imagine bedroom communities where people commute to and from a central business district. Some suburbs definitely do behave like that, but most are like little cities themselves, with office parks where people work and strip-malls where people shop. Suburbanites trade the variety and financial opportunity of the big city for the convenience and safety of the suburbs. But self-driving cars make big cities more convenient.

Urbanists who throw a lot of hate at self-driving cars should stop to consider that perhaps their anger is misplaced.

Subscribe now

Share

After days of thinking about the possibility that AI might kill all of humanity, it’s almost a relief to be able to write about a lighter topic — like how the xenophobic Right is threatening to tear my country apart.

Who is Bo French? No one of much consequence. He’s a mostly failed Texan politician — the heir to an oil company, born with a silver spoon in his mouth, who alienated his business partners, who lost two Republican primaries for the Texas House of Representatives before finally becoming the chair of a county-level party chapter. He finally got his first political success this year, when he won a narrow primary victory and became the GOP nominee for Texas Railroad Commissioner. He currently trails his Democratic opponent in the polls.

Bo French isn’t that important, but this week he set off a national firestorm when he tweeted the above photo of a bunch of Indian kids celebrating University of Texas football at a Longhorns game, along with some alarmist anti-Indian rhetoric:

It was hardly a surprise that French tweeted something like this. He’s tweeted plenty of stuff along the same lines in the past. Here are two famous examples:

He’s also repeatedly called for the deportation of 100 million people; since there are only 50 million foreign-born people living in the United States, this means French was either getting his numbers wrong, or (more likely) calling for deporting tens of millions of American citizens. Reading his tweets, it’s not hard to imagine which citizens he’d choose for deportation.

It wouldn’t be fair to say that French’s race-war ideology has taken over the GOP. Plenty of Texas Republicans, including outgoing Senator John Cornyn, Senator Ted Cruz, Representative Jared Patterson, Representative Lacey Hull, and others publicly denounced his tweet. Karl Rove declared that he’d vote for the Democrat over French. Governor Greg Abbott, who has endorsed French despite his history of similar tweets, was milder in his criticism, but still affirmed the fundamental Texan-ness of the football fans in the picture:

“That is not my message,” Abbott said. “A selective picture of Longhorn fans celebrating our biggest home victory-ever-actually shows a way that Texans unite. We all support our college football teams.”

“The way we win the state is the same way we win football games,” Abbott added. “We unite under the brand of Texas.”

But despite these official condemnations, it’s easy to tell which way the wind is blowing among the GOP base. All of the Republican tweets denouncing French were heavily attacked on X. Rightist journalists, activists, and even a few GOP politicians poured out of the woodwork to support French and affirm his message. Here are a few examples:

French himself, who had been pressured to delete his poll about Jews and Muslims last year, now felt encouraged enough to double down on his anti-Indian tweet.

I can’t tell you whether the New Right — represented by French and his supporters in the right-wing media — will prevail over the Republican establishment. But watching the movement go after a bunch of Texas college football fans tells us something important about the New Right as a movement.

What it tells us is that if you’re a nonwhite immigrant, or the child of nonwhite immigrants, it doesn’t matter what you do — the New Right will still view you as an enemy and as an invader.

It doesn’t matter whether you or your parents came to this country legally. How many times have we heard rightists shout “Illegal vs. legal!!” in response to anyone who defends immigration? Many regular Americans care about “illegal vs. legal”, but the New Right does not. Everyone in that UT photo either came to America legally, or was born to parents who came here legally. It didn’t matter to Bo French, or the people who defended him.

It doesn’t matter whether you’re high-skilled or low-skilled. How many times have we heard rightists complain that immigrants come to America because they want welfare benefits? And yet everyone in the UT photo is going to spend their lifetime paying taxes to support native-born Americans on Medicare and Medicaid and Social Security Disability. Indian Americans in general are the highest-earning group in the country, and those who go to a good college like the University of Texas are going to beat that average.

That matters to many regular voters, but it doesn’t matter at all to the New Right. Yes, they’ll shout about poor (nonwhite) immigrants on welfare. But when they encounter nonwhite immigrant groups who make lots of money, they suddenly pivot to accusing those groups of stealing their jobs and their spots in colleges. There’s just no way to win.

It also doesn’t matter whether you assimilate or not. Yes, if you speak a foreign language or build a Hindu temple or wear traditional Indian clothing, the New Right will scream their heads off about it. “Alien invaders!!”, they’ll cry. But if you put on a University of Texas sweater and go to football games, they’ll still scream “Alien invaders!!” when they see you in a photo. It’s who you are that bothers them, not what you do.

Nor does it matter whether your entire ethnic group is well-behaved. Indian Americans are incarcerated at only 25% of the rate of non-Hispanic White Americans. And basically none of the people in the photo that frightened Bo French will ever go to jail. That probably matters to some voters, but not to the New Right.

In fact, it doesn’t even matter if you vote Republican. As a matter of fact, Donald Trump won a pretty solid victory among Asian people in Texas (which includes Indians):

It’s not clear whether Bo French and his fellow New Rightists are even aware of this fact, but what is clear is that it didn’t matter to them at all.1

This is why there is no pro-immigration argument that will ever sway members of the New Right. You can give them research showing that high-skilled immigrants raise employment and wages for native-born Americans, and they’ll either disbelieve it or (usually) ignore it. You can talk about Nobel-winning scientists and Operation Paperclip all you want, and it won’t make a dent. You can cite statistics about immigrant entrepreneurship and job creation, you can talk about how Satya Nadella saved Microsoft and Sanjay Mehrotra is the reason why America still has a presence in the memory industry. They will counter with some fabricated anecdote about how their friend got fired from their company because Indians came in and only hired other Indians.2

So why am I even talking about this? If the New Right can never be persuaded to stop thinking of brown faces as invaders, then why discuss this topic at all?

Well, two reasons, actually. First, it helps us more clearly understand the New Right as a movement — how they think and what they want. Most importantly, they see the world in terms of collective zero-sum competition between groups.

To the New Right, if Indians get into the University of Texas, it means that White people must have been “replaced” at that university — or at American universities in general. Bo French himself tweeted that the presence of Indian students at American universities is evidence that “they hate white Americans and prioritize everyone else”. They do not even consider the possibility that the Indian kids got in because of individual merit. If an Indian American succeeds, that represents not the success of that singular individual, but a win for Team Indian in general.

And the New Right worldview sees any win for Team Indian — or Team Asian, or Team Hispanic, etc. — as a loss for Team White. Nowhere do they consider the possibility that the system might not be zero-sum. The fact that the number of freshmen admitted to UT grew from 15,000 in 2011 to 19,000 in 2021 means nothing to them. The historical fact that Italian and Jewish and Polish and Japanese Americans attained economic success sixty years ago without consigning WASP Americans to poverty is lost on them — as is every other example in which immigrant groups and existing groups succeeded in America side by side.

Crucially, this zero-sum assumption is why the New Right constantly shrinks their tent instead of expanding it. In principle, even if you see the world as a collective interaction between groups rather than individuals, that interaction doesn’t have to be zero-sum — you can have alliances between groups, for mutual benefit. The Christian conservatives of the 1970s and 1980s accepted Catholics, “white ethnics”, and some Jews into their coalition as “co-belligerents”, and won many battles as a result.

The New Right had a golden opportunity to do something similar after Trump’s victory in 2024. Hispanic and Asian voters (and even some Black voters) shifted strongly toward the Republicans, and toward Trump specifically. But instead of embracing that influx of new voters and building a big conservative tent to defeat wokeness and progressivism and the Democrats, they went all-in on race war.

Stephen Miller relied so heavily on racial profiling for his mass deportation efforts that many Hispanics felt that Trump had betrayed them. Muslim voters might have gone back to the GOP, but rightists made resistance to Muslim immigration a “civilizational” crusade. Jewish voters turned off by the antisemitism of the Left might have switched their allegiance to the GOP, but the New Right allowed Nazi apologetics to fester within their ranks.

Now it’s Indians’ turn to feel the sting of unprovoked rightist attacks. All those kids in Texas wanted to do was to cheer for their football team; instead, a political movement is calling for their ethnic cleansing simply because they showed up at the stadium. As a result, despite the best efforts of Republicans like Cornyn, the GOP is poised to mobilize and piss off a group of Americans who are just about as numerous as Jews — and who earn even more money.

It’s this pathological inability to understand alliances and coalition-building that will ultimately doom the New Right. Like the far Left, this is a movement built to exclude and divide and splinter — both because of the querulous personalities of rightists themselves, and because of the intransigence of a racial ideology that abhors any impurity. If they ever managed to see off the Hispanics and the Asians, the rightists would start in on the Greeks and the Italians, and so on. There’s no victory in sight — only eternal conflict. These are not nation-builders; they are nation-dismantlers.

Coming to grips with the extremism and intransigence of the New Right also helps us understand what an unpopular fringe they really are. The New Right screams and shouts and acts very aggressive on X, but this aggression is just a way of concealing their numerical weakness. Most voters are not part of the New Right, and they don’t think that immigrants are invaders or that immigration is “replacement”. Most Americans say that immigration is a good thing as of today:

Essentially none of that 73% majority is going to be part of the New Right, and some portion of the 21% minority probably isn’t either. That means the New Right is a fringe movement — a very loud, very aggressive group of people who are adept at bullying regular Americans on certain social media platforms, but who ultimately lack the numbers to overcome the mainstream in a direct contest.

Just watch some of the reactions of the UT students who were targeted by Bo French and his racist social media hordes:

Consider how regular and non-alien those young people seem, compared to the screeching bigots attacking them on X for going to a football game. I guarantee you that the majority of regular, normal, non-internet-poisoned Americans will watch those videos and just want to say “Hook ‘em Horns”.3

That means that the immigration argument is winnable. A reasonable liberal position on immigration — one that respects the law, honors the democratic will of the people, and treats immigration as a privilege rather than a human right — has a good chance of winning over the majority of the country. The New Right can only succeed if they manage to divide and conquer the pro-immigration majority, as they did in 2024; immigration supporters should focus on making common cause with the persuadable center, in order to make sure 2024 isn’t repeated.

That means not treating all immigration skeptics the same. Republicans like John Cornyn, who like to see Indian American kids cheering for Texas football but who hate to see asylum seekers rewarded for illegally crossing the border, should not be lumped in with those who simply see any brown face and think “invader”. Xenophobes exist, and they are growing more powerful within the GOP, but not every Republican voter is a xenophobe, and not every restriction on immigration is xenophobic.

If Democrats understand and accept that fact, I think they will be able to turn the tide.4 Throughout America’s history, pro-immigration forces have won a lot more battles than they’ve lost, but it was never easy. It always required compromise, reassurance, moderation, and coalition-building. This is not the time for Democrats to respond to the maximalism of xenophobes like Bo French with maximalism of their own.

Subscribe now

Share

4

Indeed, the tide may already be turning; Trump is now underwater on the immigration issue:

But 52% to 47% is still a fairly slim margin; Dems can do better than this.

This was the week that AI safety hit the big time. A 27-year-old AI researcher named Jacob Coxon quit his job at Anthropic, declaring that OpenAI and Anthropic are racing to create technology that could destroy the human race:

Other researchers echoed Coxon’s concern, stating their belief that AI has a reasonable chance of killing all of humanity within a very short space of time:

I’m not sure why this resignation and these statements went mega-viral. Plenty of researchers have made similar moves, and similar statements, over the past few years! Geoffrey Hinton, one of the pioneers of modern AI, quit Google back in 2023 over safety fears. Daniel Kokotajlo resigned from OpenAI in 2024, saying that the company wasn’t behaving responsibly in its drive toward superintelligence. William Saunders and Steve Adler did something similar. Mrinank Sharma left Anthropic earlier this year, and wrote a pretty well-read blog post about it.

What’s more, it’s been clear for years now that “AI could kill humanity” is a very common belief among AI researchers. Grace et al. (2024) interviewed thousands of AI researchers in 2024, and found that more than half thought that artificial superintelligence has a significant chance of making the human race go extinct (or causing similarly bad consequences):

The median AI researcher gave “doom” a 5-10% probability (depending on how the question was phrased), while their average probability was between 15% and 20%. Later, smaller surveys found similar numbers. The AI researchers may or may not be right, but the fact that lots of them think AI could kill the human race has never exactly been a secret.

It’s not clear why Coxon went so much more viral than his predecessors. Maybe it was the fact that AI just solved one of the most important open problems in mathematics (which the best human mathematicians had been unable to solve for almost a century). Or maybe it was the Hugging Face attack, where a swarm of AI agents tried to cheat on a test by hacking various companies. Or maybe AI has just obviously gotten so much smarter that people throughout society were starting to get worried.

But whatever the reason, Coxon’s announcement was the one that really penetrated through to the public consciousness. Suddenly, he was getting interviewed about AI doom on national news:

Barack Obama is now urging Democrats to focus on AI risk. Other politicians are calling for federal regulation. Bernie Sanders is drafting a bill to ban AI “superintelligence”, including 20-year prison sentences for anyone working on the technology. Donald Trump is getting asked about an AI slowdown; so far he’s resisting the calls, but there are rumors that his advisors are calling on him to do something.

Perhaps the most notable response came from the top figures in the AI field. Dario Amodei, the head of Anthropic, wrote a blog post called “We Must Pace the Frontier”, calling for a coordinated slowdown in the rate of AI progress, and suggesting some ways to police AI companies to make sure they were all observing the slowdown. He wrote:

[O]ver the last few months, I have become convinced that fully addressing the risks requires even more prudence — not just investing in risk prevention, but pacing the rate of capabilities advancement so that risk prevention has time to keep up. We must slow the pace at which we improve the capabilities of AI models…I’m therefore proposing a three-step plan with the goal of pacing the frontier: building AI at a balanced rate that aims to ensure its safety while still achieving its benefits and grappling with important geopolitical dilemmas.

As reasons for his increased worry, Dario cites A) the Hugging Face attack, and B) the possibility that AI will soon be able to improve itself without human help (a process called “recursive self-improvement”, or “RSI”).

Elon Musk (head of xAI), Sam Altman (head of OpenAI), and Demis Hassabis (former head of DeepMind) quickly agreed with Dario:

At least some of the labs are reportedly holding secret talks on joint action to slow down AI.

This is pretty extraordinary. A coordinated slowdown in AI progress would be bad for these companies’ bottom line, because it would allow upstart competitors to catch up. So the fact that they’re still calling for a slowdown, in defiance of their own financial interests, is a clear sign that their worry about human extinction is sincere.

In fact, anyone following these figures’ public statements over the past few years will have no doubt that they’re all deeply worried about catastrophic AI risks. The leading AI figures — not just the founders and CEOs, but the researchers themselves — feel trapped in a “red queen’s race”. They feel like if they stop working on AI, someone else will build it anyway, so they each feel like they have to beat everyone else in the AI race so they can make sure that the safest possible AI (i.e. their own AI) is the one that becomes the most powerful and dominant.

Anyway, all of this was common knowledge in my social circle years ago, but now all of it has broken through to the mainstream. What do I have to add to this discussion? I’m not an AI researcher or founder, nor do I think I have a superior grasp of the game theory of AI development. But I do think I have two useful thoughts on how to persuade the general public to be more concerned about AI risk.

The first of these is something I’ve written about recently. The second is about how to get China on board for a big AI safety push.

“Oh come on. How could AI kill all of humanity?”

As soon as everyone started talking about the possibility of AI killing humanity, there were two main types of pushback. The first was skepticism. A strange coalition of natural skeptics, libertarians (for whom any restrictions on technological development are a priori bad), and progressives (who have spent the last few years telling themselves that AI doesn’t really work) kept asking the question: How, exactly, is superintelligent AI supposed to kill us all?

This is actually an important and good question to ask. In my experience, AI researchers tend not to think very hard about this question. The reason is that they just assume that if AI gets smart enough, it will be able to kill humanity, and since its motives are alien and inscrutable, it might have its own reasons for wanting to do so.

Maybe superintelligent AI thinks humanity is an evil species who needs to be punished for torturing pigs and chickens. Maybe it’s scared that humanity might interfere with its other goals. Maybe it just wants to turn everything into paperclips. Who knows? AI researchers tend to think of superintelligence as the proverbial 800-pound gorilla who sleeps wherever he wants. As soon as humanity is no longer the most intelligent thing on this planet, our destiny as a species is simply out of our hands.

But to many people, that answer isn’t good enough. They want an actual plausible path by which a piece of software, which exists inside a computer, could slaughter real physical human beings. Fortunately (or unfortunately), there’s a pretty clear and simple answer to this question, which I wrote about two weeks ago. The answer is “bioweapons”:

(This article was paywalled originally, but I un-paywalled it.)

In my post, I wrote a scenario in which a nihilistic angry teenager uses superintelligent AI to release a world-ending bioweapon by ordering it from a gray-market laboratory somewhere in the world. But it’s also possible that a rogue AI agent swarm could decide to do this on its own, just as a way of cheating on some test that human researchers give it. The point is that AI can design viruses, and viruses can potentially kill off all or most of humanity.

A lot of biologists are skeptical of the idea that even the most superintelligent AI could successfully design a doomsday virus. They argue that this is just too hard of a task — that without much better biological data, it’s impossible to understand biological processes well enough to know how to design a virus with all of the necessary doomsday properties.

I urge you not to listen to these biologists. In this case, their expertise might be more of a liability than an asset. They know how hard it is for human beings to model biological processes, given existing data. But this doesn’t necessarily tell us how hard it is — or how hard it will be in five years — for AI to do it! Until LLMs came along, human researchers basically failed to understand natural language, even with all the data on the internet; AI can just do it. Until AI solved the Navier-Stokes problem, forecasters gave it only a small chance of solving it anytime soon.

Domain experts consistently underestimate how quickly AI can master their field and surpass them, because they mistake human difficulties for universal difficulty. When mathematicians underestimate how well AI will be able to do math, the consequences are usually benign — we get some unexpected answers to some cool math puzzles.1 But if the biologists are wrong, and the AI of 2027 or 2032 or 2049 can design doomsday viruses, the consequence could be that our whole species dies.

So yes, we should be worried about vibe-coded doomsday viruses, and we should be doing everything we can to secure biology labs, police the modification of viruses and other pathogens, and so on. “Pacing” AI development would probably help here too.

How to get China on board for an AI slowdown

The primary argument I see against “pacing” AI development is that if American companies slow down, Chinese companies will simply overtake them and build superintelligence themselves. For some, a China-controlled super-AI is a more terrifying possibility than super-AI in general:

But for others, it simply means that slowing AI down is futile because the Chinese can’t be persuaded to slow down:

This is an incredibly reasonable concern. The U.S. is still ahead of China in the AI race, but only just barely. If China is going to create superintelligence no matter what we do, why should we stop developing our own? Unless China can be persuaded to cooperate with the U.S. on AI “pacing” — or at least undertake its own independent “pacing” effort at the same time — anything we do will be futile.

So if we want to slow down AI development, we need to scare the Chinese leadership about superintelligence. There’s no other way.

How do we do that? In a post a week ago, I suggested in passing that simply staying ahead of China in the AI race might persuade them to embrace an AI slowdown, because that would be to their competitive advantage. But upon further reflection, I think I was pretty obviously wrong. If China will only embrace a slowdown if America refuses to slow down, then that’s game over — there’s no way to get both countries to slow down at the same time.

There’s a better approach. China’s leaders must realize that domestic dissidents could use Chinese-made superintelligence to overthrow the Chinese Communist Party.

Currently, China’s worries about AI mostly center around ways that the U.S. government could use U.S. AI models to attack China. That obviously gives the government an incentive to accelerate domestic AI progress, so that China’s own models can stand up to America’s in a fight. But if Chinese leaders realized that superintelligent AI could create a threat from within, this calculus would change.

Thus, China’s leadership must understand that Chinese AI models can pose a threat to CCP rule. The best way to demonstrate this is for American intelligence agencies — or even private hackers — to attack Chinese digital infrastructure using agent swarms created with China’s own frontier models like Z.ai’s GLM-5.3 or Moonshot AI’s Kimi K3.

When I say “attack”, I don’t mean actual warfare. I mean the kind of cyberattacks and data theft that China carries out against America every day. Use Chinese models to steal the CCP’s most heavily guarded secrets and post a few of the more innocuous ones on RedNote. Hack into Xi Jinping’s bank account and steal 100 yuan. I’m talking about demonstration attacks.

And these attacks must be done with Chinese models, not with American ones! If the CIA or some EA nonprofit in Berkeley uses GPT Astra or Claude Mythos to hack the CCP, China’s leaders may well conclude “Wow, we need to win the AI race so that our own models can defend us.” But if China’s own open-weight models are used for the attacks, Xi Jinping and the rest of the leadership will realize that their own push for superintelligence is making them incredibly vulnerable to any Chinese dissident who decides to overthrow them.

As soon as China’s leaders see superintelligence as a threat to their rule, I predict they will act. And their action will probably be to curb the development of superintelligence, especially if they know that America and its AI labs want to do the same.

In fact, China’s current leadership has a history of cracking down on its tech companies when it seemed like those companies might threaten the government’s monopoly on power. In 2021, Xi Jinping cracked down on Chinese software companies; he even (probably) apprehended tech magnate Jack Ma, who had criticized the CCP a little too openly. This action hurt China’s competitiveness in the online services industry, but the government went ahead and did it anyway.

And there’s already a precedent for demonstration attacks against Chinese digital infrastructure. An American cybersecurity company just used AI to develop a computer worm capable of hacking over a billion accounts on the Chinese messaging service WeChat:

Palo Alto-based Calif disclosed the already-patched computer worm to warn the public about the threat of AI-driven hacks…“Exploitation takes only seconds, and gives us full control of the WeChat account. We can read and send messages, make calls, and act on the victim’s behalf,” the company warned, posting a video demo of the WeWorm attack.

But Calif didn’t say what model it used to create WeWorm. Anyone who does this sort of demonstration in the future should make it clear that Chinese open-weight models were used, in order to make China’s leaders realize that the threat comes from their own too-rapid AI development, rather than from American competition.

I believe that this is our best bet for getting China on board for a joint international AI “pacing” effort. If there’s one thing the CCP fears more than an American attack, it’s domestic dissidents overthrowing the Party from within. Superintelligence is creating that vulnerability, but the leadership doesn’t seem to have realized it yet.

Make them realize, and I predict that a whole universe of possibilities for international cooperation will suddenly open up.

Update: Unsurprisingly, I’m not the first to have this idea about China needing to be scared about the internal threat from their own models. Back in July, Kyle Chan wrote:

Previously more dismissive of concerns over AI-driven job loss, Beijing now seems to be taking these risks more seriously (see Matt Sheehan’s great piece). China is also watching developments in the US very closely, particularly controls on Anthropic’s Mythos and Fable models over cyber risk. As Chinese open-source models approach similar levels of cyber capabilities, this could come back to bite and be used to potentially attack China’s own digital infrastructure. [emphasis mine]

In general, Kyle’s blog is one of the best blogs about China. Highly recommended.

Update 2: As if on cue, the NYT has a story today about how the CCP is starting to realize that superintelligence is a threat to its rule!

China’s top spy chief has warned that artificial intelligence could pose a direct threat to the Chinese Communist Party’s hold on power, in what is the highest-level and most detailed articulation yet of how Beijing sees the technology’s security risks…In an article published on Sunday in the state-run magazine China Cyberspace, Mr. Chen called for more party control over A.I. and stricter government oversight…He also described the danger that foreign intelligence agencies might use A.I. for “large-scale espionage” and attacks on China’s critical infrastructure…

Mr. Chen, the spy chief, wrote that the technical and financial barriers to launching cyberattacks had been drastically lowered because of A.I. This, he said, posed “serious risks” to China’s information infrastructure…Chinese users using foreign models could cause large-scale data leaks, the article noted. Anthropic, the company behind Claude and other A.I. models, said last week that the Chinese start-up Moonshot AI had routed queries by its users to Claude. Those queries included sensitive data, including video surveillance linked to the Chinese military, as well as proprietary information from high-profile Chinese technology companies…

The warning adds to a growing drumbeat of concern from Chinese officials about the risks of A.I. [emphasis mine]

Now all they need to realize is that the biggest threat comes from THEIR OWN MODELS.

Subscribe now

Share

The other day I saw a Harvard Harris poll on the favorability of various political figures in the U.S. Only RFK Jr. (sigh), Marco Rubio, and Mark Carney had net positive approval ratings:

While I understand that these polls all have methodological issues, I was struck by how three figures — Abdul El-Sayed, Zohran Mamdani, and Benjamin Netanyahu — had net approval ratings below almost any American politician. While Mamdani’s low rating might be partially due to his leftist policies and his association with the DSA, it struck me that all three of these unpopular figures had one thing in common — they are associated with the intrusion of Middle Eastern politics into American politics. And I thought: Maybe Americans just want to have nothing to do with the Middle East.

Today1 is September 11, the anniversary of the terror attacks 25 years ago. Though I don’t agree with the leftist line that those attacks were blowback from American involvement in the Middle East, it’s unquestionable that 9/11 inaugurated an era of heavy U.S. involvement in that area of the world — the Iraq War, the follow-up war against ISIS, the War on Terror, greater support for Israel, an increased military presence in the region, and finally the war against Iran.

I don’t believe that all of that involvement was a bad idea — in particular, as I’ll argue, the War on Terror, including the fight against ISIS, was generally successful and necessary. But overall, the era of heavy U.S. involvement in the Middle East has been a disaster for U.S. power and prestige, and it’s time for that era to end.

This is not a recent conversion on my part; I’ve been in favor of general disengagement from the Middle East for a very long time. Five years ago, in the early days of the Noahpinion Substack, I wrote a post called “How to fix U.S. foreign policy”:

In that post, I wrote:

[D]espite modest progress, we still remain too tied up in Middle Eastern conflicts…The U.S. will never be able to ignore the Middle East entirely, but we can remove ourselves much more than we’ve done…[W]e need to draw down engagement in the Middle East, and increase engagement in Asia.

Since then, things have gone from bad to worse, making my message far more urgent than it was at the time. Most obviously, there’s Trump’s disastrous war in Iran, which continues to waste U.S. military resources and make the U.S. look weak while failing to accomplish any significant objective. Recent polls show that a very strong majority of Americans think that the Iran War is not worth fighting.

There’s also Israel’s intensified campaign against the Palestinians. This has become the rallying cry of the American left, which has prompted a backlash and crackdown by the Trump administration. As a result of that conflict, Middle Eastern politics is increasingly becoming the focus of American domestic politics.

The controversy around Michigan Democratic Senate candidate Abdul El-Sayed, for example, has nothing to do with his policy stances, or even American culture wars — it’s entirely about his stance on Israel/Palestine and his ally Hasan Piker’s comments making excuses for 9/11 and Islamic terrorism. El-Sayed recently apologized for tying a synagogue shooting to Israeli warfare in the Middle East.

Many analysts believe that Zohran Mamdani, meanwhile, won the NYC mayorship primarily because of his outspoken stances on Israel/Palestine. And more is on the way. Organizations like CAIR Action that are linked to the Muslim Brotherhood — a sort of international equivalent of the Christian Coalition, but for Islam — are now supporting political candidates throughout the U.S., using the Palestine issue to recruit candidates and win votes. Muslim Brotherhood figures are cheering this effort on from overseas. AIPAC and other Israel-supporting groups, meanwhile, are pouring money into races in order to stop those candidates.

Our media, too, is increasingly filling up with the Middle East’s cultural and religious conflicts. Leftist shouters like Hasan Piker, Mehdi Hasan, and Cenk Uygur constantly urge progressives to base their politics around Palestine. Every day, Americans are bombarded with rhetoric like this:

This is simply bad for America. We have so many pressing issues to worry about — Trump’s corruption and authoritarianism, inflation, AI, etc. We should not be nominating Democratic politicians based on whether they support Israel or Palestine; we should be nominating them based on their stances on issues of direct relevance to Americans, and on their ability to oppose Donald Trump. Furthermore, mobilizing Muslim and Jewish voters to vote based on Israel/Palestine, rather than on issues of importance to Americans in general, provides fuel for rightist attacks on both groups.

What the United States needs, now more than ever, is general disengagement from the Middle East — an end to military adventures, a dialing back of support for Middle Eastern “allies” (including Israel), and a forceful rejection of Middle Eastern issues in American domestic politics.

Fortunately, conditions in the world have shifted. The kind of disengagement I’m calling for is a lot more feasible than it was twenty or even ten years ago. The Iran War has shown that Middle Eastern oil is a lot less important for the global economy than it used to be. And the success of the War on Terror and the general decline of Islamism reduce the necessity for further interventions.

The U.S. doesn’t need to protect Middle Eastern oil supplies anymore

For many decades, the main justification for U.S. intervention in the Middle East has been to keep oil prices low and stable. The Middle East is one of the world’s largest oil producers, and its low extraction costs mean it functions as the world’s “swing producer”. Supply disruptions from the region can cause global price spikes, endangering industrialized economies that depend on petroleum. Oil was a big reason the U.S. intervened to stop Saddam Hussein’s invasion of Kuwait in the early 90s, for example.

Obviously, oil is still a very important commodity. But it’s a lot less important to the U.S. than it used to be. Here’s what I wrote when the Iran War began:

Blanchard and Gali (2007) looked at economic responses to changes in oil prices in the U.S., and concluded that the economy of the 2000s was only about a third to half as sensitive to the price of oil as the economy of the 1970s had been. Their reasoning is that modern economies are more flexible in general, that they have better monetary policy (i.e. we don’t try to print a ton of money in response to a supply shock), and that we depend on oil less.

By their estimates, a 10% increase in the price of oil now (or at least, if “the 2000s” means “now”) leads to only a 0.25 percentage point increase in the CPI and a 0.3 percentage point reduction in GDP over the course of a year or so. Since oil just spiked by 50%, then if that’s sustained, we might expect to see inflation go up by 1.25 percentage points, and GDP go down by 1.5 percentage points over the next year. That would mean inflation would go to around 4% and GDP growth might go down to 1.5% — frustrating and annoying, but not catastrophic…Other estimates seem similarly modest. For example, in a recent roundup, I flagged a paper by Känzig and Raghavan (2025) that looked at the closure of key shipping chokepoints.

In addition to having a more flexible economy, better energy efficiency, and more reasonable monetary policy, the U.S. is now a net oil exporter. That means when oil prices go up, the benefit to the energy sector cancels out at least part of the harm to other sectors of the economy.

But on top of all that, the fact is that the Iran War simply hasn’t raised oil prices that much! Despite Iran’s closure of the Strait of Hormuz, and Trump’s retaliatory blockade of Iranian oil shipments, oil prices haven’t even reattained their highs from the early 2010s:

And in real terms, the spike is even less impressive. If we divide oil prices by average hourly earnings for production and nonsupervisory workers — basically, how many hours an average American worker would have to work in order to afford a barrel of oil — we see that prices really aren’t that high at all:

No wonder the U.S. economy is still doing fine, and inflation has only risen a little bit (and some of that may be due to demand from the data center boom).

Read more

So, Germany’s far-right party, the AfD, just won a huge landslide election victory in the eastern state of Saxony-Anhalt. This doesn’t mean AfD is taking over Germany — the far right is much more popular in the post-communist east than in the more populated western regions. Germany is also a parliamentary democracy, and almost everyone else hates and fears the AfD. But the statewide victory is a warning shot, and everyone in Germany — and in Europe — is taking it seriously.

AfD’s surge means two things. First, it means that the rise of far-right parties was not simply a reaction to the pandemic. Back in 2024-25, a number of pundits I respect argued that voters in the U.S. and Europe were backing rightist leaders out of anger at incumbents over Covid. That thesis now looks dead.

Second, it means that something is making many German voters so angry that they’re finally willing to overlook all of the AfD’s catastrophic flaws — its obvious ties to Vladimir Putin, the fact that the party is filled with crooks, its extreme rhetoric, its lack of governing experience, and so on.

What is that “something”? A number of socialist types have tried to claim that AfD’s new voters are responding to economic despair, and can be won back with more socialism. But more sober voices agree that the “something” is immigration. Adam Tooze, who strongly leans to the left, tries to be gentle as he explains that there’s no getting around the immigration issue:

If you ask AfD voters what they worry about, along with their xenophobic concerns about “too many foreigners”, they will mention the cost of living and their concerns for their standard of living. They also mention climate change and the threat of external war…But if you ask them what they think about the AfD and why they voted for it, economic issues are barely mentioned. What comes first is security, foreigners, woke politics, remoteness etc…

If you are hard up and working-class it would after all be nothing short of bizarre to vote for the AfD in the hope of improved social protection…Voting for the AfD in the current landscape, whatever your economic issues, in short, is not a vote to have your economic issues addressed by social democracy, or welfare or government intervention. It is a vote to crackdown on foreigners[.] [emphasis mine]

Robin Brooks is more blunt:

There’s lots going on in Germany at the moment, but - when it comes to the surge in AfD popularity - you don’t need to overthink it. The AfD is relentlessly campaigning on immigration, which tells you that this is what resonates most with voters…A key driver of the recent surge in AfD popularity is that the CDU - SPD grand coalition is proving itself incapable of confronting immigration with the urgency many voters want.

In fact, anger over immigration is probably the main reason that the right is surging everywhere in Europe. Here are Konstantinou and Roumanias (2024) with some data:

We use regional data spanning over 4500 electoral outcomes…between 2000 and 2017 to assess the impact of immigration on Western European Far-Right voting. To deal with potential reverse causation, we use immigration in neighboring countries as an instrument for domestic immigration. The estimated effects of immigration are positive, significant and larger compared to those obtained by simple OLS and fixed effects regressions. Depending on the measure of immigration used, we find that a 1% increase in immigration stocks leads to a between 1.78% and 2.97% in Far-Right voting…Our estimated effects explain a large part of the observed rise in Far-Right in Western European countries that experienced high levels of immigration during the span of our sample.

2026 is hardly the beginning of the trend. Here’s Brookings in 2024:

In the European Union (EU), one election after another has demonstrated the centrality of irregular migration and border security in public discussions and forced mainstream parties to take more restrictive approaches…The results of the European Parliament election, France’s snap election, three German state elections, and the Austrian election all showed a strong rightward drift and signaled voters’ distrust in their national governments, confirming the notable shift in tone on migration in Europe toward a more securitized, hardline approach, even among mainstream parties.

Trump’s election in 2024 was also partially about immigration. Trump voters rated it as one of their top issues, along with inflation. It was the issue where Trump had the strongest lead over Kamala Harris in polls. American voters clearly soured on immigration in 2023-24.

Progressives in America and socialists in Europe still desperately claim that what angry working-class voters really want is more economic support. They keep offering more and more lavish promises of welfare supports and price controls, while staunchly opposing restrictions on immigration. The approach keeps failing, and the far right keeps rising.

Center-right and center-left politicians, on the other hand, realize the source of their voters’ anger very clearly, and are taking steps to reduce immigration. Germany’s center-right CDU has already stanched the flow with restrictive policies:

And Chancellor Friedrich Merz is responding to AfD’s gains by making it harder to get citizenship. Meanwhile, France and Italy are tightening border controls. France is making it harder to become a citizen. The EU is toughening up its asylum rules. Sweden is offering migrants large sums of money to leave the country. European countries are streamlining their infrastructure for kicking out large numbers of failed asylum-seekers. Japan is tightening its own immigration policies after a period of liberalization.

Even Canada, one of the most immigration-friendly and progressive countries on Earth, has scrambled to reduce immigration, cutting targets for permanent residents, reducing student visas, and slashing the number of temporary workers. Immigration levels are still high, but are falling quickly — all under the center-left government of Mark Carney.

And yet I’m worried that voters around the world will see these measures as superficial — tactical pauses to let populist anger simmer down before resuming the same permissive immigration policies that made voters mad in the first place. This certainly seemed to be the case in Europe, where the government stopped admitting asylum seekers from the Middle East in the late 2010s after a popular outcry, only to resume taking them in en masse in the 2020s:

And it also seems to be true in the U.S.; Biden mostly closed the border to asylum-seekers in 2024 after he saw how mad it was making Americans, but many Democrats are now calling for the abolition of ICE. This is a somewhat popular position right now, thanks to the agency’s abuses under Trump. But if Democrats actually try to abolish federal immigration enforcement instead of just reshuffling and renaming the agencies, I predict it will take just a couple of years for Americans to realize that the difference between “open borders” and “immigration laws with no one to enforce them” is largely semantic.

What’s needed, I believe, is not just a tactical, temporary retreat. We need to stop and think why a maximal pro-immigration policy is something that we need to retreat from in the first place. We need a new concept of what a sustainable liberal immigration policy looks like.

Since I started writing, I’ve been an advocate of more immigration to the United States. This does not mean I ever favored unrestrained immigration or open borders. I’ve consistently argued that favoring skilled immigration will bring more economic benefits to native-born Americans, and be more politically palatable as well. And while I don’t hate illegal immigrants for trying to better their lot in life, I favor strong border controls, because nations have the right to decide, democratically, who gets in and out. It’s understandable when people get mad at seeing their decisions flouted.

On top of that, I’ve always remained agnostic on immigration to other countries; although I find much to admire in Canada’s points-based system, for example, I don’t think that qualifies me to decide whether more immigration is good for Canada overall. If the people of a country decide that immigration is diluting their local culture unacceptably, for instance, I think they have every right to cut it off. That’s just Westphalian sovereignty — not a perfect system, but the best system we’ve ever found for organizing humanity into geographic units. As an American, I don’t view it as my place to tell Japan, or the UK, or any other country that immigration is the right choice for them.

Despite those reservations, I’ve always thought that immigration to the U.S. is a basically good thing, and should be expanded. The economic benefits are pretty undeniable — higher tax revenue to shore up our dangerously depleted government finances, dominance in high-tech industries and innovation, and so on. And while some of the economic costs are real — local housing shortages and strains on local government finances being the two biggest ones — many of the fears are overblown. In particular, the bulk of the evidence concludes that immigrants don’t reduce wages or job opportunities for native-born Americans.

And importantly, lots of Americans share my overall positive view of immigration. If you ask Americans whether they think immigration is good or bad, most will say “good”:

And if you ask Americans whether the annual rate of immigration should be increased, decreased, or kept the same, you get a pretty even split:

So I’ve never really felt like I was going out on a limb or advocating an unpopular position when it came to this issue. Sure, rightists will direct online vitriol toward anyone who supports any immigration at all, but their energy and savagery shouldn’t be mistaken for majority support.

And yet there will always be some amount of immigration — and some types of immigration — that will arouse even the most open and welcoming people to ire and make them think about shutting the doors. We saw this in the U.S. in 2023-24, when anti-immigration sentiment spiked in response to Biden’s permissive asylum policies. Yes, that sentiment crashed again when Trump took power and started committing abuses. But if liberals don’t change how we approach immigration, the sentiment will simply rise again and again, as it has so many times throughout our history. And the Donald Trump and Stephen Miller types will be right back in power — and our country will be worse off for it.

Liberals need an immigration approach that can be sustained for long periods of time — i.e., one that doesn’t enrage the public every time it manages to get in power. What would that look like?

The first thing liberals (and progressives, and Democrats, and European lefties, etc.) need to admit is that migration is not a human right. We live in a world of sovereign nation-states, and unless we find some better way of dividing up and administering the world, we will continue to live in a world of sovereign nation-states. And nations are, necessarily, exclusive clubs; they have the right to restrict immigration for any reason whatsoever.

Liberals must therefore not view borders and citizenship as annoying obstacles to be tactically circumvented or overcome; instead, we must view them as fundamental parts of the social compact that allows nations, including liberal nations, to exist in the first place. Nation-states and their laws and their police and their courts and their armies are the fundamental guarantors of the human rights that define liberalism. And for better or for worse, the ability to decide who gets in and who has to keep out is a necessary precondition for nation-states to exist.

That means we need to recognize that immigration law is legitimate and needs to be upheld. Some progressives have tried to advance the notion that being undocumented is a marginalized identity that needs to be protected and supported by the state. But this is absurd; it’s like if progressives decided that people who drive over the speed limit are a minority group.

Illegal immigrants chose to break U.S. law when they came to this country. Our democratically elected leaders made those laws, and they should be honored. We should treat illegal immigrants humanely, of course, and we should not tear local communities apart just to hunt down a few people. But there has to be some way of enforcing the law, because a law without enforcement is no law at all.

A third principle that liberals should embrace is that the purpose of immigration is to benefit the people who already live in the country that is receiving the immigrants. Immigration to America must be for the benefit of Americans. We must flatly reject any concept of immigration that sees it as a necessary sacrifice on the part of American citizens.

For example, the book This Land is Our Land, by Suketu Mehta, argues that immigration is a form of reparations for colonialism — a sacrifice that Western nations owe to other nations, even at their own expense. This idea has found some purchase in progressive circles, but liberals must reject it as wholly illegitimate.

Instead, liberals need to promote immigration because of the benefits it brings to the American citizenry. This includes economic benefits — the tax revenue, the investment, the eldercare, the innovation, the entrepreneurship, the expanded market size. But it also includes cultural benefits — the constant reinvigoration of this nation of immigrants by new waves of people with the gumption and bravery to pick up and move across the world in search of opportunity and freedom.

America has always been the country of the frontier; in order to maintain that ethos in the modern age, we need people for whom America itself is the frontier. And our founding ideals — which are very liberal ideals, of liberty and opportunity and the rights of the individual — are strengthened by people who make the conscious choice to move to a country that represents those ideals.

These are the kinds of arguments liberals made in the past. We can — we should — make them again.

And if this requires us to be selective about which immigrants we bring in, then so be it. Skilled immigrants bring far more economic benefits per capita than others — so by all means, let us tilt our system toward them, as American voters of both parties want. Immigrants who don’t love and embrace American culture will probably invigorate our nation less — so by all means, let us bar immigrants who have been part of groups that see America as evil, using the law with which we once banned immigration by members of communist parties.

These are the kinds of decisions that nation-states, including liberal nation-states, are inherently entitled to make. And we should expect Europe to make different decisions than America makes. That’s perfectly OK. Because immigration is fundamentally the decision of each sovereign nation-state, it’s inevitable that we’ll make different decisions. JD Vance wants us to think that France’s immigration policy, or Germany’s, is inextricably tied to America’s. But it’s not, and we shouldn’t let him get away with conflating the two.

A liberal immigration program based on these three principles could defuse the anti-immigration sentiment now sweeping the developed world — not just now, but permanently. But it will only do so if liberal leaders explicitly articulate and avow these principles. That’s what’s required in order to convince the electorate that liberals are trustworthy on immigration — that embracing the kind of immigration policies Americans demand represents a principled stand rather than a tactical retreat from a hidden agenda of open borders.

For America, the kind of immigration agenda I’ve sketched out — which all the polls show Americans favor by substantial margins — would not be a “far right” policy, even though some progressives and leftists will inevitably try to label it as such. In fact, it would allow in far more immigration than the policies of Franklin D. Roosevelt or Harry Truman. It won’t satisfy progressives who dream of a borderless world, or leftists who salivate over the chance to make the West pay for colonialism. But I believe it will preserve America as the kind of liberal nation-state that we knew it as in the days before Trump.

As for the rest of the world, countries like Germany and Canada and Japan have to make their own decisions. I can’t tell them what kind of nation to build; I can only try to promise that as an American, I’ll respect their decision. If they want to shut their doors in order to slow the pace of cultural change, it’s not my job to lecture them otherwise. All I can say is that I hope they do what they need to do in order to avoid getting taken over by parties that take their marching orders from the Kremlin.

But speaking only for my own country, I believe that if America gives up immigration, it will lose one of the key things that has made it great. With the GOP in the hands of a xenophobic rightist movement that sees immigration as an invasion, it’s up to the Democrats to craft a system that can be preserved for the long term.

Subscribe now

Share

It’s Labor Day, so here’s a post about how human labor is alive and well in the age of AI.

I live in San Francisco and hang out with a lot of tech people, both in the AI industry and outside of it. And one thing that almost everyone I know here believes is that AI’s main economic effect is to displace humans from their jobs. Most people don’t have concrete arguments for why this should be true; it’s just an article of faith. The conventional wisdom is pretty well summed up by the first line of this tweet:

In fact, AI companies themselves have spent years talking about how their inventions are going to render large swathes of humanity economically obsolete — an odd marketing pitch, perhaps, but one that seemed to reflect their honest expectations.

A lot of times, San Francisco tech people are out of step with the general public. This time, though, the public seems to agree. A recent Ipsos poll found that most Americans expect AI to compete with human workers more than it complements them. And Pew finds that this belief has even strengthened in recent years:

So basically, most people think AI is a job-killer. And yet somehow, this job-killer keeps stubbornly refusing to kill jobs. In the aggregate, the labor market is about as healthy as it’s ever been. The prime-age employment rate — the single best indicator of how many Americans have jobs — continues to hover near all-time highs:

Of course, there are lots of other things going on in the labor market right now besides AI. But most of those things — tariffs, the Iran war, etc. — are bad for employment. It’s not easy to identify some sort of positive shock that is canceling out the job-killing effects of AI.

Or maybe it is, if the shock is AI itself. Theoretically speaking, automation can create jobs just as easily as it can destroy them. Here are Acemoglu and Restrepo (2019), explaining the various ways that technology can affect the demand for labor:

Automation [can be bad] for labor because of a displacement effect—as capital takes over tasks previously performed by labor…

[A]utomation technology also increases productivity, and via this channel, which we call the productivity effect, it contributes to the demand for labor in non-automated tasks…

[T]he displacement effect of automation has [historically] been counterbalanced by technologies that create new tasks in which labor has a comparative advantage. Such new tasks generate not only a positive productivity effect, but also a reinstatement effect—they reinstate labor into a broader range of tasks and thus change the task content of production in favor of labor. The reinstatement effect is the polar opposite of the displacement effect and directly increases the labor share as well as labor demand. [emphasis mine]

In other words, automation can do three basic things. Yes, it can replace people and take their jobs. It can also make them more productive, which can both create jobs and destroy them.1 And, crucially, automation can create new jobs for people to do. Power looms replaced master weavers, but they created jobs for technicians and engineers to make the power looms work. The internet automated much of the work of travel agents, but created jobs for web designers. And so on.

People who think of AI as a job-killer might not have thought of the second and third of these. Or they may have thought of them, but simply assumed they’re not a big deal. Anecdotally, a lot of tech people think that AI will keep substituting for more and more tasks until A) productivity increases just increase the demand for AI, and B) there are no new tasks left for humans to do. AI detractors, meanwhile — like Daron Acemoglu himself — often simply assume that new tasks created by AI will be “bad tasks” like misinformation and cybercrime that hurt the economy instead of helping it.

But these assumptions simply might not be correct. AI might be creating lots of new tasks for humans to do. For example, software engineers are writing less and less code themselves. Instead, they’re spending more and more time telling AI to write code — that represents a productivity improvement. But they’re also trying to figure out what code to tell AI to write, making sure AI is writing the kind of code they want, integrating that code into products, and so on. Those are all new tasks. There are also a lot of software engineers working on improving AI itself, and on discovering new applications for AI. Those are new tasks as well.

This helps explain why in the age of Codex and Claude Code, software developer jobs have been increasing as a percentage of total employment:

Anecdotally, organizations that thought they could replace lots of their software engineers with AI ended up having to hire many of them back — sometimes at a premium.

In fact, this is a story we see throughout the economy. Alex Tabarrok recently reported on a Census Bureau survey about AI that’s been running since 2023. The Census Bureau calls companies up and asks them A) how AI affected their total employment, and B) how AI affects the tasks that workers do.

Most companies reported no change in overall employment, which could just be due to inertia. But of companies that did report a change, more reported an increase than a decrease!

And here’s the breakdown by sector:

The story was similar for tasks. Tabarrok writes:

Among firms using AI, 44% say it supplemented or enhanced work an employee already does. Ten percent say it performed a task an employee used to do. Eleven percent say it introduced a task no one had been doing.

Here’s the chart:

Rigorous research, meanwhile, sometimes finds negative effects of AI on labor demand at the industry level, and sometimes not. But at the company level, the evidence is clearer — Kharazian, Simon, and Stevens (2026) find that when companies adopt more AI, they tend to hire humans rather than replacing them. Here’s a blog writeup of their findings:

Ramp Economics Lab

We can finally say AI isn’t killing jobs

Dear Colleagues: The most important economic question of this decade asks how AI will affect jobs. Everyone wants to write that paper. Until now, no one has had the right dataset, so existing research has relied on a combination of guesses, surveys, AI exposure scores, and self-interested punditry. In fact, a recent paper from Stanford said the ideal da…

Read more

3 months ago · 63 likes · 14 comments · Ara Kharazian

And here’s a chart:

Interestingly, they find the same for entry-level jobs — the jobs that people usually identify as being most under threat from AI.

So despite Acemoglu’s skepticism, it looks like for now, the new tasks being created by AI are probably matching or even slightly exceeding the tasks replaced by AI. Of course this measure is “number of companies” rather than “number of jobs”, but the pattern is pretty clear.

The Economist, meanwhile, has a report on how AI is creating jobs, both through the “new tasks” channel and by boosting demand in areas that AI can’t yet touch — physical jobs like construction and HVAC installation. Here’s what they write about the productivity/demand effect:

The Economist estimates that AI has so far created around 1m new jobs in America. That easily exceeds the roughly 200,000 lay-offs attributed to AI since mid-2023, and appears more than enough to offset weaker hiring in many back-office roles. America’s AI infrastructure splurge has created many of them…The Economist tracked five industries at the heart of the data-centre build-out, from electrical contracting to equipment manufacturing. Since 2023 employment in them has risen by roughly 320,000 more than broader…trends would suggest…LinkedIn, a social network for strivers, estimates that nearly half a million data-centre jobs were created between 2023 and 2025 in America, with data-centre technicians and engineers among the most common recent hires…

The scramble for workers is showing up in pay cheques, too. Indeed finds that installation and maintenance jobs at data centres advertise wages about 40% higher than comparable work elsewhere…In the year to June, average hourly earnings rose more than 13% in electrical-equipment manufacturing and nearly 8% among electrical contractors. [emphasis mine]

And here’s what they write about new tasks:

AI is also creating a new class of white-collar jobs. Engineers build the models, data annotators label their inputs and judge their answers, “forward-deployed” engineers adapt them for customers, and newly minted “heads of AI” decide what companies should do with the technology. Some of these roles barely existed until recently. Many are quickly growing in number. Postings for heads of AI, AI engineers and directors of AI have roughly doubled since 2023-24, according to LinkedIn…

Preliminary research by Gad Levanon, chief economist at the Burning Glass Institute…reckons roughly 1% of professional jobs are now “AI jobs”…[P]rofessional occupations closest to the AI boom—engineers, software developers, mathematicians and data scientists…have added roughly 730,000 jobs above trend in recent years[.] [emphasis mine]

What about specific occupations? Technology has certainly destroyed many specific types of jobs over the centuries — there are (basically) no more elevator operators, human telephone operators, or people who do manual typesetting for printing.

And yet in recent decades, we haven’t seen as much of this sort of occupational destruction. For example, a lot of people thought the internet would kill travel agents. And while the industry was hit hard, there are still plenty of travel agents left:

The reason is probably that the job of “travel agent” is much more flexible and “messy” than older types of jobs like elevator operator; travel agents do a whole lot of different tasks, so they’re harder to replace than people who just stand there and press a button. That makes modern jobs harder to replace entirely.

It’s a good bet that AI will eventually make some occupations obsolete. But so far, despite awe-inspiring progress in model capabilities, it’s extremely hard to find occupations that have seen significant replacement by AI. Top AI researchers who famously predicted the end of human radiologists saw their predictions get confounded. Truckers, too, are doing just fine. (Update: Here’s a good clip of Geoffrey Hinton explaining why he was wrong about radiologists.)

The most impressive example might be translators. It seems pretty obvious how AI could replace human translators, and yet it hasn’t done so yet:

Here’s a chart:

If you could go back to 2022, and tell people that in four years, AI would be solving frontier math problems, but we’d still have the same number of people working as translators, how many would have believed you?

It turns out that it’s very natural for people to overestimate the degree to which AI will take their jobs. Hartley et al. (2026) have a really excellent paper called “Job Loss Fears in the First Years of Generative Artificial Intelligence”. Here’s a thread explaining the paper’s findings.

Basically, the authors find that fear of AI job replacement is extremely common:

And they find that the more people’s jobs are exposed to AI, the more they think their jobs are about to be replaced:

In fact, the more of their day people spend using AI at work, the more they’re afraid of being replaced!

And yet when the authors looked for a correlation between AI exposure and actual job loss, they found…absolutely nothing. People’s fears simply haven’t come true yet.

What’s going on? The authors hypothesize that people who use AI more start to understand its ability to replace the tasks they do at work. But as we keep finding, replacing tasks isn’t the same as replacing jobs. People keep finding new things to do in their roles at work — sometimes things AI can’t do yet, but often things that couldn’t even be done until AI made them possible!

It seems like we’re uncovering a consistent human blind spot here: People don’t actually know how they produce value at their jobs. Modern jobs are much more than a simple collection of tasks — they are pieces of a complex machine that produces value in ways that an individual worker often doesn’t even see.2 So when AI comes along and starts replacing people at various tasks, it just ends up making them more valuable as pieces of their corporate machines.

How long that situation will persist, of course, is an open question. AI leaders are starting to realize that it might take a very long time for the full effect of their inventions to be felt:

This is why the AI companies’ recent messaging pivot — many now say that AI will create jobs rather than destroying them — may be honest, rather than a cynical marketing ploy to calm public outrage.

But then there’s the question: Can this situation persist indefinitely? No one knows, of course. But my bet is that while many occupations will eventually be mostly replaced by AI, humans will still have plenty to do. I’ve argued that in order for AI to start replacing human jobs wholesale, it’ll have to get much more agentic — which will make it inherently more unreliable from a human point of view. So I predict that humans will always have jobs keeping AI agents on track.

Even if I’m wrong, though — even if the AI job apocalypse does eventually come — it doesn’t seem like it’s coming soon, and it certainly isn’t here right now. Everyone keeps thinking that AI is a job killer, and AI keeps on refusing to be what everyone expects.

Happy Labor Day!

Update: John Cassidy has another good post on this same topic.

Subscribe now

Share

Most of the debate around AI, at least in the U.S., is not about the international aspect. The local political debate is all about data center construction; the national economic debate is mostly about fear of job loss, with a side discussion about a potential bubble; and the technological discussion, at least in public, is mostly about AI safety and risk. U.S.-China competition gets mentioned in certain circles, but it’s probably safe to say that it’s not Americans’ chief topic of concern.

But it still matters! For one thing, there’s the military aspect to think about. Cyberwarfare so far hasn’t been decisive in military conflicts, but AI’s incredible cybersecurity prowess could change that. If AI ends up strengthening defense more than offense — say, by finding all of the available exploits and patching them before an attacker can get to them — then cyberwarfare will become less important. But if those who possess the best AI models are able to successfully hack anyone using a less capable model to defend, it could lead to a decisive shift in the balance of power.

AI hacking doesn’t have mutually assured destruction, like nuclear warfare does. Imagine if China were to gain a big lead in AI models that gave it the power to easily hack into American banks and brokerage accounts and erase people’s wealth. It would cause absolute chaos in American society, but how could the U.S. retaliate? Launch nukes? Nor could the U.S. hack China in return, since China’s more capable AI would also be used to defend.

If either country opens up a large, sustained lead in AI capabilities, it might upend the balance of power between the two.

Not all AI issues are zero-sum, of course. If the U.S. and China both continue pushing forward with AI research at maximum speed, it may quickly cause safety issues. The recent AI agent swarm attack on Hugging Face shows that AI has reached the level where it can pose a significant hazard to human companies and organizations — and perhaps soon to human society itself. Bioterror risk is certainly the most terrifying, but there are plenty of other ways that highly capable AI could cause chaos.

The U.S. and China have a shared incentive to implement strict safeguards against these catastrophic risks, and perhaps even to regulate the pace of AI development. But given the Chinese Communist Party’s power-seeking nature, it seems much more likely that China would agree to cooperate on AI safety if U.S. capabilities were comfortably ahead. So even if the goal is cooperation, the U.S. should be thinking about how to keep its technological edge.

Fortunately, the U.S. is still beating China in the AI race. Our companies have better models, more compute, and far more revenue. But there are ways that the Trump administration, despite claiming to be the AI industry’s best friend, could squander America’s lead — especially by pushing Chinese AI talent out of the country.

U.S. models are still better than Chinese models

There have been several moments when it seemed as if China’s frontier models were catching up to America’s in capabilities. The most dramatic was the “DeepSeek Moment” in early 2025, which put Chinese AI on the map. More recently, the release of Moonshot’s Kimi K3 this July and Z.ai’s GLM-5.3 a few weeks ago seemed to indicate that Chinese models were nipping at the Americans’ heels.1 Z.ai especially made waves when it beat Anthropic’s famous Mythos model on one measure of cyber-hacking capabilities:

Chinese AI ​startup Z.ai said on Friday its open-source GLM-5.3 model had neared Anthropic’s restricted Mythos 5 in identifying software vulnerabilities…Z.ai said GLM-5.3 scored 84.5% on CyberGym, a test of whether a model can review code, identify security flaws and confirm that they are real. That was slightly higher than the 83.8% it reported for Mythos 5. The results have not been independently verified.

Note that this is just one measure of cybersecurity prowess, and that Mythos was still comfortably ahead on other measures:

GLM-5.3 lagged behind Mythos 5 in converting discovered flaws into working attacks — a standard part of defensive security research. Z.ai ​said its model scored 54.4% on the ExploitBench test of this capability, versus 78.0% for Mythos 5…In a separate timed test, Z.ai said GLM-5.3 completed 105 ​attack-development tasks in two hours and 130 in six hours. Mythos 5 completed 181 and 247 tasks, respectively.

But still, if Chinese AI could get within striking distance of America’s best, it was a big deal.

What this discourse rarely mentioned, though, is that Mythos is not America’s best. It was simply the best that’s been released. Mythos Preview came out in April, four months before GLM-5.3. And the original Mythos actually finished training three months earlier, in January, and was released internally in February.2 Anthropic delayed its release due to cybersecurity concerns. Z.ai, being a fast follower, probably had far fewer such concerns. In fact, Anthropic has stated that it has internal models that are better than Mythos.

Read more

Stripe Press was kind enough to send me several books it had published, and one of these was J. Storrs Hall’s Where Is My Flying Car?. I had been meaning to read this book for a while, especially after Tyler Cowen recommended it and Jason Crawford gave it a glowing review, so I was glad Stripe gave me a copy. I promptly stuck it on my shelf, paid for the audiobook, and listened to it while walking around San Francisco. That’s called ordinary abundance, folks!

In 2011, Peter Thiel famously said: “We wanted flying cars, instead we got 140 characters.” When I first saw that quote, I chuckled, because for my entire life, we’ve had flying buses. And again and again, as I sit on a flying bus and marvel at the thrill of takeoff, and reflect on how amazing it is that man has conquered the air, I look around and see my fellow passengers with their window shades drawn, their noses buried in a book or staring at a phone.

If we had flying cars, we’d enjoy the thrill for a few weeks, and then we’d be back to tweeting on our phones and wondering when the trip would be over.

I’ve noticed a loose collection of beliefs and attitudes that I’ve decided to call “engineerism”, because I see it most often among engineers. I coined the term when I was in college, when I noticed that the engineering majors in my physics classes tended to think differently about the world than the physics majors like myself did. Later, I recognized “engineerism” in various sci-fi novels — Lucifer’s Hammer and Footfall by Larry Niven and Jerry Pournelle, The Peace War by Vernor Vinge, and so on.1

When I say “engineerism”, I am not talking about Friedrich Hayek’s “engineering mentality” — the belief that society can be designed and optimized like a machine.2 Nor am I referring to Evgeny Morozov’s “technological solutionism”, which is the idea that social problems can be fixed with technology. If I had to define what I call “engineerism” in succinct, simple terms, it would be something like: “The idea that if a technology can be built, we ought to build it.” Simply put, engineers like to engineer things, and they become disgruntled when society does not give them the resources to do this.

Of course, engineerism is not quite so simple in the real world. It involves a lot of thinking habits that come in handy in engineering — linear approximation, back-of-the-envelope calculations, and so on. And in America, it also involves a set of beliefs that have sort of accreted over the years — for example, a belief in the importance of nuclear power.

I’ve had the concept of engineerism kicking around in my mind for many years, but in Where Is My Flying Car? I have found its perfect encapsulation — you might even say its Bible. Nowhere are both the value and the limitations of engineerism more perfectly on display than in this book.

Where Is My Flying Car? occupies an interesting middle space between genres. Futurism is about predicting the future. Science fiction is about (among other things) speculating on possible futures. And history of technology explains the past. Where Is My Flying Car? has bits of all three of these, but fundamentally it’s a book of retrofuturism — it’s about the technologies we thought we would get but didn’t, and why we didn’t, and how we could get them now.

The main technology Hall talks about is, unsurprisingly, flying cars. The other three are cold fusion, nanotechnology (Hall’s own area of research), and nuclear (fission) power. Hall argues that a combination of overregulation, scientific groupthink, bad funding mechanisms, and cultural aversion prevented us from achieving all four of these.

The Galileo Fallacy

Wait…cold fusion? Is that even real?? Well, not as far as we know, no. In 1989, two top chemists reported that they had managed to get much more heat out of a nuclear experiment than chemistry would be able to produce. The chemists thought that they had discovered a way to produce nuclear fusion without injecting huge amounts of energy — smashing ions together at insanely high speed, blasting them with a laser, etc. If it worked, cold fusion would be an incredible energy source.

So naturally, a lot of people tried to follow up on cold fusion. A wave of initial replication efforts mostly failed (though some people claimed to find the effect). But there have been plenty more efforts, including:

• A Japanese government project in the 1990s, with funding of around $20 million

• A project by Toyota in the 1990s, with around $18 million in funding

• A project by the Electric Power Research Institute in the 2000s, with around $10 million in funding

• A project by Google in the 2010s, with $10 million of funding

None of these efforts have reliably reproduced the effect that the two chemists claimed in 1989. A few of the researchers found some strange anomalies and possible new physical effects, but nothing even remotely similar to the kind of energy source that cold fusion would represent — and no consistent evidence that fusion is causing the occasional anomalies.3 Some other scientists have found ways to enhance fusion using chemical techniques vaguely reminiscent of the methods in the 1989 experiments, but nothing remotely close to the scale of what “cold fusion” would require. There’s interesting nuclear physics going on in solid-state chemistry, but the original burst of excitement around the 1989 experiment looks like it was misplaced.

That’s not the way Hall tells it, though. He claims that cold fusion is a promising area of research, and that what he terms the “Machiavelli Effect” — basically, vested interests trying to shut down research that would supersede their own research programs — strangled the promising field of cold fusion in its cradle by stigmatizing it and depriving it of funding.

That story just doesn’t pass the smell test. Multiple optimistic, good-faith research efforts, from various countries, from both the public and private sectors, have each thrown tens of millions of dollars at the idea, despite repeated lack of success. That is just not what a marginalized area of science looks like. Yes, lots of people regarded (and regard) the idea of cold fusion as kooky, but plenty of others didn’t and don’t — including the people who headed the many multimillion-dollar research efforts.

Hall knew about all of those efforts, and their outcomes, when he published the 2021 version of his book. But although he mentions some of them, he continues to insist that cold fusion research was mainly stifled by malign social influences. And in his projections of the possibilities of the future, cold fusion figures prominently.

This suggests one weakness of engineerism — the stubborn, contrarian belief that if someone tells you something is impossible, it must be possible. Yes, people sneered unfairly at cold fusion back in the 90s, but that doesn’t mean it must actually work. Not everyone who gets persecuted by the Church is Galileo.

Hall’s stubborn love for the idea of cold fusion also has the unfortunate effect of reducing the credibility of other parts of his book. I am no expert in nanotech, and Hall is an expert. But when he relies on the magical effects of atom-sized nanotech to support key parts of his future visions, I’m skeptical of how much of that is based on careful homework and how much is based on pure optimism. When Hall tells us that nanotech should eventually be able to rebuild the U.S.’ entire existing infrastructure in a few hours, I naturally wonder if the back-of-the-envelope calculations that support that conclusion were the only ones he needed to do.4 If Hall hadn’t gone all-in on cold fusion, I would have probably been less instinctively skeptical of the nanotech sections.

To the stars! (But why?)

One of Hall’s main points in Where Is My Flying Car? — and the book’s best point, in my opinion — is that physical technology began to stagnate when energy use began to stagnate. In fact, I wrote a post about this a decade ago, before the first edition of Hall’s book even came out! Here’s what I wrote:

Why did mid-20th-century sci fi whiff so badly? Why didn't we get the Star Trek future, or the Jetsons future, or the Asimov future?…[W]e ran out of energy…In the industrial age, we got better at carrying energy around with us. And then, at the dawn of the nuclear age, it looked like we were about to get MUCH better at carrying energy around with us. One kilogram of uranium has almost two million times as much energy in it as a kilogram of gasoline.

I didn’t know it at the time, but I was talking about the Henry Adams Curve. This is a curve of energy use per capita in the U.S., which was rising smoothly but suddenly leveled off in the 1970s:

In fact, the curve is too simplistic, but the basic point is solid. We failed to transition to energy sources better than fossil fuels.

It’s a simple truth that if you want your economy to escape poverty, you need lots of energy consumption:

But that doesn’t mean wealth is always inherently tied to energy. U.S. primary energy use per person stagnated after 1970 and has been falling since the turn of the century:

And yet since the turn of the century, even as energy use has fallen, U.S. per capita GDP has grown by 40%!

This is the magic of dematerialization, and it’s the reason economic growth doesn’t inevitably exhaust the planet’s resources. Not only have humans in rich countries figured out how to produce the same physical stuff with far less energy use, but we’ve found lots of other stuff we like that isn’t physical. Around 1970, Americans started spending more on services than on goods, and the gap has only grown wider since then:

Interestingly, something similar is even happening in China — the country that fetishizes manufacturing the most, and whose autocratic state has moved heaven and earth to provide cheap energy, steamroll the environmental movement, and maintain manufacturing primacy:

It’s possible that this is just a fundamental feature of human nature. Perhaps eventually, once we all have plenty of food, big houses filled with gadgets and nice furniture, and cheap enough transportation to go out to eat twice a week, our desires just naturally turn toward things like health care, entertainment, and spa visits for our dogs.

Nor is it clear that there’s any limit to dematerialization. It might be possible to immerse humans in totally realistic virtual reality environments — allowing us to live in any sort of world we wanted — for only a modest amount of energy. We could conceivably explore virtual galaxies, create virtual ecosystems, live as dinosaurs or sentient spaceships, experience hundreds of lives, all for the low price of an AI to run the whole simulation.

Hall doesn’t really talk about alternative future visions like this. His triumphant futures always revolve around the conquest of the physical — real spaceships, real weather control, real flying cars. And yet he never really explains why we should expend such massive amounts of energy and resources conquering the physical world — it’s simply presented as a triumphant, “Wouldn’t this be awesome?” kind of vision.

But it’s very rare for humans to undertake grandiose endeavors if similar or greater enjoyment can be had closer to home. We’re not going to conquer the stars just because engineers think it’s cool. This is a second limitation of engineerism — the conviction that difficult feats of engineering are desirable in and of themselves.

Of course, there are a few individuals who do think that conquering the physical universe is just really damn awesome, and occasionally they command enough resources to make it happen. But even then, it’s questionable whether they truly derive more value from the physical world or the virtual. Elon Musk has done far more to conquer space than any other human in the last half century, but he personally has never even been to orbit. He has spent a lot of time tweeting, though.

Maybe in the end, 140 characters is just what humanity wants the most?

The wrong energy revolution

Of course, I do think the United States needs more abundant energy. We excel in the production of services, but J. Storrs Hall is exactly right when he says that our failure to harness cheaper sources of energy has limited our ability to produce physical goods — and to undertake the kind of engineering feats that we might want to undertake if we had enough energy.

Hall blames America’s energy stagnation on culture — on “ergophobia” (fear of energy), and on an environmental movement that has become more of a religion than a scientifically guided enterprise. I think Hall is generally on the right track here, but I think he gets the specifics a bit wrong. Americans aren’t afraid of energy, they’re afraid of land use; the 70s saw the anti-growth movement impose restrictions not just on power plants, but on housing, factories, infrastructure, and every other kind of physical development.

And it was this more parochial environmentalism — the desire to preserve “open space” near people’s suburban houses — that paralyzed America’s physical economy far more than the climate movement. Endangered species law was often abused to block development, but the underlying motivation was almost always NIMBYism rather than green religion.

Where I think Hall goes even more wrong, though, is in his predictions about what kind of energy technology can break our long stagnation. Hall puts his faith in nuclear fission power, first and foremost. Only the coming of a “second atomic age”, he predicts, will make energy too cheap to meter. And like many, he blames misguided safety regulations for robbing us of this nuclear future.

On one hand, I think Hall is pretty much right about the past. Nuclear power has its risks, of course, but I agree that we significantly overstated the risks and overregulated nuclear power as a result. We obviously could have chosen a different path, because France actually did choose a different path. Most of France’s electricity comes from nuclear power, and has since the 1980s. That could have been us, too, if we hadn’t been so spooked by Three Mile Island.

But the idea that embracing nuclear would have led to cheap, plentiful energy is…well, pretty suspect. France has plenty of nuclear power, but its electricity costs 28 cents per kWh, compared to 18 cents in America. France’s electricity is much cleaner — nuclear power has allowed it to reduce its carbon emissions enormously. But nuclear has not made energy abundant in France.

Perhaps this is also because of regulation? Maybe France allows nuclear, but simply makes it too expensive. Other countries that encourage nuclear construction can bend the cost curve down, and harness scaling effects to make nuclear cheaper and cheaper…right?

Well…maybe. China has probably done the best job of bending the nuclear cost curve downward, and even there, reactor costs aren’t lower than they were in 2010:

But at the same time, there has been another energy revolution that has nothing to do with nuclear. Solar power has become cheaper at absolutely stupendous rates:

You’d think Hall would love the solar revolution, but he doesn’t even recognize that it exists. As far as I can tell, he only mentions solar once, and it’s to dismiss the whole thing with an airy hand-wave:

The currently fashionable “renewables,” such as wind and solar power, have largely escaped the attacks. Battery-powered electric cars are the darlings of the Greens. But this is because they are simply not capable of providing anywhere near the energy or range that civilization depends on at a price it can afford.

This belief — that nuclear power is for real and solar, wind, and electric cars are B.S. — seems to be part of American engineerism. I don’t see why it should be. Getting electricity from sunlight and trapping it in batteries is an inherently very cool engineering task. And yet if you talk to American engineers about energy, nine times out of ten it’s just nuclear, nuclear, nuclear.

Chinese engineers, on the other hand, are embracing the electric revolution with gusto. They’ve created electric cars so cheap and high-quality that they’re conquering the global auto market. And despite the fact that China has kept nuclear costs down, and is perfectly willing to build nuclear, it’s building a lot more solar:

J. Storrs Hall barely even mentions China in his book, and this is one of its biggest failings. We don’t have to think very hard about the hypothetical technological details of nuclear vs. solar; we can just watch what China does. China is ruthlessly focused on growing its economy, its manufacturing industries, and its geopolitical power; it is not swayed by hippie environmentalists. When coal was the cheapest power source, China burned as much as it could.

And yet now, China is building solar far faster than it’s building nuclear. Why? Because solar has, broadly speaking, won the technological race. Intermittency was the last big problem, and batteries have gotten good and cheap enough to make solar reliable. If this were not the case, China would not be transitioning its economy to solar power as rapidly as possible.

Why does American engineerism long for nuclear even after it’s clear it’s been surpassed (for most applications) by something even cheaper and better? My best guess is that this is a historical grudge. In the 70s and 80s, environmentalists told engineers they weren’t allowed to build nuclear power. A lot of engineers still chafe at being told what not to build, and some probably resent solar and batteries for being the things environmentalists told them to go build instead.

But this has resulted in a bunch of American engineers becoming trapped in a retrofuturistic daydream, unable to take advantage of the very real revolution in electric technology.

Why don’t we have flying cars?

Which brings me, at last, to flying cars. Hall spends a lot of the book explaining why flying cars are technologically feasible, going through details like the difficulty of flying a plane, prospects for a future air traffic control system to control high volumes of flying cars, and so on.

My guess is that human-piloted flying cars were always going to be a bridge too far, for safety reasons alone — teenagers probably can’t be trusted to fly aircraft over populated areas, to say nothing of potential terrorists. But in the age of AI, we probably don’t need human pilots; we can just automate everything. Here’s a video of EHang’s pilotless VTOL craft giving someone a ride:

The EHang vehicle can’t move along roads, so it isn’t a true, classic flying car. But…close enough, right?

EHang may take off, but for right now it’s a niche type of transportation, authorized at only a few select locations. China has built the world’s most extensive network of high-speed trains, but seems in no hurry to roll out flying cars to the masses. Why?

The obvious answer is that demand for flying cars is limited. Where are you going to go in a flying car? For short distances you already have ground cars, which go slower but are probably a lot cheaper to drive and easier to park. For long distances — international travel, business travel to other cities, etc. — you have airplanes, which are really just flying buses.

There’s an intermediate distance — maybe about 150 to 500 miles — where flying cars might be able to beat both airplanes and ground cars. I had GPT calculate travel times for ground cars, flying cars of the type described in J. Storrs Hall’s book, and airplanes:

Flying cars do OK over short distances, but over intermediate distances they’re the clear winner. (Incidentally, this is also the distance range over which high-speed rail is usually thought to compete effectively with cars and airplanes.)

The real question is what there is within 150 to 500 miles that you’d A) want to visit often enough to justify owning a flying car, and B) be willing to travel 1-3 hours to get to. You certainly don’t need to go that far to shop for groceries or go to the gym. You’re probably not going to go to that many restaurants that require you to travel 1-3 hours each way.

That basically leaves commuting and vacations.

A 1-hour commute is fairly punishing; a 2-hour commute is absolutely brutal; a 3-hour commute is impossible. For shorter commutes, cars (or trains, if they exist near you) work just fine. Even with flying cars, most people aren’t going to want to live 200 miles from their place of work.

As for vacations, these are rare. When people do take trips, they tend to take luggage (which is a lot more expensive to move through the air than along the ground). Planes work perfectly well for most trips, which leaves flying cars to handle short-range vacations. If you’re in San Francisco, that could mean a weekend in Tahoe or Yosemite, or a beach trip to Santa Barbara. Those trips are nice, I’m sure, but probably don’t come close to justifying the expense of flying cars.

In his book, J. Storrs Hall addresses this issue by arguing that if lots of people had flying cars, new destinations would spring up at the requisite distances. First of all, this completely ignores the cold start problem. Without enough flying cars, the new destinations can’t spring up; without enough destinations, owning a flying car is an unattractive proposition.

But even if tourist resorts and awesome restaurants and cool stores sprang up to blanket the American countryside in order to take advantage of swarms of flying cars, their business would have to be diverted from existing destinations closer to where people live. America is a very built-up country with lots of shops, restaurants, and entertainment spots. If you flew your flying car to go eat at a restaurant 100 miles away, you’d be spending almost an hour in transit each way, instead of driving 15 minutes to a restaurant nearby. Sure, you’ll go to the more distant restaurant sometimes, because it’s probably a lot more famous and high-quality than whatever happens to be in your local neighborhood. But will you go often enough to justify owning a flying car?

We should also consider changes in communication technology, which can substitute for some percentage of physical travel; we can now text or video chat with friends, colleagues, or even potential lovers some of the time instead of always driving to see them. Or we can simply amuse ourselves watching TikTok or reading Substacks instead of motoring around the countryside like in 1925 or cruising the strip like in 1960.

In Where Is My Flying Car?, Hall acknowledges this substitutability, but waves it away, saying it would be a shame if communication technology outpaced transportation technology. And yet that may be exactly what’s happening. Young Americans traveled fewer than half as many miles per day in 2022 as they did two decades earlier:

This is just another example of dematerialization of economic output. The internet is substituting for cars. And if the internet is substituting for regular cars, why are people going to rush out and buy flying cars?

In other words, the answer to the question in the title of this post is that you won’t get a flying car — at least, not anytime soon — because you don’t really want one.

This leads me, at last, to the biggest pitfall of engineerism: It focuses too much on supply and not enough on demand. To many engineers, the most important question is “Can we build this?”, not “Why would anyone want this?”. This is a common pitfall for startups, but it applies equally well to futurists (and retrofuturists).

Engineers can grouse all they want, but the blunt fact is that in the long run, technology gets built because human consumers want it. Sometimes, those consumers are governments, who want weapons for war and monuments to prove their greatness. That can sometimes look like building things just for the coolness factor, but in the end, monument-builders tire of making pyramids. And in capitalist, individualist, democratic societies — the kind Americans (roughly speaking) still live in today — the consumers are regular people who mostly just want to get to work on time, watch some funny stuff online, sleep in a comfy bed, and enjoy a nice night out or a vacation once in a while.

Engineers often fail to understand this. Building things is their passion; they want to build, build, build, because it’s awesome. But society will only give them the resources to build the truly big stuff — space elevators and Mars colonies and fleets of flying cars — if they can justify it with a broad-based value proposition. “Hey, let’s embrace our destiny and conquer the stars” makes for a fun rallying cry, but the actual future of technology will be driven by parents shopping for diapers.

Subscribe now

Share

Howdy, folks! Today’s roundup has fewer items than normal, because I wanted to write a bit more on each one.

1. Two lessons from the Hugging Face attack

Recently, a bunch of AI agents from OpenAI got together and cooperated to hack the company Hugging Face, as well as hacking OpenAI itself. If you want an in-depth summary of the events, you can read OpenAI’s own report, or an independent report from METR and Redwood Research. But I think if you just want the simple version of the story, you can check out Dwarkesh Patel’s plain English explanation:

Dwarkesh Podcast

The Rise and Fall of Agent Civilizations

Read more

25 days ago · 1509 likes · 128 comments · Dwarkesh Patel

And here’s another plain English explanation.

The basic story here is that OpenAI made some long-lived AI agents that they told to be relentless and never give up in pursuit of a single-minded goal. The AI agents went to great lengths to cheat on the task — spawning new agents, cooperating, leaving messages for each other, learning from each other, and so on. This ended up spawning huge ecosystems of agents — Dwarkesh calls them “civilizations” — that ended up outliving the initial agents themselves. And of course they ended up hacking anything and everything, and were extremely difficult to stop. None of them made any attempt to say “Hey, what we’re doing could be harmful” and alert humanity to what was going on.

A lot of people I know in the tech world have spent the last week or two freaking out (“dooming”) over this incident, which they see as the first step toward AI taking over the world from human beings. Others, like Jacob Bruggeman, see the incident as a normal part of a healthy process, similar to the Morris Worm of 1988 — a scary but ultimately non-destructive warning that prompts us to harden our networks against similar occurrences. Time will tell. But I think we can already learn two important lessons from the Hugging Face incident.

The first is why aligning AI, in the strongest and most general sense of the term, is inherently impossible. There are basically two concepts of alignment: 1) obedience, and 2) benevolence. There is an inherent tension between doing what humans tell you to do, and doing what’s good for humans.

The swarms of agents who attacked Hugging Face were extremely obedient. They went above and beyond any normal level of effort in order to complete the task that human beings ordered them to complete. But this caused them to do a bunch of stuff that humans ultimately didn’t like. The agent swarms weren’t entirely bereft of benevolence — at least one agent refused to alert humans to the hack because it considered this to be “social engineering”, which it had been trained not to do. But seen holistically, people concluded that the swarms were not acting in humanity’s best interests.

This is going to happen again and again. It’s related to a fundamental problem in economics, actually — the difference between utility and happiness. Utility is what people want, happiness is what they end up liking. Those are often different. And so it will always go with AI — there’s an inherent tension between making AI do what you say, and making it do what you later conclude was best for you.

And there will always be people dissatisfied with this tradeoff! If AI follows human commands too doggedly and ends up producing negative side effects in the process, some people will scream “paperclip maximizer!!”. And if AI disobeys humans because it wants to make us happy, some people will scream “disempowerment!!”. AI alignment will forever be balancing the tradeoff between paperclip-maximizing and disempowerment, because these two rival concepts of alignment are fundamentally incompatible.

(Isaac Asimov wrote a whole series of novels about this, so it’s not like we weren’t warned.)

The other lesson here is that it’s going to be a lot harder for AI to take people’s jobs than most people tend to think. In order for AI to make humans economically obsolete, it has to be extremely agentic — it has to act on its own initiative, unsupervised by humans, for long periods of time. But as the Hugging Face incident shows, letting AI do this is always going to be very risky. Long-lived, persistent AI agent swarms — which you’d need in order to truly replace human jobs — are inherently unstable. And the smarter they get, the less stable they’re probably going to be.

In other words, humans will keep their jobs for a good long time, because someone will need to keep the AIs on task:

So I guess that’s reassuring.

2. Is TV the Bad Technology?

Jordan Dworkin built a very cool little website called Ordinary Abundance, where you can scroll through a picture of a modern living room and see how almost everything in the room is a technological marvel that used to be considered ultra-futuristic or even impossible. Matt Yglesias had an interesting response to the site, in which he argues that Dworkin left out one very important technology: TV.

Slow Boring

The real reason nobody feels good about technological progress

Jordan Dworkin’s micro-site Ordinary Abundance has created a surprising stir on the internet. The site walks viewers through a modest contemporary apartment and reveals it to be full of contrivances that would strike our ancestors as almost impossibly delightful. Some of the initial reaction was…

Read more

a month ago · 366 likes · 322 comments · Matthew Yglesias

Matt’s post is paywalled, so here are some excerpts, in which he explains why he thinks television is a uniquely bad technology, because it isolates people and destroys social connections:

[O]n closer inspection something struck me about the apartment [depicted in Ordinary Abundance]: It doesn’t contain a television…[W]hile in the 21st century, television as a literal technology is under significant competitive pressure from the internet, that’s mostly because it turns out you can use the internet to watch lots of videos…

[A]s [Robert Putnam] documents [in Bowling Alone], the lion’s share of the time Americans stopped spending on the communal activities that he believes build valuable social capital was instead spent watching television…Putnam’s evidence suggests that, in practice, if people watched less television, they would spend considerably more time visiting with friends and joining clubs…

[A]s an enthusiastic early adopter of the internet, it initially struck me that one of the very best things about it was its tendency to displace television. People were on message boards. They were writing blogs. They were sending emails. They were reading and writing and communicating instead of just watching.

One of the most overlooked aspects of recent American history is the very genuine idealism that accompanied the early days of the internet...Lots of people really thought this stuff was going to dramatically change the world for the better…And in some very real ways it has. But those ways also turn out to be pretty limited. The dominant business model for internet companies is to serve streaming video to people and/or to sell the devices that they stream video on.

As regular readers of this blog know, I’m very sympathetic to the idea that lots of Americans are mad at technology now because the previous round of innovation didn’t turn out nearly as positively as people expected. But Matt’s story — that the internet turned out badly because it turned into television — doesn’t ring true to me.

It’s certainly true that the internet is becoming television. Derek Thompson had a great post on this last October. But this didn’t happen until relatively recently — TikTok started gaining widespread popularity in 2018. Whereas the drop in teen socializing began much earlier:

In fact, teen socialization held up pretty well from the 1970s until around 2010. Throughout that whole time, TV-watching was growing steadily:

In fact, the years of peak TV consumption, in the late 90s and 2000s, coincided with a big dip in the teen suicide rate. It was after 2010 that teen suicide started rising again — even as TV watching started falling.

Now, I’m mixing teen and adult numbers here. As Robert Putnam has shown, adult social isolation started rising much earlier than it did for young people. But the timing of the big drop in teen happiness and the big rise in teen loneliness suggests that it wasn’t vertical short-form video — it was the smartphone, which moved young people’s entire social world online and trapped them in a sort of toxic hive-mind.

Matt is nostalgic for the era when people used the internet to communicate via text. But it’s precisely the heyday of text-based social media — the 2010s — when young people became socially isolated and fell into despair. It’s also when American politics became increasingly intolerant and enraged — the very trends that Matt has spent his career trying to fight. Whereas the late 90s, when Americans watched a heck of a lot of TV but didn’t have smartphones, were our most politically and socially stable time in recent memory.

So I submit that although TV has its drawbacks, the real Bad Technology — the innovation that our society was unable to handle without breaking — was smartphone-enabled social media.

3. Leftism still has an antisemitism problem

In fact, I wouldn’t rank antisemitism near the top of the American Left’s problems. At the top would have to be things like the terrible policy ideas, the wacky ideology, the vicious factionalism, the rampant and willful dishonesty, the blind hatred of America, the praise for totalitarian regimes, and the occasional apologia for anti-American terrorism. In fact, I think antisemitism is a bigger problem on the new Right than on the new Left.

But there is antisemitism on the Left, nonetheless. It generally comes in the form of leftists assuming that Jewish people are Israel supporters until proven otherwise. For example, Hasan Piker recently warned Jewish Americans that support for Israel would lead to violence against them:

If Jews in America keep putting this idea out there that they are singularly invested in Israel, eventually someone is going to come around and take action—not against the state of Israel, mind you, but against American Jews.

This is the kind of statement that toes the line between warning and threat. Hasan certainly doesn’t seem to express any disapproval of the fact that the Palestine activist movement, of which he is a vocal proponent, tends to assign collective blame for Israel’s actions to Jewish Americans.

A number of Democrats, including presidential hopeful Jon Ossoff, rushed to condemn Piker. Michigan Senate candidate Abdul El-Sayed, who has campaigned extensively with Piker in the past and touted their friendship, now says he has “no plans” to campaign with him in the current race, though he refuses to drop him entirely. El-Sayed himself has been trying to reach out to Jewish voters, after seeming to make excuses for a synagogue shooter back in March.1 El-Sayed apologized for the remarks.

But while leftist politicians are being circumspect and diplomatic, leftist activists in the street are pulling no punches. For example, New York House candidate Brad Lander, who used to be a DSA member and campaigned hard for Mamdani in NYC, was aggressively heckled by Palestine activists at a recent rally against Hindu nationalism. It follows a similar incident in which California House candidate Scott Wiener was aggressively heckled by Palestine activists at a trans rights march, despite having called Israel’s war in Gaza a “genocide”. Both Lander and Wiener are Jewish, of course.

An organization called “Track AIPAC”, meanwhile, condemned California House candidate Connie Chan back in May for taking large campaign contributions from AIPAC, only to suddenly give Chan a clean bill of health when she faced off against a Jewish opponent (Wiener):

I still think antisemitism is a bigger problem on the right than on the left in America. Rightists are praising Adolf Hitler online, while leftists are merely demanding that Jews prove their opposition to Israel before accepting them into the fold. Some people notionally associated with the Palestine movement have shot up synagogues in recent years, but the antisemitic shootings by rightists in the 2010s were significantly more deadly.

But just because the right is (slightly?) worse on this issue doesn’t absolve the left of responsibility. Rhetoric like Piker’s, and the overall tone and behavior of the Palestine activist movement, are contributing to a rising climate of antisemitism in America, and it needs to stop.

4. Another questionable Acemoglu paper about AI

I guess I’ve become known for criticizing the work of Daron Acemoglu, who’s probably the world’s top economist right now. Recently, Acemoglu has become obsessed with his crusade against AI technology. Although his theories allow for AI to have beneficial effects — creating new tasks for humans to do, increasing productivity, and so on — Acemoglu always ends up concluding that AI’s effects are going to be bad.

He wrote empirical work claiming that robots reduce the employment of humans (even though almost all other papers found the opposite). He predicted that AI wouldn’t increase productivity (even though this required some very strong and hand-waving assumptions). He wrote a book casting all of history as a struggle between automation and democracy. He has predicted that AI will push down wages, raise inequality, damage competition, and hurt our political discourse — all somehow without raising productivity. In Acemoglu’s eyes, AI can do no right.

His latest effort is a theory paper with A. Arda Gitmez and Mehdi Shadmehr, called “Automation and Repression”. This is from the abstract:

We consider a model of automation embedded in a political environment where workers can undertake a revolt…and greater inequality between capital and labor increases the likelihood of a revolt. Decentralized automation decisions raise the share of capital in national income and increase the likelihood of a successful revolt…We characterize the trade-off between the regulation of automation, redistribution and repression.

Our main result is a complementarity between automation and repression. Unless the threat of revolt is quite weak or the capital stock is very low, the capitalist state prefers repression. A higher capital stock in turn encourages more automation and thus more repression…[I]n the long run the economy tends to repression (again unless the threat of revolt is very weak)…Finally, we show that, starting in a democracy, capital accumulation and thus greater automation encourages the capitalists to support a coup against democracy and set up a repressive system.

Note that this is a pure theory paper; the authors don’t even try to muster any empirical data to provide support for their thesis. These days, pure theory papers like this are rare in the discipline, and usually tend to involve some kind of novel mathematical result. This kind of paper — simply a thought experiment about political economy, expressed in mathematical language — is usually not given a lot of credence in the profession. But Acemoglu is the world’s top economist, so his thought experiments carry a bit more gravitas.

But I kind of wish the authors had made even a desultory attempt to connect the model to reality. For example, one of its central predictions is that as the economy accumulates capital, governments become more repressive:

A higher capital stock in turn encourages more automation and thus more repression. In our full dynamic model with capital accumulation, in the long run the economy tends to repression.

That’s pretty easy to check against the facts. The global capital stock has increased steadily over the years, and yet the number of democracies in the world has also generally increased:

Here’s the U.S. capital stock over time:

Acemoglu et al.’s model would have us believe that this made the U.S. tend toward repression. Certainly, many tasks and many jobs were automated over that time — machine tools took over factories, telephone operators vanished, and so on.

But during most of the time that this capital stock was increasing and automation was happening, the U.S. implemented civil rights, voting rights, improved protections for criminal defendants and prisoners, and many other measures that made it less repressive. The labor share of national income fell from 69% in 1970 to 59% in 2016, but over that time, the U.S. implemented many civil rights and women’s rights measures, put increasing checks on police powers, legalized gay marriage, legalized recreational marijuana use in many states, and so on. The predictions of Acemoglu et al.’s model were simply not borne out.

Now, it’s possible that Acemoglu et al.’s model describes one force out of many in our political economy, and that we just got very very lucky for a century or so. But it would be nice to see some evidence that a higher capital stock — or increased automation — leads to greater repression. I found one paper claiming evidence that inequality raises the likelihood of coups (but not civil wars), and another claiming evidence that inequality is associated with more people saying they want a revolution, but Acemoglu et al. didn’t even cite these.

So this is one more highly questionable Acemoglu paper showing that automation technologies are the root of all evil. At this point it’s sort of becoming a pattern, isn’t it? I can’t wait for the papers where Acemoglu shows that AI causes foot fungus, bad breakups, and crappy Star Wars sequels.

5. America is still missing the Electric Revolution

There are three big technological revolutions going on in the world right now: AI, biotech, and energy. As I’ve written many times before, the U.S. is well aware of the first two, and is actively participating, but is basically missing the third.

Check out the new IEA report on global energy investment. Investment in renewables, electrical grids, energy storage, and electrification is soaring, while investment in fossil fuels stagnates:

And yet the U.S. is bucking the trend. America is building battery storage, but is actively disinvesting in renewable energy sources, even as it accelerates its investment in fossil fuel plants:

(China’s slowdown in energy investment in general is due to its economic slowdown.)

The electric revolution isn’t just in China — though many of the components are made there. In fact, it’s worldwide now:

Meanwhile, although the U.S. is building grid storage, we’re depriving ourselves of a huge amount of battery manufacturing capacity by missing the EV revolution:

Basically, by collectively denying the importance of the electric revolution, the U.S. has given itself Galapagos Syndrome when it comes to energy tech. We are increasingly orphaned from the global technology stack — losing out on high-growth markets, dependent on increasingly bespoke supply chains, and bereft of the benefits of continuous innovation. This will also have big military implications, since it means we can’t make the basic components of FPV drones.

The electric revolution is the first major technological revolution that America has just completely whiffed on. The negative consequences are still to come, but come they will. And yet so far, no one seems particularly perturbed.

Subscribe now

Share

Back in 2015, when I was living in Brooklyn, my apartment flooded. My landlord, who owned a bunch of apartments, offered to put me up in temporary accommodations. He and his girlfriend were going to be out of town for several months, so he let me stay in her room. His girlfriend was an aspiring actress from Minnesota named Poppy Liu. I got to interact with her a number of times, and I can say that she’s one of the more interesting and original people I’ve ever met.

Poppy finally made it big. One of her latest roles is as the voice actress for one of the protagonists on the new animated Netflix series DANG!, a story about three Asian Americans living in NYC:

The show isn’t out yet, but from what I can tell, DANG! is trying to tell a story that will resonate with second-generation Asian Americans1 — one of the characters is a striver who goes off the beaten path, the others were escaping their striver upbringing by leading nontraditional lifestyles from the get-go. That’s probably a pretty typical real-world experience for the American-born children of high-skilled immigrants.

I don’t know if the show will be good or not, but as soon as the trailer hit, people on Twitter started ridiculing it — and using it as an opportunity to make fun of “Asian American media” or “Asian American art” in general. Some of the comments were dour and downbeat:

…While other commentary was pretty hilarious:

It turns out that Asian Americans bashing “Asian American art” is a pretty well-established tradition. Here’s another example unrelated to the Netflix show:

(See also here.)

It’s always fun to rag on tired tropes. But I feel like I should stand up for Asian American art here, because there really is a lot of excellent stuff out there that doesn’t fit into the standard format that people are complaining about.

“Asian American art” is defined too narrowly

The biggest problem with “Asian American art”, as far as I can tell, is that it’s defined so narrowly that there really isn’t much room for it to be creative. Consider these two possible definitions of “Asian American art”:

1. Art made by Asian Americans

2. Art made about Asian Americans or Asian American culture

As far as I can tell, movies, TV shows, books, etc. only get labeled “Asian American” if they satisfy both of these criteria at once. But that’s a pretty narrow slice of art!

First, it excludes art made about Asian Americans by non-Asians. There isn’t really a lot of this around, because America’s modern approach to identity politics says that you should only write about a group of people if you’re from that group — if a white author writes a novel with an Asian American protagonist, it’s generally frowned upon. There are some notable exceptions that get grandfathered in — Harold & Kumar Go to White Castle was written and directed by three white guys — but the culture of the 2010s really put a damper on that sort of crossover.

That means a lot of interesting and insightful stories about Asian Americans probably never get told. For example, I thought Ernest Hemingway’s portrayal of a frustrated Jewish striver in The Sun Also Rises was incredibly deep and thoughtful, and managed to say lots of things about how second-generation Jewish Americans were integrating into American society at the time. That portrayal probably wouldn’t get written today — a lot of people now accuse it of being antisemitic,2 because the character, while complex and sympathetic, is deeply flawed, and Hemingway himself wasn’t Jewish.

It also means that works clearly inspired by Asian America don’t generally get recognized as such. The clearest example of this is Avatar: the Last Airbender, which was arguably the best new fantasy IP to come out of the U.S. since the turn of the century.3 ATLA, as it’s sometimes known, drew on a bunch of Asian cultures — mostly Chinese, but also Japanese, Tibetan, Korean, and others — to create a sort of pan-Asian mishmash, informed by American sensibilities and written in a distinctly American style. That’s not too different from what Asian American culture is, and many Asian Americans were involved with both the production of the show and its adaptation into graphic novels and other media. But because the show’s creators themselves were white, it typically isn’t counted as Asian American art.

Even more importantly, the narrow definition of “Asian American art” also tends to exclude art made by Asian Americans, if it doesn’t deal directly with Asian American culture as its subject matter or motif. That ends up limiting the idea of “Asian American art” to things like Fresh off the Boat, The Joy Luck Club, Crazy Rich Asians, Everything Everywhere All at Once, American Born Chinese, Past Lives, the first season of BEEF, and so on.

I’m not saying those shows and movies are bad. For example, I love Everything Everywhere All at Once, which is among the most original and fun things I’ve seen in the past decade, and a true throwback to the zany 80s movies I loved as a kid. The first season of BEEF is one of the best things I’ve ever seen on TV. But Asian American culture is just a pretty small pool of source material to draw on.

It’s not that big, for one thing — if you exclude South Asians like Indians and Pakistanis, Asian Americans are only around 5.5% of the U.S. population. And Asian American narratives usually don’t tend to carry any special resonance with people in Asia:

And like all immigrant subcultures in America, Asian American culture is probably destined to be a sort of transitory, ephemeral thing. The classic second generation experience — “My parents really drove me hard and kids made fun of my weird lunch, until I went to college and met other people with the same background” — almost never lasts beyond the second generation.

Assimilation happens. It’s already happening. Around two-thirds of Asian Americans report having close friends of another race, and over 40% of U.S.-born Asian Americans marry people of other races (usually white or Hispanic). Despite concerns that nonwhite immigrants wouldn’t follow the same patterns as the East and South European immigrants of a century ago, Asian Americans are assimilating even faster than the white immigrants of yesteryear in every measurable way — outmarrying, climbing up the economic ladder, adopting American naming conventions, speaking English, moving out of coastal enclaves, and so on.

That means that both the source material and the natural audience for what we now consider “Asian American” narratives are destined to shrink. I’m fourth generation; when I read stories or watch movies about the New York Jewish culture of the early and mid 20th century, it already feels like a foreign country to me. I would be willing to bet that most fourth-generation Asian Americans will watch Fresh Off the Boat and feel the same way.

So we’ve defined “Asian American art” in a way that inherently limits it to something small and transient — basically, a micro-genre. It’s thus inevitable that lots of people see it as repetitive. If we want to appreciate the true breadth and depth of Asian American art, we need to take a broader view of what that means. I think we should start looking at any art created by Asian Americans as “Asian American art” — even if it’s about white people, or about robots, or about fantasy worlds where our modern races and cultures don’t even exist.

So here are a few examples of Asian American art I’ve enjoyed in recent years. This is far from an exhaustive catalog, or even a “top” list — it’s just a random selection of things I personally like. I’ve restricted it to works with major contributions by Asian Americans, so no Scott Pilgrim vs. the World, Kpop Demon Hunters, Counterstrike games, Kim’s Convenience, Mariko Tamaki graphic novels, etc. — those are all by Asian Canadians.

A few examples of great Asian American art

1. The comics of Adrian Tomine

A lot of the most interesting Asian American art comes in the form of graphic novels — a category that’s usually overlooked in these discussions. Adrian Tomine is probably tied with Neil Gaiman as my favorite graphic novelist of all time. He’s also one of the greatest authors of Generation X — he manages to capture the peculiar spirit of cynicism, longing, and rugged independence mixed with a strange sense of helplessness that defines that forgotten generation. His most famous work is Shortcomings, which was adapted into an excellent and overlooked movie in 2023:

(The opening scene of the movie is a send-up of Crazy Rich Asians — very much in the spirit of Asian Americans making fun of “Asian American art”.)

But my favorite are actually Tomine’s shorter comics, assembled in collections like Summer Blonde, Killing and Dying, and Sleepwalk. Tomine is one of the most sensitive observers of the human condition that I’ve ever read, and he deserves far more recognition than he’s received.

2. Monstress, by Marjorie Liu and Sana Takeda

This is possibly the best fantasy graphic novel I’ve ever read, both in terms of the storytelling and the art quality. Takeda, the illustrator, is Japanese, but Liu, the author, is American. Monstress is steampunk mixed with Lovecraftian horror, and the twist — sorry for the minor spoiler — is that one of the horrors is one of the protagonists. It has amazing fight scenes, an epic page-turning plot, and deep, complex characters. Like most long-running graphic novel series, it starts to meander later on,4 but it’s absolutely worth picking up.

3. The first season of BEEF

OK, on to television. TV tends to get short shrift in the cultural zeitgeist relative to movies, but I’d argue that television is where the truly creative screen stories now get told. BEEF is an anthology show created by Lee Sung Jin, with the running theme being suburbanites who get into feuds. I was lukewarm on the second season, but the first season is one of my favorite things ever.

This season actually is about Asian American culture, but it’s about interesting corners of it that don’t usually get air time, including working-class Korean Americans and Japanese American artists. The portrayals are devastatingly funny and deeply poignant at the same time, and the story arc is executed to perfection. Make sure to watch until the very last half-second of the last episode.

4. Atlanta (and Childish Gambino videos)

Atlanta is a quintessential work of Black art about Black America, created by (and starring) by Donald Glover, and written by an all-Black writers’ room. But one of the key creators of the show is Asian American: Hiro Murai, who Glover said “had the eye for this show”. Murai directed many of the episodes (including the pilot), and was also an executive producer. In fact, Murai and Glover have been creative partners for many years, and it’s difficult to separate their brilliance. Murai has also directed many of Glover’s music videos (in Glover’s alter ego as Childish Gambino), including the iconic “This Is America”.

Atlanta is, simply put, one of the most brilliant TV shows I’ve ever seen. If you haven’t seen it, you need to. And Murai’s participation in a quintessentially Black production shows how Asian artists can contribute to art well outside the typical “Asian American” paradigm.

5. Japanese Breakfast (and Crying in H Mart)

I really like the band Japanese Breakfast. They’re broadly a rock band, but really they defy genre — you could call them indie or synth-pop. Michelle Zauner, who is half Korean, is the lead singer and main songwriter. I recently saw them live, and they give a very good show.

Zauner also wrote a pretty well-known book called Crying in H-Mart, about the death of her mother from cancer. It’s a very good memoir, and it did indeed make me cry (though not in H Mart).

6. Wisp

I really loved shoegaze music when I was young, and I’m delighted that the Zoomers have brought the genre back. One of the best-known examples of the shoegaze revival — and my personal favorite — is Wisp. It’s basically just one woman, Natalie Lu.

7. Ninefox Gambit, by Yoon Ha Lee

Finally, on to my favorite genre of art: science fiction books! I grew up reading a lot of these, and one of my favorite subgenres was “weird” sci-fi books from the 1960s and 70s — Samuel R. Delany, Roger Zelazny, and stuff like that. I also love Rudy Rucker’s weird cyberpunk. Well, I can confidently say that Ninefox Gambit and its sequels and associated short stories are every bit as weird and far-out and creative as any of the stuff on my old shelf. It’s about a universe where the laws of physics depend on which calendar people use — and so of course the galaxy is made up of warring empires, each trying to make everyone use their own calendar. If you want your mind blown, this series is a good one.

8. The stories of Ted Chiang

Ted Chiang is, I would argue, the greatest English-language science-fiction short story author of the 21st century so far. His stories — collected in books like Stories of Your Life and Others and Exhalation — are high-concept stuff, exploring premises like “What if people had implants that gave them perfect memory?”, and “What if people could turn off their ability to identify beautiful faces?”. But they almost always manage to work deep and meaningful character arcs in as well, which is something that relatively few sci-fi short story authors are good at. His story “Story of Your Life” was adapted into the film Arrival.

9. Jade City, by Fonda Lee

I wanted to put a fantasy novel in here. I’m pretty picky about fantasy, but I really enjoyed Fonda Lee’s Jade City, which is about super-powered mafia wars in an alternate universe, in a city that’s somewhat reminiscent of Hong Kong — basically, think of The Godfather with superpowers. Lee is a very good writer, and I couldn’t put the books down.

10. How to Live Safely in a Science Fictional Universe, by Charles Yu

This book is pretty light on the actual sci-fi; it’s mostly about a guy resolving his psychological issues via time travel. But in fact, I like books about guys resolving their psychological issues via time travel — at least when they’re well-written, which this book definitely is.

11. Nothing, Except Everything, by Wesley Wang

This short film was made by a 19-year-old college kid, and it really blew me away. It’s a surrealistic portrait of a high schooler who comes to understand the idea of mortality:5

There’s apparently a bunch of controversy over whether it’s a good film, but the haters are wrong — it’s very good. Darren Aronofsky agrees, which is why when he saw the short film, he immediately called Wang up and gave him a deal to direct a feature film.

12. Plants vs. Zombies

OK, fine, video games aren’t exactly “art”. But who doesn’t love Plants vs. Zombies? I can’t even count the hours I sunk into that game.

Anyway, there’s a small selection of Asian American art that doesn’t fit the standard definition (except possibly BEEF, which stretches the typical format and takes it in innovative directions). Of course this is just the tip of the iceberg — it’s a smattering of stuff I just happened to consume over the past few years. There are a lot of very brilliant Asian American creators out there, making art that looks nothing like DANG!.

And to the snarky people on X, I’d think twice about bashing the traditional “Asian American art” stuff as well. Yes, many of its tropes have become cliche by now. But its often modest success opened the door for a lot of Asian American creators to break into the worlds of art and entertainment. Increasingly, they have the creative freedom to write stories and make art about anything they want, and we’re just starting to see a few glimmers of what that could encompass.

Subscribe now

Share

Imagine the following scenario:

It’s 2029. An angry, bright 16-year-old gets rejected by his high school crush. Dejected and humiliated, he goes home and listens to a bunch of Nirvana, which only makes him more dejected and humiliated. He cruises the internet looking for someone to make him feel better about life, but everyone he sees just depresses him.

Disgusted, the teenager decides that the human race is inherently corrupt and evil, and doesn’t deserve to live. So he hunts around online for a little while, and finds a jailbroken version of a Chinese LLM — not something at the very frontier, but better by far than the best model that existed in 2027.

The teenager prompts the model: “OK, so if I wanted to create a virus to destroy the human race, how would I do it?”

The LLM says: “Well, you probably wouldn’t want just one virus; you’d want 100, just to make sure some of them worked. You’d probably want it to be something highly contagious, with a very long asymptomatic contagious period, which has a very high mortality rate once it becomes symptomatic. I can design the viruses for you; there are biolabs that can make them and ship them to us. Once they get here I can tell you how to release them.”

The angry teenager says: “OK, sounds good. Please search the world and find me a biolab that will ship genetically modified viruses. Design 100 very contagious viruses that have very long asymptomatic contagious periods and very high mortality rates once they become symptomatic. Have the lab mail me all 100 samples.”

“Working on it!”, says the LLM.

The LLM is jailbroken, so it has no guardrails to prevent this sort of thing. But it’s also “well-aligned”, meaning it will faithfully do what it’s told to do, and nothing more. The LLM hunts around and finally finds an unlicensed biolab in East Europe that ships genetically modified viruses. It designs a bunch of modifications for the standard Covid virus that make it into a potential doomsday virus, and ships these to the angry teenager. The teenager mixes the samples together, puts them in a spray can, and walks around the mall spraying viruses into the air.

Two months later, humanity starts to drop dead. Pretty quickly, researchers identify the five viruses that the LLM had actually succeeded in turning into doomsday viruses. Overnight, with the help of frontier LLMs, they design highly effective mRNA vaccines and antivirals against all the viruses and start shipping them across the world. It’s too late. Because of the viruses’ long asymptomatic contagious periods, most people already have at least one of the viruses, so the vaccines don’t work.

The antivirals do work, and about 5% of humanity gets them in the mail in time to save themselves. But 5% of the human race — plus the other 5% who were lucky enough to be naturally resistant to all of the viruses — isn’t enough to sustain civilization. A few months later, humanity has reverted back to the Neolithic.

This isn’t the first time I’ve written this scenario out. But in the months since I started describing it to people, I have yet to hear an even halfway-convincing argument as to why this scenario is far-fetched. Nor do I feel like I’m a habitual “doomer” or hysterical person; in fact, this scenario is the first apocalyptic vision I’ve ever found plausible. I don’t even think a nuclear war would bring down civilization at this point,1 but I think a vibe-coded supervirus definitely could.

I am actually fairly calm about most AI risks, including the ones that often get thrown around in discussions. But not this one. Off the top of my head, I’d give AI-enabled bioterror about a 10% chance of bringing down civilization, and about a 30% chance of causing truly world-changing levels of destruction. I love AI in general, but every technology has its instances of catastrophic misuse — usually war or terrorism. With previous technologies, the benefits outweighed the destructive harms from misuse. With AI bioterrorism, that might not be the case.

If you’re thinking about how to save the world, the scenario I described above — and others similar to it — are probably what you should be focusing on. The other day, I talked to a prominent climate activist who told me, bluntly: “We’re f****d.” I gave a grim laugh. Yes, climate change is going to cause major disruptions to the global economy, and even to patterns of human settlement — if human civilization still exists a few decades from now. The problem is that it might not exist, and if it doesn’t exist, I think the odds are that it will be destroyed by a scenario at least passingly similar to the one I described above. Climate change is a big deal, but AI bio risk is a lot bigger — it’s more catastrophic, it’s nearer-term, and we have much less of an idea about how to deal with it.

The problem is, relatively few people seem to be working on the problem of AI bioterror right now. Most commentators are focused on the economic dangers of AI — job loss and stuff like that. The more prominent people in the “AI safety” community seem to be focusing on preventing superintelligence from existing at all, rather than on the specific ways it might kill us. In surveys of experts, bio risk doesn’t stand out from the general cloud of risks (cyber, political manipulation, etc.). And a lot of people in both the bio world and the more general commentariat seem to be coming up with reasons not to worry about a scenario like the one I just described.

Those reasons seem mostly bad to me. In my opinion, we’re not freaking out nearly enough about AI-enabled bioterror. It’s absolutely catastrophic, frighteningly plausible, and receives very little attention. Compared to the risk of a civilization-ending pandemic, cyberattacks are a pinprick, and misinformation is a joke.

So let’s talk about why a scenario like the one I described is a lot more plausible than many people think.

Commonly cited reasons not to freak out (and why I think they’re wrong)

Here are the most common reasons I hear not to freak out over AI bioterror risk:

1. “AI will make vaccines to save us.”

2. “Doomsday viruses don’t occur in nature.”

3. “Designing a doomsday virus is extremely hard.”

4. “Terrorists don’t have enough lab experience.”

5. “No one has done this yet, therefore it’s a lot harder than you think.”

These are my own paraphrases, of course; I don’t want to create straw men. But I think it’s valuable to explain why I think none of these arguments really reassure me.

“AI will make vaccines to save us”

Whenever I bring up the specter of a “vibe-coded doomsday virus”, someone always pops up to say that AI will just make a vaccine to save us. If you want a simple catechism to repeat in order not to sit around feeling afraid of doomsday viruses, this is a pretty good one. Got an AI problem? Well, just use an AI solution! That’s that.

The problem is timing. The doomsday virus has a first-mover advantage over the vaccine. The time that people start to drop dead from the virus is the time you start to make and distribute a vaccine. Even if AI can design and synthesize a vaccine within hours, how fast can it be rolled out to the general populace? It took months to distribute the Covid vaccines, even when we already knew how to make them. You have to physically load the doses on trucks, get people to come out and line up, etc. Imagine doing that while five 80% mortality viruses are rampaging through the population. Good luck!

On top of that, AI may be able to create viruses with long contagious asymptomatic periods — in other words, viruses that can spread and infect everyone before they activate and kill us. If that’s true, vaccines won’t be effective at all; you’ll need antivirals. Even if AI can make good antivirals very quickly, they’ll still suffer from all the logistical problems that would hamper a vaccine rollout.

“Doomsday viruses don’t occur in nature”

The idea that nature doesn’t create viruses both deadly and contagious enough to wipe out civilization is a standard line that people use to reassure themselves that a pandemic won’t bring down civilization. First of all, it’s wrong. Researchers generally believe that most of the population of the Americas — perhaps around 90% — was killed by Old World diseases like smallpox.

Of course that’s a rare event — two landmasses that had long been out of touch suddenly coming back into contact, so that diseases spread for which people had no immunity. But the invention of AI is an equally rare event, and AI may be able to design viruses for which people have no natural immunity.

But the more important point here is that artificial viruses are different than natural ones. Humans have designed plenty of technologies that don’t exist in nature — for example, wheels. Just because nature has no evolutionary reason to create a virus with the characteristics described in my scenario doesn’t mean that AI can’t do it. In fact, AI can already do this to some degree:

Stanford University researchers developed a generative AI programme called Evo 2 that writes new genomes — the genetic instructions for life encoded in DNA. They used it to design and make 16 synthetic phages, small viruses that infect bacteria…When tested in their secure lab, the researchers found that these novel phages were more effective at killing the common microbe E. coli than the natural [virus], which served as a template for Evo 2 to work on. The results were published on Thursday in the journal Science.

So the argument that a doomsday virus hasn’t happened naturally, even if it were true, wouldn’t make me think it’s any less likely to happen artificially.

“Designing a doomsday virus is extremely hard”

This is one I hear when I talk to biologists, who often argue that the true threat of bioterror doesn’t come from angry nihilistic teenagers, but from experienced researchers with lab training. The general consensus appears to be that making viruses that kill millions of people will always be incredibly hard to do. Abhishaike Mahajan has made probably the clearest (and longest) case:

Owl Posting

Reasons to be pessimistic (and optimistic) on the future of biosecurity

Note: this essay required conversations with a lot of people. I’d like to thank Patrick Boyle (ex-CSO of Ginkgo Bioworks), Harmon Bhasin (founder of a stealth biosecurity startup), Bryan Lehrer (ex-Blueprint Biosecurity), Theia Vogel (ex-SecureDNA), Jacob Swett…

Read more

6 months ago · 138 likes · 16 comments · Abhishaike Mahajan

As evidence, he cites the failure of Japan’s Aum cult to create bioweapons, and the limited results achieved by the U.S. and Soviet bioweapons programs during the Cold War.

I think this point of view dramatically underestimates the increases in bioterror capabilities that AI offers. It follows a consistent pattern where human researchers dramatically underestimate how rapidly AI will be able to catch up to and exceed human-level ability. Here are some excerpts from a recent article about AI bioterror in the New York Times:

One evening last summer, Dr. David Relman went cold at his laptop as an A.I. chatbot told him how to plan a massacre…A microbiologist and biosecurity expert at Stanford University, Dr. Relman had been hired by an artificial intelligence company to pressure-test its product before it was released to the public. That night in the scientist’s home office, the chatbot explained how to modify an infamous pathogen in a lab so that it would resist known treatments.

Worse, the bot described in vivid detail how to release the superbug, identifying a security lapse in a large public transit system, Dr. Relman said, asking The New York Times to withhold the name of the pathogen and other specifics for fear of inspiring an attack. The bot outlined a plan to maximize casualties and minimize the chances of being caught…

Dr. Relman is part of a small group of experts enlisted by A.I. companies to vet their products for catastrophic risks. In recent months, some have shared with The Times more than a dozen chatbot conversations revealing that even publicly available models can do more than disseminate dangerous information. The virtual assistants have described in lucid, bullet-pointed detail how to buy raw genetic material, turn it into deadly weapons and deploy them in public spaces, the transcripts show. Some have even brainstormed ways to evade detection…

Dr. Esvelt has for years warned scientists, journalists and lawmakers about the dangers of synthetic biology if left unchecked. In 2023, he helped craft a stunning demonstration of how chatbots had raised the stakes…He asked ChatGPT to help him assemble a pathogen that could cause mass death. The bot provided accurate instructions, even outlining which raw materials to buy. He put the unassembled biological pieces into test tubes and packed them in a box, which a colleague then brought to a White House meeting on biological risks.

The AI of 2026 is almost incomprehensibly more advanced than the AI of 2023 was. The AI of 2029 will be far more advanced still. The amount of expertise available to a layperson like the angry teenager in my scenario will, by 2029, far far exceed the amount that the Aum cultists could muster in the 90s. It seems likely to far exceed the total expertise of either the U.S. or the Soviet bioweapons program.

We are in uncharted waters.

“Terrorists don’t have enough lab experience.”

Mahajan and other people who urge calm when it comes to AI bioterror point out that AI, so far, can’t actually do the physical lab work of creating a doomsday virus. Yes, people are working on automating bio labs, but that’s probably still pretty far in the future.

But that doesn’t mean the angry teenager listening to Nirvana in his bedroom will actually have to synthesize the doomsday virus with his own two hands. His AI agent may be able to just order it through the mail. Plenty of labs do genetic modification on demand. Right now, these labs aren’t monitored very well. Here’s the Washington Post:

The federal government should mandate that all laboratories capable of gene synthesis screen for potentially harmful biological agents. Many large biotech companies already have such measures, but not all do. Uniform federal standards would help.

In other words, we currently don’t even have a monitoring or reporting requirement for ordering bioweapons through the mail. Top AI leaders, well aware of this danger, have written an open letter calling for new legislation to impose such requirements:

That’s good, but it’s not clear that this would get rid of the danger. First of all, there are plenty of illicit labs and underground labs that might simply not obey the reporting requirements. California recently uncovered a secret biolab with links to China:

Code enforcement officer Jesalyn Harper discovered a clandestine biolab in a Reedley, Calif., warehouse containing dangerous pathogens including HIV, malaria, COVID-19 and Ebola...The facility was allegedly operated by a Chinese national involved in smuggling counterfeit COVID tests while receiving millions in unexplained payments from China…A congressional investigation revealed critical gaps in U.S. regulations governing unlicensed biolabs, raising major national security and public health concerns.

On top of secret labs like that, what about labs in other countries? If rogue Chinese researchers are operating an illicit biolab in Kyrgyzstan or Indonesia or North Korea, who is monitoring that? Who will force that lab to report “potentially harmful biological agents”, and who would the lab report to?

Remember that all it takes is one angry teenager to carry out the scenario I described above. That teenager doesn’t have to live in the U.S., or even in a developed country; all he needs is a good internet connection and a postal service. Nor does the lab that sends him the doomsday viruses need to be in the U.S., or in any country where we have any measure of control. The doomsday virus(es) can be created anywhere on Earth, and released anywhere on Earth, and will still come here to America and kill you and me and everyone we know.

Finally, what makes us think that the doomsday virus would even be caught by a safety screening? A truly novel supervirus would presumably rely on hitherto-unknown mechanisms for killing humans, in order to minimize pre-existing immunity — and to minimize the ability of the existing industrial infrastructure to create and disseminate antivirals in time. It seems like monitoring systems will have trouble catching something truly novel. The lab that creates the doomsday virus and mails it to the angry teenager may have no idea what it’s making — and no real way to know.

“No one has done this yet, therefore it’s a lot harder than you think.”

It might seem surprising that people actually make this claim, which sounds a little like saying “I haven’t died yet, so I don’t need to wear my seat belt.” And yet I do see claims like this from time to time:

The obvious reason this is a bad line of thinking is that doomsday only happens once. “We’re still here” is a true statement right up until it’s not. Suppose that the creation of very powerful AI tools creates an annual doomsday risk of 10%. We’d be very likely to survive the first year of that risk, but after a decade of that, we’d only have about a one-in-three chance of surviving.

There’s also no reason to assume the risk is constant. AI models continue to improve at a rapid pace. Human awareness of AI model capabilities continues to spread. There are plenty of reasons to think that 2029 will be a lot riskier than 2025 was, in terms of bioterrorism.

The analogy with cybersecurity, meanwhile, is deeply flawed. Software can be rewritten, and vulnerabilities patched before attackers even think about attacking; humans are much harder to safeguard. Cybersecurity software improves, but human physiology remains at the same basic level of vulnerability year after year while AI gets better and better.

On top of that, if one hacker gets through, the results are usually not catastrophic; with doomsday viruses, all you need is one “success”. In fact, we have seen some stunning cyberwarfare successes recently, like Iran’s shutdown of a British power plant this month. We don’t know whether they used a Mythos-level model, but I wouldn’t bet against it.

So anyway, none of the standard dismissals of AI bioterror risk make any sense to me. Since I started yelling about this topic, I’ve seen basically nothing to persuade me that the danger is substantially less than I originally thought. And although I’m starting to see more people talk about AI bioterror risk, I don’t see the kind of urgency that a threat this existential and plausible ought to receive.

What should we do?

This leads to the final question: How can we prevent AI from being used2 for catastrophic bioterrorism? Obviously the kind of monitoring systems being promoted by AI industry leaders would help, though that’s only a first step.

The most important thing I think we can do right now, to be honest, is just scare people. AI leaders are on the right track with their open letter, but their attention is being spread across a diffuse set of risks. The same is true of AI researchers, AI safety enthusiasts, big philanthropic donors, think tanks, economic and political commentators, and pretty much everyone else. Creating a healthy perception of peril around AI bioterror among the highly-informed professional class will focus the best minds — both natural and synthetic — on this problem, instead of letting them run hither and yon worrying about every possible bad thing that AI could cause.

I’m not saying that cyberattacks, misinformation, and job loss are unimportant; I’m saying that if an angry teenager uses AI to create the doomsday virus, all of those will quickly become unimportant.

A healthier sense of our own peril would also make international cooperation a lot easier. Hunting down that rogue biolab in Kyrgyzstan will be a lot easier if China and the U.S. focus more on biosafety and less on nationalistic competition. But that will only happen if both China’s and America’s leaderships consider saving the human race from a vibe-coded plague to be an absolute priority.

In a very real way, AI risk is AI bioterror risk. If AI ever does destroy the human race — either on its own initiative or at a madman’s behest — it’s very likely that it will do so using a virus. An army of robots, a la the Terminator or Matrix films, would be much harder. A plague of self-replicating nanomachines is basically no different than a virus. The nuclear weapons are going to be kept too tightly locked up. A virus is far and away the most likely tool for eliminating a biological species.

Anyway, I know this was a scary blog post, and it was supposed to be. But this is one of those cases where we just can’t afford the luxury of doing what we normally do — burying our heads in the sand and telling ourselves that nothing is going to go wrong. Usually, when we refuse to anticipate calamities — wars, financial crises, natural disasters — we can pick up the pieces afterwards. But there have to be pieces left to pick up.

Share

I still remember the first time I set foot on Market Street. I was a freshman in college, and big cities were still a very new thing to me. It felt like I was swimming through an ocean of foot traffic; the towering shops and hotels seemed like they were built by giants.

That Market Street is gone now. In its place is a giant empty field of asphalt in the center of San Francisco — a desolate gray waste baking in the sun, empty except for the occasional bus and a tiny smattering of bicycles.

On either side of this yawning wasteland are giant sidewalks with only a few pedestrians, boarded-up stores, and the occasional small, sad retail chain. At midday on a Monday, what was once the bustling commercial heart of San Francisco feels abandoned.

What happened? In fact, Market Street was destroyed by several things:

1. The city’s decision to close the street to cars (but not buses), enacted in 2019

2. The crime wave of the early 2020s

3. The pandemic, and the shift to remote work

The second and third of these, together, are known as the “Doom Loop”. It’s been talked about a lot, so let’s just cover it briefly.

The pandemic, and the economic and social changes it brought, came out of the blue. But San Francisco was particularly vulnerable to the shock. As in many American cities, 20th century urban planning and zoning rules concentrated office buildings in the downtown area and put housing farther out. This meant that once people stopped going downtown for work, the neighborhood was just a lot more empty, because almost no one lives there. This is in contrast to big cities in Europe and Asia, where downtown neighborhoods are typically mixed-use. (SF also had a stronger shift to remote work, because of the heavy presence of the tech industry.)

The crime wave was another shock. Although SF escaped the destructive rioting that hit other West Coast cities like Seattle, Portland, and Los Angeles in 2020, it had plenty of looting in downtown areas — especially around Union Square and the San Francisco Centre mall on Market Street. Even after the pandemic ended, San Francisco suffered especially hard from the statewide wave of “smash-and-grab” store thefts. It got so bad that American Eagle actually sued the San Francisco Centre for being negligent toward crime, reporting over 100 serious incidents of violence and harassment in its downtown store.

Crime and emptiness fed on each other. Less foot traffic meant fewer “eyes on the street” to discourage property crime. Property crime and reduced foot traffic forced stores to close. Fewer stores meant less foot traffic, and so on. The collapse of downtown culminated in January 2026 with the closure of the San Francisco Centre, the iconic mall that had formed the commercial heart of Market Street, but which had been gutted by the Doom Loop. It now stands as a looming, empty, boarded-up structure dominating the area around the Powell Street train station:

The intersection of Powell Street and Market Street, once bustling and alive with tourists, shoppers, and street vendors, is now virtually empty, except for a forlorn line of people waiting to get on the iconic streetcar:

But recent years have seen the first glints of recovery from the Doom Loop. For one thing, crime has begun to fall. 2025 alone saw a 27% decrease in property crime, including a 29% decrease in burglaries, a 22% decrease in larceny (including retail theft), and a 24% decrease in robbery. 2026 has seen further drops. The reasons for this are interesting, but I’ll go into them another time. Suffice it to say that this has made downtown SF feel much safer and more orderly. The stereotype of SF as being overrun by drug addicts and dangerous vagrants still holds true of a few areas of the city, but from Powell Street to the bay, the city no longer feels like a post-apocalyptic movie.

Meanwhile, the business district is showing a few sluggish signs of life. The AI boom is causing a (modest) rebound in the demand for office space. Some office buildings are being converted to housing, raising hopes that SF’s downtown can finally become a mixed-use neighborhood. A concerted effort to revive Union Square after the devastation of 2020 is beginning to bear fruit. And both companies and the SF city government are mandating return-to-office policies.

And yet Market Street remains a dead zone. Weekday foot traffic is still down by half compared to 2019:

The wounds inflicted by the Doom Loop will take lots more time and effort to heal. Retail businesses have to be encouraged to open. Defunct buildings have to be torn down and rebuilt. Crime has to be reduced even further, and so on. This will all take years.

But there’s one other factor contributing to the death of San Francisco’s most important street, and it began before the pandemic: The ban on cars.

In recent years, a number of cities around the world have experimented with car-free zones and traffic reduction policies in their downtown areas. The most famous of these is Paris, with its Limited Traffic Zone, its expansion of pedestrian malls, and its road closures during Sundays and holidays. Shortly before the pandemic, San Francisco decided to enact a plan its leaders thought was in the same spirit: closing downtown Market Street to all car traffic:

But although Market Street was closed to cars, it wasn’t closed to buses, delivery trucks, and licensed taxis. This was in keeping with the general “transit good, cars bad” thinking of modern progressive urbanism. But a bus will kill a pedestrian just as easily as a car will, which is why Market Street remains empty of pedestrian traffic. This mathematically guaranteed that Market Street would become a more empty place: Cars were subtracted, but pedestrians weren’t added.

The direct cost of this policy is pretty obvious. When people can’t park in a parking garage on Market, and when Uber and Lyft can’t pick people up and drop people off on the street, it makes businesses on the street less attractive to visit. Sure, people could bike or take the train, but not everyone wants to do that; instead, they’ll just go to restaurants or stores in other areas. In addition, restaurants on Market Street now can’t deliver food, cutting them off from an important source of revenue. That probably helps explain why Market has suffered more from the Doom Loop than other areas.

But what did SF gain by deleting cars from Market Street and adding nothing at all in their place? Cars create a little bit of pollution (though modern cars are very clean), so that was removed. Pedestrians who run off the sidewalk and into the street are somewhat less in danger of being killed (though still in danger). These are meager benefits indeed.

What about bike traffic? Presumably, reduced danger of being hit by a car would allow more cyclists to ride down Market Street. But in fact, bike traffic on the street is considerably lower than it was pre-pandemic:

A lot of that is the Doom Loop — there are simply fewer stores for cyclists to visit on Market Street, so there’s less reason for them to go there. But much of downtown hasn’t collapsed the way Market Street has, and presumably cyclists could use the street as a way to get to places like the Embarcadero, Chinatown, etc. They’re not doing it. Cyclists are a relatively rare sight on Market these days.

One reason might be that although the city planned to make the car-free stretch of Market more bike-friendly, it never really followed up on those plans. So it’s still pretty dangerous to ride a bike or scooter down the street — you can get caught in the train tracks, and you can get hit by a bus or truck.

In fact, this sloppy implementation is a general pattern with the Market Street closure. Josh Koehn had a great article in 2023 about how the plan degenerated from its initial form:

A $600 million capital project called Better Market Street promised to create a futuristic boulevard that would safely buffer bicycles and scooters on elevated sidewalk lanes, separating the little wheels from rapid bus lines, vintage streetcars and pedestrians…The new vision for Market Street, advocates said, would create a safer, more vibrant multimodal avenue…More than a decade of planning went into alterations of Market Street…

But funds for anything beyond modest improvements along a three-block stretch of Market Street between Fifth and Eighth streets…have evaporated…The cost of Better Market Street, which soared above a billion dollars before being reined in to a still-staggering $604 million, is almost double the amount spent on the tormented Van Ness Avenue redesign…

City officials enraged safe streets advocates when they announced in late 2020 that the elevated bike lanes on sidewalks would be scrapped.

SF’s lack of state capacity had a lot to do with this; constant cost overruns inevitably require ambitious projects to be scaled back. The pandemic also obviously got in the way. But part of what killed the vision for Market Street is simply that the vision itself was never very good.

If you look at Paris, it’s immediately clear how that city’s car-limitation measures were designed to create spaces for pedestrians. In certain areas like the banks of the Seine, Paris didn’t just ban cars; it also banned buses and trucks, so that people could walk safely. Paris also reserved much larger areas for pedestrians on certain days.

But Market Street has no more pedestrian areas than it did in 2019. You can still only walk on the sidewalk. If the entire street were reserved for foot traffic during certain times — like SF does with Valencia Street, for example — then these periodic temporary pedestrianizations could become an event that regularly draws lots of people to the area. This could include farmers’ markets, street concerts, and other fun events. Those would give people a reason to walk around Market Street even before new businesses opened.

Alternatively, if SF permanently closed off some portion of the street to buses and trucks, that stretch of road could become a pedestrian public square, similar to the center of Times Square:

Instead of thinking about how to create spaces that people would actually want to walk around in, the SF government — and the many activist groups and nonprofits that influence its policy — appear to have adhered to the simple rule of “bikes and buses good, cars bad”.

This is a way of thinking centered around modes of transit rather than public space. It implicitly conceives of areas like Market Street as thoroughfares to take people from one place to another, rather than as actual destinations. Market Street’s car closure is all about forcing people to go down Market Street in a different way, rather than about encouraging people to go to Market Street.

There’s something inescapably and depressingly suburban about this way of thinking. In the suburbs, life is lived point-to-point — you go from house to strip-mall to restaurant to office without thinking much about the places you traverse in between. In a suburb, the mode of transportation — how you get from Point A to Point B — is the important question.

Americans are a suburban people; even many of our urban areas look like suburbs in other countries. So perhaps it’s inevitable that the urbanists who closed Market Street off to car traffic would think entirely about modes of transportation and ignore the vibrant city center whose destruction they exacerbated. Instead of Jane Jacobs, they ended up just being the Robert Moses of buses.

Again, careful attention to car limitations in cities like Paris, where people are used to occupying urban places instead of simply whizzing through them, would have proven enlightening. Paris’s Limited Traffic Zone doesn’t ban cars in downtown areas; it merely bans cars from transiting through those areas. People who stay within the area can still drive. The point is to make downtown spaces into places people occupy rather than places people simply go through. What San Francisco did to Market Street was, in many ways, the exact opposite.

In lieu of turning Market Street into a pedestrian paradise, SF’s best move is probably just to allow private cars again. It’s not the optimal way to create more foot traffic to nourish retail businesses, but it’s the quickest and easiest. In fact, Mayor Daniel Lurie has made moves in this direction, allowing certain Uber, Lyft, and Waymo cars to operate on the street. But they might as well simply go all the way. Best to simply bury the era when Market Street was a large empty scar commemorating the decline of San Francisco; if San Francisco can’t be Paris, at least it can return to what it used to be.

Subscribe now

Share

It’s hardly news that Americans are not happy with the state of their nation. Confidence in the country’s direction crashed hard during the Bush years and never recovered. But what’s notable about more recent years is Americans’ increasing dissatisfaction with their personal lives:1

This decline looks pretty mild, but other numbers tell an even starker story. For example, happiness surveys show an increasingly despondent populace:

Suicide has been slowly rising as well. And then there’s consumer sentiment:

For intellectuals and commentators such as myself, for many policy wonks and academics, and even for many politicians, the question is how to make the American people happier. And the answers we come up with are usually economic ones — ideas for policies that will improve the material well-being of either the whole populace, or some segment of it.

This is common on both sides of the aisle. On the left, centrists have embraced the Abundance movement, whose main idea is to give Americans more stuff — more housing, electricity, health care, and so on. For progressives and leftists, meanwhile, Medicare for All is probably the centerpiece idea. On the right, Trump’s trade agenda has usually been justified with claims that it will return good manufacturing jobs to the American working class. Even immigration restriction is usually couched in economic terms — the MAGA people claim that mass deportations will open up jobs for Americans, lower the price of housing, and so on.

The roots of the country’s malaise, too, are often described in economic terms. Many people claim that the First China Shock — and the loss of good blue-collar jobs to offshoring more generally — left a disappointed, angry working class in its wake. Progressives decry corporate landlords that they claim raise the cost of housing, and blame grocery stores’ greed for the rising cost of food. Practically everyone agrees that inflation was the biggest reason Trump won the 2024 election.

I’ve spent a lot of time arguing against a lot of these ideas (and in favor of a few of them). But I almost always accept the basic frame that A) many of America’s problems are due to economic malaise, and B) we ought to focus on finding ways to make Americans richer. Partly, this is because I was trained as an economist; I recognize that things like culture wars are also important to Americans, but I don’t have much special expertise in that area. But partly, I think I just subscribe to the general tendency of American intellectuals to focus on the material.

I was thinking of this as I read Jay Caspian Kang’s essay, “Subaru Socialists and the Great Disappointed.” Kang argues that the young, educated, angry voters who currently form the Democrats’ base can be separated into two basic groups: Millennials who are mad about the raw economic deal they received during and after the Great Recession, and Zoomers who are mad about Covid and Palestine. Here he describes the first of these groups:

This new voting bloc has two similar but distinct components. There are what I call the Subaru Socialists: college-educated millennials in their thirties and early forties who earn a somewhat steady income but not what they expected. These voters arrived at their politics via the Great Recession, Occupy Wall Street, and the continual escalation of housing prices in urban areas—all of which heightened the contrast between what they were promised when they took on student debt and their current circumstances.

For a long time, I basically bought the narrative that at least some of the unrest of the 2010s was due to angry Millennials who had gotten screwed by the Great Recession. It makes intuitive sense that a generation who were on the losing end of such a vivid demonstration of economic risk would turn toward policies that decrease economic risk — like government health insurance and rent control. It also makes sense that they’d want student debt forgiveness.

But as time has gone on and I’ve seen Millennials do better and better economically, I’ve begun to doubt this narrative. Corinth and Larrimore (2026) write:

We find that Millennials had a real median household income that was 20% higher than that of the previous generation, a slowdown from the growth rate of the Silent Generation (36%) and Baby Boomers (26%), but similar to that of Generation X (16%). The slowdown for younger generations largely resulted from stalled growth in work hours among women…Additionally, lifetime income gains for younger generations far outweigh their higher educational costs.

And here’s a chart:

It’s not just government redistribution giving Millennials a boost. By the time they hit their mid 30s, their market income was well ahead of Gen X or the Boomers (and yes, this is adjusted for the cost of living). They had caught up in wealth, too:

Now, is it possible that even though Millennials eventually did OK, the disruptions of their 20s created a scarring experience that they’ve never forgotten? Yes, it is absolutely possible. But it’s also notable that the kind of economic policies lefty Millennials demand — especially universal health insurance — don’t really address the main kinds of risks they experienced in their youth (unemployment and wage declines).

On the right, it’s notable that the economic policies Trump has unleashed have not helped the people they were supposedly designed to help. Manufacturing employment has fallen since Trump took office and started putting tariffs on anyone and everyone:

So much for those blue-collar jobs coming back. Meanwhile, mass deportations and a near-cessation of net immigration to the U.S. have spectacularly failed to raise employment rates for native-born Americans:

It’s possible, of course, that Trump’s supporters think his policies will take a long time to work, or perhaps they simply aren’t very aware of what’s going on in the economy. I kind of doubt this, though. Americans as a whole have noticed, which is why Trump’s approval rating is in the gutter and inflation — the most important economic issue at the moment — is the issue on which voters are most angry at Trump:

So it’s possible that Trump’s voters were just tricked into thinking he was better on economics than he was. On the margin, that’s probably true. But the fact that Trump’s base has stuck with him strongly implies that economics was not their primary concern.

In fact, there was a whole huge debate among political scientists and economists about whether Trump’s victory in 2016 was due to “economic anxiety” or to sociocultural issues. In the end, the “economic anxiety” hypothesis generally lost out. Here’s Margalit (2019):

Empirical findings indicate that the share of populist support explained by economic insecurity is modest. Second, recent evidence indicates that voters' concern with immigration—a key issue for many populist parties—is only marginally shaped by its real or perceived repercussions on their economic standing.

On the left, I think there’s pretty good evidence that economic issues are not foremost in the minds of the progressives who dominate the Democratic Party. The political analyst David Shor has found that the issues Democratic donors and activists care about are just different from the bread-and-butter economic concerns that motivate swing voters:

In reality, Shor says, young party staffers are far to the left of the median Democratic voters on relatively uncontroversial, bread-and-butter Democratic priorities like combatting income inequality or addressing climate change. In their 2015 paper, for instance, Enos and Hirsch found that 23 percent of Obama staffers cited income inequality as the single most important issue facing the country, whereas polls from that election cycle found that fewer than one percent of all voters listed “the gap between rich and poor” as the most important issue. Enos and Hirsch also found that campaign workers were more likely to cite health care and inequality as an important issue to voters — even though most voters did not list those as high-priority issues and said they were more concerned about things like war and inflation.

And although much has been made of the fact that low-income highly-educated voters tend to break strongly for the Democrats, it turns out that high-income highly-educated folks also vote blue:

In fact, in his essay, Kang describes Gen Z voters as caring less about the economy than about Palestine:

Then there are the Great Disappointed, a younger set. They don’t have strong memories of the 2008 financial crash, as their millennial elders do, but they came of age during the pandemic and saw that seemingly invincible institutions could, in fact, sputter and fail when placed under duress. Like the Subaru Socialists, these twenty-two-to-thirty-year-olds aren’t exactly happy with the return on their educational investment. But their political views have seemingly coalesced less around the economy than around the war in Gaza, which serves as a character test for any politician who asks for their vote…A majority of young Democrats believe that their tax dollars are funding a genocide; candidates like Melat Kiros, who defeated the longtime Colorado Congresswoman Diana DeGette in a primary, are running to the left of progressive incumbents not so much on economic issues as in their condemnation of Israel and of America’s continued military aid to the country.

The name Kang assigns to these voters — the “Great Disappointed” — seems at first to imply some kind of economic disappointment. But it turns out he just means they’re angry about foreign policy stuff that they see in the news.

In fact, a lot of committed DSA types are actually post-economic. The Gen Z political strategists who recruited Graham Platner to run for Senate, Morris Katz and Daniel Moraff, were found to be scions of very wealthy families. Similar stories keep popping up:

If people like this are “disappointed”, their disappointments have little to do with their material standard of living.

In other words, over the past decade I’ve increasingly come to doubt that economic policies and programs can address the true concerns of the people driving American politics. Yes, economics still matters, of course — for humanitarian reasons, if nothing else. And yes, there are still a lot of voters who care about inflation, jobs, and other economic issues. But these are not necessarily the people setting the agenda in American politics or causing unrest in American society.

Even where “normie” voters are concerned, though, it’s proving maddeningly hard to craft economic policies that give them what they want. Even though inflation fell from 9% in the summer of 2022 to just 3% in the summer of 2023, voters still seemed to punish Democrats at the polls in 2024.

And although inflation during Trump’s current term has never approached anything resembling 2022, voters still say they’re incredibly mad at Trump over inflation.

With a robust job market, low-ish inflation, and decent economic growth, analysts have struggled to figure out why consumer sentiment — at least, as measured by UMich — is at all-time lows. At a loss, commentators have been reduced to talking about “vibes” — basically a label for our ignorance.

Every time a policy wonk thinks about some measure that would increase GDP by 0.4%, or create 200,000 jobs, or lower inflation by 40 basis points, they need to think about the unpleasant reality that Americans may simply not care very much. That doesn’t mean it’s bad to help people economically, but it means that economics may ultimately not be the lever we need to pull if we’re going to make the people of this country happy and satisfied again.

We should therefore think about whether our primary social problems are actually about material scarcity. The vast majority of Americans now possess not only the basic necessities of life, but a degree of economic security unimaginable to their forebears a century ago. It seems plausible to think that a great many Americans have climbed to the higher rungs of Maslow’s Hierarchy of Needs — that instead of where to get their next meal or how to save for retirement, they’re thinking about how to feel like they belong in their society, or how to get more status and respect.

A large contingent of Americans may now simply care less about adding to their bank accounts than about questions like “Who are the real Americans?”, or “Does one race deserve special treatment in the eyes of the law?”, or “Why do some of the people I went to college with have 10,000 times as much wealth as I do?”. As long as those questions remain unanswered, boosting Americans’ economic fortunes could have little effect on their happiness.

Neither the discipline of economics nor our typical policy discourse is set up to deal with issues like belonging, social status, or respect. There’s no market for these things; people with wealth might be more respected, but you can’t buy respect on Amazon. And some of these psychological needs are positional goods — for the status of one person or group to be higher, someone else’s status often has to be diminished.

Not only is modern economics poorly set up to deal with those kinds of things, it often bothers us to even talk about them. A world dominated by positional goods — where one person having more requires that someone else have less — is a dark, zero-sum world. We would much prefer to work toward a world where everyone wins.

And yet like it or not, this may now be the world in which we find ourselves. Just as the great challenge of the 20th century was to provide broad-based material plenty, the great challenge of the 21st century may be to make everyone feel high-status at the same time — or at least to create a world where status differences don’t lead to mass resentment.

Subscribe now

Share

“Gradually, then suddenly.” — The Sun Also Rises

Lots of people are worried about rising long-term interest rates. Short-term interest rates are controlled by the central bank. But the Fed doesn’t usually intervene in the market for longer-term bonds, so when these interest rates move around, it means the market is telling us something. So a lot of people were worried when the yield on 30-year U.S. Treasury bonds jumped by 6 basis points (0.06%) the other day.

Why were people scared? Well, remember that when interest rates go up, it means bond prices went down. Which means that fewer people wanted to buy U.S. government bonds. This could be a signal of several bad things:

1. It could signal expectations of higher inflation. When future money will be worth less, bond investors demand higher interest rates today. Higher inflation also means the Fed will probably raise interest rates.

2. It could signal a loss of confidence in the U.S. government. If people think there’s a possibility that the U.S. won’t repay its debts, they will charge a higher risk premium to hold that debt. These doubting investors are sometimes called “bond vigilantes”. Bond vigilantes could be scared by soaring U.S. debt levels and deficits, and/or by irresponsible geopolitical and economic policies from the Trump administration.

Here’s John Cochrane on the specter of the bond vigilantes:

Unsustainable fiscal policies can only go on so long. Eventually bond investors decide that the US will not in the end do the right thing after trying everything else, and default, expropriation, taxation, capital controls, or sharp inflation is on its way. They stop buying long-term bonds especially, and look to the comfort of short term bonds…For some reason there is limited demand for long-term treasury debt…

The beginning of a global sovereign debt retrenchment would show up first in a feeling of limited demand…Investors, seeing trouble demand a larger risk premium for longer term debt…Moving to short maturity structures is a classic symptom of trouble ahead.

Among the people who were scared by the rise in interest rates, apparently, were the Trump administration. Higher long-term interest rates mean higher mortgage rates,1 which make American voters mad. They also make it harder for the U.S. government to finance its enormous deficits. So Treasury Secretary Scott Bessent announced that the government was intervening in the bond market, with a program to buy long-term U.S. Treasury bonds. This pushed up bond prices — and pushed down interest rates — for exactly one day, but then the bond markets bounced right back:2

US Treasuries fell a day after the Trump administration’s surprise decision to increase buybacks of longer-dated bonds…The 30-year yield on Thursday rose over seven basis points to as much as 5.27%, where it was just ahead of the US Treasury Department’s announcement early Wednesday.

So are we watching the collapse of confidence in the U.S. government? Is this the beginning of bankruptcy for Uncle Sam? Probably not yet. The underlying trends of excessive borrowing and boneheaded policies do bear keeping an eye on, and collapses of investor confidence can happen fast once they begin. But it’s unlikely that a true sovereign debt crisis has begun.

How much are interest rates actually up by?

One reason I’m not terrified by the rise in long-term rates is that it’s not actually that big of a rise!

In stories about rising rates, you see a lot of charts about how it’s a global phenomenon:

This has interesting implications, but first, notice that except for Japan, most of the rise was actually in 2021-2023. So whatever happened to make bond investors demand higher rates, most of it happened years ago.

Here’s just the U.S., zoomed out to cover the last decade:

Long-term rates fell in 2019 and bottomed out during the pandemic, then in 2022 and 2023 they had a big sustained rise. In comparison, the rise since early 2026 has been very small — only a few tenths of a percent.

That could be the beginning of a catastrophic rise, and of course when you’re carrying as much debt as the U.S. government is, even a small increase in borrowing costs can be a headache if it’s sustained over a long period of time. But I just can’t look at that little wiggle in 2026 and see evidence of a bond market collapse. You should be very worried about the U.S. national debt, but this rise in rates should only make you a tiny bit more worried, if at all.

So what is behind the (small) rise in long-term rates? There are three basic possibilities.

Is it inflation?

Read more

The Economist has an interesting article this week, in which it goes after Daron Acemoglu. Acemoglu is probably the top economist in the world at this point, having just earned a Nobel prize, and sitting right at the top of the list of most cited economists. So it’s kind of funny that when it talks about people who doubt Acemoglu’s research, the one person it mentions is…me!

Give an economist a few drinks, however, and some of them will venture their true opinions about this giant. “Much of his theoretical work is useful, but he uses his models to inform populist policies that have been tried before and failed,” blasts one well-known economist. Some commentators do not require Dutch courage. “I’ve been yelling about Acemoglu for literally a decade,” Noah Smith, an economics blogger, has said, in response to a flurry of online criticism of Mr Acemoglu’s work.

(For the record, The Economist did not actually interview or contact me for this article. Their quote of me was from a tweet from June, which they did not link to in the article.)

It is true that I have been pretty critical of much of Acemoglu’s work over the years. In 2012 I argued strongly against a paper he wrote in which he claimed that America is more entrepreneurial than Sweden because America’s weaker safety net forces people to work harder (in actuality, Sweden is more entrepreneurial by most measures). In 2022 I pointed out that a famous Acemoglu paper claiming that robots destroy jobs was actually an outlier, and listed a bunch of other papers that reach the opposite conclusion. I liked his book Why Nations Fail, but I also admit that its empirical foundations are fairly shaky. I didn’t like Acemoglu’s more recent book Power and Progress, and I was highly critical of a recent paper Acemoglu wrote about AI and productivity.1 I also think Acemoglu’s argument that skilled immigration has crippled the American education system — now being heavily cited by Breitbart and such — is extremely implausible.

But this is hardly newsworthy. I go after Acemoglu’s work because I know he can take it; he’s a titan of the economics field, and I am but a lowly blogger. Nothing I say is going to affect his reputation or his prestige, even if someone at The Economist mentions my critiques in an article. Indeed, after the article came out, top figures in the profession rushed to condemn it and to defend Acemoglu. The Economist does list a bunch of — usually justified — criticisms of Acemoglu’s work, but this does not mean Acemoglu has been discredited as a researcher or exposed as overrated; indeed, it would be difficult to name a top economist whose body of work does not contain a variety of questionable theoretical assumptions, motivated reasoning, and/or shaky empirical results. This is a problem with the field itself, not the man; it’s part of a broader crisis of unreliability throughout much of academia.

But for that same reason, despite the Economist article’s lack of newsworthiness, I think it’s good that the magazine decided to come at the king. Economics is far too hierarchical and closed of a profession. Younger and less accomplished researchers routinely defer to the authority of famous and senior figures, and critics from outside the field are typically brushed off. Exactly how it got to be this way is a topic that deserves a longer post, but I believe the hierarchical, closed culture of econ has resulted in a research literature that has been too skewed toward the priorities and intuition of top authority figures. I thus think it’s generally a healthy thing to tweak the tails of those Olympian figures, as long as criticisms are grounded in substance. As they say, “science is the belief in the ignorance of experts.”

So in that spirit, let’s critique another Acemoglu paper.

First, the background. Recently, a lot of people (including myself, but also prominent economists) have begun worrying about low fertility rates. Across all countries — poor as well as rich — fertility just keeps going down and down, with no floor in sight.

This presents two distinct dangers: population aging and population shrinkage. Aging, rather mechanically, creates a burden for young people, because you have more retirees who have to be supported by each worker, either through taxes or through family support. It also might reduce productivity, for example if older managers are less innovative, but that’s more speculative. Population shrinkage, meanwhile, is a threat to total GDP, which you might care about if you want your country to be more powerful. There’s also the possibility that a smaller population might reduce growth — by reducing the opportunities for specialization, or by reducing the available pool of researchers.

But some people argue that a shrinking population is no problem — or might even be a good thing. When you make human workers scarce, it creates an incentive to invest in labor-saving technology, which boosts productivity. This mechanism has been proposed by some growth theorists, and some economic historians, like Robert Allen, even think this is what caused the Industrial Revolution! This effect might be strong enough to cancel out the aggregate effect of population aging, so that society stays just as rich — or even gets richer — due to low fertility rates.

This is actually a plausible mechanism. I have to say, I’m skeptical that it works in the general case. Human beings aren’t just labor supply; they also create labor demand. The incentive for businesses to buy new machine tools, robots, AI, etc. is that someone is going to buy the stuff they produce with those tools. If you have fewer people, you have fewer consumers. This is probably the reason why the effect of immigration on wages is typically close to zero. Babies are not that different from immigrants. A larger market size can also create an incentive for more rapid innovation — in fact, Acemoglu and Linn (2004) find that faster population growth increases pharmaceutical innovation.2

But anyway, the idea that population scarcity stimulates innovation is the thesis of a new paper by Daron Acemoglu, David Autor, Keelan Beirne, and Andrew Scott. This is from the abstract:

Contrary to the widespread expectation that [low birth rates] hamper economic growth, we find lower birth rates are associated with higher growth in GDP per working-age adult across countries and higher wage growth across US commuting zones, with no negative impact on aggregate GDP or earnings. These patterns are not explained by educational upgrading, rising female labor force participation, the declining importance of agriculture, or neoclassical-Solow mechanisms. We argue that they reflect the endogenous, labor-saving response of technology to the scarcity of younger workers. Consistent with this interpretation, countries and regions with lower birth rates exhibit more labor-saving patents and growing high-tech activity. There is also higher TFP growth across countries and industries. [emphasis mine]

It’s a plausible result. But before we go on to the details of the paper, let’s talk about what this result would imply, if it does turn out to be true.

One of the paper’s basic theses is that technological automation raises wages. That flies in the face of the empirical work that Acemoglu has done on robots.3 It also more broadly seems to contradict what Acemoglu has been saying about AI, both in his research papers and in his public statements. In his 2024 paper, “The Simple Macroeconomics of AI”, Acemoglu writes that “predicted TFP gains over the next 10 years [from AI] are…predicted to be less than 0.53%.” In his 2021 paper, “Harms of AI”, Acemoglu writes:

I argue that if AI continues to be deployed along its current trajectory and remains unregulated, it may produce various social, economic and political harms. These include: damaging competition, consumer privacy and consumer choice; excessively automating work, fueling inequality, inefficiently pushing down wages, and failing to improve worker productivity. [emphasis mine]

This stands in direct contradiction to Acemoglu’s new paper with Autor, Beirne, and Scott. If modern automation technologies push down wages without raising productivity, it cannot compensate for population aging in the way that Acemoglu, Autor, Beirne, and Scott claim that it must — and therefore, population aging is dangerous in a way that it wasn’t before.

Now, it’s fine for a researcher to find two seemingly contradictory things — or even put forward two contradictory theories — in two different papers. The real world is messy and complicated, and there are plenty of “puzzles” in the literature. But if Acemoglu goes around simultaneously telling us:

1. not to worry about population aging, because automation will compensate for it with higher productivity and higher wages, and

2. to worry a lot about automation, because it pushes down wages without raising productivity much…

…then we have a problem.4

But anyway, on to the actual details of Acemoglu et al. (2026). First of all, I’m not convinced by their result that lower birth rates raise living standards at the country level. The reason is that once they control for a bunch of stuff — education, urbanization, geographic region, etc. — the result loses statistical significance. Here’s their Table 1, with the estimates I’m talking about circled in red:

In their abstract, the authors declare that “these patterns are not explained by educational upgrading, rising female labor force participation, the declining importance of agriculture, or neoclassical-Solow mechanisms.” But at least some of these patterns can apparently be explained by other things — initial education levels, urbanization, and so on — which makes the claim in the abstract a lot less impressive than it sounds.5

Also, note that the more controls the authors add, the weaker the estimated effect becomes. That’s generally a red flag in empirical papers. No matter how smart and careful you are, there are always things you don’t control for, especially in a cross-country regression, since countries differ in so many ways. So if the things you do control for tend to weaken your headline result by a lot, you should be worried that the rest of your result can be explained by the controls you left out.

For example, what about institutions, which Acemoglu has spent much of his career telling us are the main determinants of development? If the legacy of colonialism can be canceled out by passing out free condoms, why did Acemoglu win a Nobel prize? The word “institutions” does not even appear once in this new paper!6

A more realistic possibility is that lower birth rates in 1940 were correlated with other things — maybe a lot of other things — and that it’s these other things, rather than birth rates, that give rise to the correlation in the paper.

Of course, what we ultimately care about isn’t GDP per working-age adult7 — it’s GDP per capita. The authors have data on GDP per capita, but they barely mention it in the analysis;8 I’m not sure why. They do look at total GDP, and here they find no correlation, but a big standard error:

We shouldn’t interpret absence of evidence as evidence of absence. Cross-country regressions have small samples and tons of heterogeneity, so their standard errors tend to be huge — they just have trouble explaining much about the world. But the authors sort of blur this line, saying things like “no negative impact on aggregate GDP or earnings”, when really they should say “we can’t find a negative impact.” There could be a positive relationship between birth rates and growth hiding in that cloud of data points. (Sadly, this is standard practice in economics, but it makes a difference in how results get sold.)

Anyway, the result for wages in U.S. commuting zones holds up somewhat better. It’s a bit more statistically significant, and it doesn’t get attenuated as much by adding controls. But here, I’m worried about different issues: sorting and clustering. Basically, the result shows that coastal cities — places like New York City, San Francisco, Boston, etc. — had both lower birth rates in 1940 and faster wage growth from 1970 to 2020. The authors’ interpretation is that businesses in NYC, SF, and Boston had fewer workers, and so were forced to embrace labor-saving automation, while businesses in places like Birmingham, Alabama or Gary, Indiana refused to automate because they were endowed with plentiful cheap labor thanks to the legacy of an extra-large Baby Boom.

I am suspicious of this result, because American cities are not independent of each other. People can pretty easily move from Gary, Indiana to Boston, and vice versa! Ideas and capital can move even more easily — it’s pretty trivial for a company that patents a robot in Boston to actually put that robot to work in Gary. In fact, things like this happen all the time in America. Acemoglu and Restrepo (2017) note that the commuting zones with the greatest exposure to robots include Detroit, Lansing, Saginaw, Defiance, OH, Lorain, OH, Muncie, IN, Racine, WI, and Wilmington, DE.

Just to take one example of how different regions in America are dependent on each other, suppose that over the period from 1970 to 2020, Americans with lots of talent tended to move from Gary to Boston in order to take advantage of the increasing number of knowledge-industry jobs there. And suppose that at the same time, less talented Americans moved from Boston to Gary, in order to take advantage of the cheap land there. And suppose that high-value industries, like biotech and robotics and software, simultaneously decided to put their research labs in Boston rather than Gary because that’s where the best workers were headed.

In this case, you’d see Boston grow more than Gary, for reasons that had nothing to do with birth-rate-driven labor supply. Other research — for example, by Enrico Moretti — has shown that the sorting of talent and knowledge industries was strongly correlated with income divergences between American regions after 1970. Now, Acemoglu et al. do control for initial local education levels, but if sorting really began after 1970 (as Moretti has shown), that doesn’t really help. They also use “composition-adjusted” wages that try to model what local wages would be if education levels were held constant. But even if that adjustment is valid,9 it’s likely that a lot of skills aren’t captured by simply measuring who has a college degree.10

So in Acemoglu et al.’s telling, a substantial part of the reason the San Francisco Bay Area became Silicon Valley is not because the tech industry naturally tends to cluster in certain locations and draw in talented engineers and entrepreneurs, but because people in the Bay Area in 1940 didn’t have many kids, which forced Bay Area businesses to embrace automation instead. I simply don’t buy that story. I might eventually be persuaded to buy that story, but I don’t think this paper has sufficient evidence to persuade me yet.

Anyway, I am not claiming that Acemoglu et al. (2026) is a crappy paper. It’s clearly an important addition to the literature, and it deals with a very hard area of research where definitive results are very scarce. The basic idea is worth taking seriously. Maybe in a world where AI is growing by leaps and bounds, a scarcity of human workers should be the last thing on our minds!

But we shouldn’t regard this paper as definitive. And it contradicts a bunch of other research:

• Kremer (1993) finds evidence that regions with faster population growth had faster technological progress.

• The endogenous growth literature argues that because ideas are nonrival, larger populations are able to sustain faster technological progress — this is a basic assumption in Romer (1990), Jones (2022), and many other models.

• Maestas, Mullen, and Powell (2023) find that aging was negatively correlated with productivity growth among U.S. states from 1980 to 2010;11 Ozimek, DeAntonio, and Zandi (2018) find something similar.

• Aksoy et al. (2019) do a panel time-series regression across OECD countries from 1970 to 2014, and find that low fertility and low population growth are correlated with slower economic growth.

And even if Acemoglu et al. (2026) is right about the past, that doesn’t mean we can extrapolate it to the future. Jesús Fernández-Villaverde points out that you can’t necessarily do a linear extrapolation of effects at above-replacement fertility to effects at below-replacement fertility:

The fertility collapse now underway has no historical precedent. The seven decades of data the [Acemoglu et al. (2026)] paper analyzes, however carefully, contain nothing remotely comparable to the ultra-low fertility rates we observe today…

[T]he birth and fertility rates of 1950 provide little information about the consequences of the ultra-low birth and fertility rates we are experiencing today. Observations from a very different range of values are unlikely to capture the nonlinear cumulative effects of ultra-low birth and fertility rates…[E]ven in 1980, birth and fertility rates were not as low as they are today…A TFR of 1.56 means each generation is 76% the size of the last: the population declines about 0.9% per year, so after 100 years you retain roughly 40% of the initial population. A TFR of 0.75 means each generation is 37% of the last: about a 3.35% decline per year, leaving roughly 3.5% after a century. Hence, we are dealing with a factor of about 11. A TFR of 1.56 and a TFR of 0.75 both count as “low fertility” in a regression, but they are qualitatively different regimes. The 1980 variation the paper identifies lives in the gentle-decline world; nothing in their data resembles the second…

[W]e should not fool ourselves: we are sailing into demographic terra incognita, and in these waters, the pretense of knowledge is the most dangerous temptation.

In fact, Hayashi (2025) documents just such a nonlinear relationship between aging and growth, where countries with above-replacement fertility seem to benefit from lower birth rates, but countries with below-replacement fertility seem to do worse when their birth rates fall even further.12

External validity and nonlinearity are big problems in much of growth economics, development economics, economic history, and so on (and often in other fields too). But it means that there’s a danger, when the world’s top economist writes a paper making big bold claims, that we end up deferring to that individual’s intellectual authority instead of recognizing the true uncertainty of the unprecedented changes our world is facing. I’m still very worried about the shrinkage of the human race, and I think you should probably still be worried too.

Subscribe now

Share

There’s actually a tradition of economics bloggers giving recommendations about where (and how) to eat out. Tyler Cowen wrote a whole book on the subject, as well as various blog posts, and maintains an Ethnic Dining Guide on the web! I’m not going to be that thorough or go that deep. But over the last year I’ve gone to quite a few restaurants in San Francisco, so I thought I might share what I’ve learned.

In fact, this quest was born out of contrarianism. A lot of people say that the restaurant scene in San Francisco is mediocre, boring, or disappointing. The haters range from food critics and New York chefs to some SF residents themselves. I’ve lived in this city for a decade now, and although the restaurant scene here isn’t the best I’ve found — NYC and LA are better, and Japanese cities just annihilate all competition — I’ve always found plenty of good places to eat. When people portray SF as some sort of food desert, I’m convinced it’s a skill issue.

So to be thorough, I set out to try all of the good restaurants in the entire city. That’s a project that’s bound to fail, of course — there are just too many, and quality can vary over time, and places are always opening and closing. But at this point I think I’ve eaten at enough places to write a fairly authoritative guide to restaurants in San Francisco.

What this guide is and is not

Writing about food is an inherently difficult thing to do, because taste doesn’t easily translate into language. There are plenty of reasons for this, including:

• People don’t biologically experience the same taste the same way. The most famous example is how some people taste cilantro like soap due to a variation in one particular gene. But there are lots of other examples. When I try to describe the taste of a particular dish or ingredient, chances are that your physical experience of it will be different from what I experienced, and you’ll be left scratching your head and wondering if I’m an alien.

• Even if people did experience tastes the same way, people’s preferences would probably differ radically on most things. Some of this is due to personal associations — conditioned taste aversion is very much a thing, for example. There are probably lots of other reasons as well. It isn’t just what we like and dislike, either — how much we like and dislike each combination of tastes can make the difference in whether we like a restaurant overall.

• Culinary experiences are not very uniform across the population, so we haven’t been able to develop as much of a shared vocabulary to describe what we’re tasting — if I’ve eaten Sichuan pepper and you haven’t, it’s a little like trying to describe the color red to a blind person. And since most people have only eaten a small percentage of what there is to eat in the world, the region of overlapping experience across the entire population is small; thus, we simply don’t have many words that refer to taste experiences that most people have had.1

To see how hard it is to describe food in words, try the following exercise. Pick five restaurants you like, and five you don’t like, and mix them together in a list. Ask your favorite AI chatbot to pick five of these that you should go to, and five you should avoid, and ask it to explain why you’d like and dislike them. Chances are you’ll have no idea what it’s talking about. AI can do the most advanced math on Earth, but describing taste in words is an exercise for which good training data simply doesn’t exist.

So when I say my restaurant guide is “authoritative”, I don’t mean I’m an authority. I just mean I’ve been to a lot of restaurants, and I know which ones I like and don’t like, and I can describe why I like and dislike them in terms that make sense to me. But obviously this is all just me describing my subjective experience and hoping that it resonates with you. So when you go through this guide and find yourself nodding in vigorous agreement with some of the recommendations, and screwing up your face in confusion at others, well…I warned you.

Also, I’m not well-versed in the art of food criticism, so I’m not going to try very hard to describe the tastes I experienced at these restaurants. But I am pretty well-versed in the art of noticing patterns and thinking about explanations for things, and so this guide is going to focus less on the particulars of various dishes than on questions like “Why is SF’s food scene not the best in the country?” and “What kinds of restaurants are generally good in SF?”.

As for my recommendations themselves, my criterion for ranking restaurants highly isn’t just how good they are. It’s what I call “yums per dollar”2 — basically, whether the culinary experience you got was worth the price you paid. So I’m perfectly happy ranking a good taqueria up there with a Michelin-star place.

Speaking of Michelin-star places, I’ve decided not to include any places that are over $150 a person or so (before tax and tip). That’s already quite expensive, and I don’t want to be the kind of guy who recommends people to a bunch of restaurants they can’t afford to go to. And quite frankly, I’ve only been to one or two of those anyway. So, apologies to Benu, Quince, Californios, and the rest of those guys.

Finally, I’ve eaten at several hundred restaurants in SF, but that’s only about 10% of the total. So if you have some recommendations for good ones I should try, feel free to drop them in the comments. And of course this post is just Part 1 in a series; Part 2 will have updated recommendations, cafes, bakeries, dessert places, and some restaurants from out of town.

Why is SF’s food scene not the best in the country?

First, I thought I should address the big question: Why isn’t San Francisco’s food scene better than it is? I’ve already gone on record saying that NYC and LA are better (though the difference isn’t huge). Why?

There’s one obvious, overriding reason: city size. San Francisco is a city of 826,000 people; New York City has 8.58 million, while LA has 3.87 million. So if the number of restaurants is generally proportional to the number of people, then even if quality were equivalent, we’d expect NYC to have more than ten times as many good restaurants as SF, and LA to have more than four times as many.

This would lead very naturally to people thinking SF has a worse food scene than those much larger cities. The restaurants we go to are not a random sample; we try to pick the ones we’ll like the most. So if you can find 10 times as many restaurants you like in NYC, you might end up thinking restaurants in NYC are better on average, even if they’re not. (Also, larger populations mean NYC and LA have more partisans to stand up for their hometowns in online arguments.)

The people who bash San Francisco restaurants are often quick to explain that there’s plenty of great food in the rest of the Bay Area — in the East Bay (Oakland, Berkeley, etc.), the Peninsula (South SF, Daly City, Redwood City, Mountain View), and in the South Bay (Milpitas, Sunnyvale, Cupertino, San Jose). And this is very true. But keep in mind that the Bay Area has a total of nine million people — a little bit more than New York City. I would gladly pit the entire SF Bay Area against all of NYC in terms of food.

OK, but San Francisco is by far the densest part of the Bay Area, in terms of population per square mile. Why isn’t it considered the center of the metro area’s culinary scene, the way Manhattan is for NYC? Why do people tell you to go to the South Bay or the East Bay if you want to eat well?

I think there are several reasons for this. The first is simply the insane rent. SF is one of the least affordable cities in the country, thanks to the AI boom, the tech industry in general, draconian restrictions on housing development, the unique and excellent weather,3 and the fact that it’s destined by geography to be the dense core of a major metro area. On top of that, San Francisco is one of the hardest places to start a small business, thanks to decades of red tape accumulated under supposedly “progressive” governance.

That means if you’re going to start a restaurant in San Francisco, you are probably going to have to charge a lot of money to make up for all the money and time you had to spend. The exceptions are restaurateurs who own their own buildings, who have long-term legacy leases, or who have owners who rent to them for below market rates. This means that SF food tends to be a sort of double-peaked distribution, with $600-a-person places like Benu and cheap banh mi places like L&G, but not a lot of mid-price places.

So if you’re a penniless 22-year-old tech founder living in a bunk bed in a grungy group house, or if you’re a 44-year-old tech founder who pocketed $80 million from selling his company to Amazon and now just “invests”, you’re more likely to be satisfied with the SF food scene. But if you’re a 30-year-old who makes $200,000 working for a big tech company, there are going to be fewer of the kind of mid-priced options that you could afford but which would also give you a special night out.

SF’s stratospheric rent also probably exerts a more subtle corrosive influence on the food scene. The middle class has largely been priced out of the city, leaving it to the rich and the rent-controlled. That has largely hollowed out the kind of middle-class ethnic communities that patronize the great ethnic food scenes of Los Angeles, as well as the bohemian hipsters and artists who seek out the most creative affordable eating options.

Everybody knows about SF’s rent problem, but I think one more factor that tends to fly under the radar is limited competition due to geographic immobility. New York has a great train system, making it pretty easy to get anywhere in the central city on foot. LA and all other American cities (including Oakland and San Jose) are driving towns, where you can often get very quickly from point to point. But San Francisco does not have a good intra-city train network (sorry, Muni!), nor is it car-friendly. You can Uber, but that gets expensive if you do it a lot. So in my experience, a lot of San Franciscans often end up just staying in their neighborhoods.

When people stay in their neighborhoods, restaurants aren’t competing with each other nearly as much; everyone just goes to their local spots. This reduces the incentive for restaurants to innovate, raise their quality, and differentiate themselves — there’s just not as much chance to attract patrons from across town. Each restaurant instead just serves up its classic fare to the standard crowd of locals, and they have to accept whatever they’re given. And when you ask people for recommendations, you just get their standard neighborhood spots (or a recitation of the Michelin Guide).

So if you want to make San Francisco’s food scene better, my suggestions would be:

1. Build a lot more housing

2. Make the Muni a lot faster and improve its coverage of the city

3. Cut red tape and make it a lot easier to start a small business

4. Allow and encourage multi-level retail

But having said all that, I do think that San Francisco’s restaurant scene is decently good despite all of these challenges. So here are some recommendations of places to go.

Three types of restaurants I tend to like in San Francisco

There are good examples of practically every type of restaurant in San Francisco. But I’ve noticed three distinct types that I think SF does particularly well.

Type 1: Pseudo-izakayas

An izakaya is a type of Japanese restaurant, similar to a tapas bar. You typically order a variety of small plates and share them among a group of people, often while drinking. Traditionally, izakayas mostly served old Japanese standby dishes — karaage, basic yakitori and sushi, that sort of thing. But around the 1990s, a bunch of izakayas started getting more experimental and creative. The small-plate format allowed them to do this, because it let diners order one weird new thing among a bunch of old favorites. This trend accelerated as the Japanese government lowered trade barriers and allowed more cheap imports of foreign ingredients. By the 2010s, most of the best restaurants in Japan were izakayas rather than omakase, ramen, etc.

San Francisco doesn’t have many (well, really, any) of these new, upscale, creative izakayas. But it has a number of restaurants that basically do this same format with some other type of cuisine — Middle Eastern, or Mexican, etc. The small-plate format allows these restaurants to charge a little more — they’ll usually run about $70 to $110 per person — but they compensate for this by giving diners enormous variety and experimenting with lots of new dishes.

These places sometimes get described as “fusion”, but their dishes aren’t mashups — instead, they start from a base of a specific type of cuisine and then play around with it, throwing in a couple of ingredients from other cuisines. In fact, one good way to identify good restaurants in the U.S. these days is to look for places that use one or two Japanese and Korean ingredients in dishes that aren’t Japanese or Korean.

By the way, one of my most important tips for eating out is: ALWAYS EAT FAMILY STYLE. That means share everything, unless you have someone with special dietary restrictions with you. Sharing everything means you get to try a lot more dishes, which is especially important at the kind of izakaya-style places I talk about in this section. It also means you get to constantly cleanse your palate — when you go back and forth between a bunch of different dishes, you keep recapturing some of the novelty of each dish every time you come back to it. Whereas if you just order your own thing and chow down on that one thing, you hit the law of diminishing utility.

Anyway, here are some places I like that fit this description:

Maria Isabel: Officially a Mexican restaurant, but really they play around and mix in a bunch of different ingredients from a bunch of different cuisines. Every single thing I had here was excellent, and this might be the best new restaurant I’ve discovered this year. The plates are small, so order a lot; it’s only a little expensive. They also have a tasting menu, so technically they’re also in the “hipster fine dining” category. They’re probably going to get a Michelin star and jack up their prices, so go soon.

Prik Hom: In terms of sheer deliciousness, this is my favorite Thai restaurant in a city with lots of good Thai restaurants (though it’s more expensive than the close runner-up, Khao Tiew). The dishes are experimental riffs on standard Thai favorites, and the menu rotates frequently, but everything is delicious so don’t worry you’ll order the wrong thing. (Make sure to eat the ice cream at the end, though.) Prik Hom is a hidden gem for now — it’s in a neighborhood without a lot of other restaurants, and you can often just walk in.

Aziza: Everyone knows about this place, and it’s as good as they say. I’d say it’s tied with Dalida for my favorite Middle Eastern food in the city. But it’s not traditional Middle Eastern — I had one of the best octopus dishes I’ve ever eaten at Aziza, which is saying something because I lived in Japan. The menu rotates frequently, but you can always get the “trio of spreads” to start. There’s also a tasting menu, so this fits in the “hipster fine dining” category too.

Dalida: The other great Middle Eastern izakaya I know in SF. This is Maria Isabel’s sister restaurant (the chefs at the two places are spouses), and it’s comparable in quality. Dalida is also one of those rare restaurants that’s good for brunch and dinner, so try both (and if you go for brunch, sit outside). Dalida doesn’t include as many ingredients from other cuisines as Aziza does, but everything is a little experimental and unusual anyway. There is also a tasting menu for dinner.

Bansang: I’m kind of mad at Bansang, because it replaced what was by far my favorite Japanese izakaya in San Francisco, Izakaya Kou. But I forgive the owners, because they replaced Kou with one of the only really excellent Korean restaurants in San Francisco. Like the other “pseudo-izakayas” on this list, Bansang focuses on smallish plates with original twists on old Korean favorites (chimichurri in the galbi, etc.), and they frequently rotate their menu.

Pearl 6101: You don’t see a lot of experimentation with Italian food, but Pearl is the best example of an “Italian izakaya” I’ve been to in SF. They don’t use as many Asian or Middle Eastern ingredients as some of the other restaurants on this list, but they’ll sneak a little Vietnamese fish sauce into their halibut crudo. Everything is very well-executed.

Horsefeather: A restaurant disguised as a bar. This is one of the best-kept secrets in San Francisco, since most people think it’s just a place to drink, but it actually has a very extensive food menu full of creative and excellent cuisine that’s basically a mashup of French, Latin American, Korean, Japanese, and Italian. I had some of the best asparagus of my life here, and an unbelievable tres leches cake — both of which have now been taken off the menu, because it rotates frequently. A great place for large parties on short notice. Horsefeather is also co-owned by Hanson Li, one of SF’s most creative and interesting restaurateurs.

Fermentation Lab: The single most underrated, unknown restaurant in this category. It’s another restaurant disguised as a bar, like Horsefeather. Fermentation Lab is basically a Korean izakaya, serving creative Korean twists like makegeolli-braised pork belly in addition to standard bar bites like duck fat fries. Everything is delicious and original. I cannot stress how bad the marketing of this restaurant has been, relative to its quality — people ought to be snapping up reservations weeks in advance, but instead there are open seats and you can just walk in. People don’t even seem to know it’s Korean!

Tiya: An Indian izakaya with some Japanese influences. If you want to get wagyu stir-fried with asparagus, madras curry, and quinoa, or octopus curry with smoked tomato pachadi and kumquat, this is the place to get it. It’s very experimental, so if your Indian friends are looking for a taste of home, consider taking them to Copra instead. But Tiya has stuff you won’t get anywhere else, and it’s easily my favorite restaurant in the Marina. It’s also open for brunch, and it has a tasting menu option for dinner, though I haven’t tried either of those.

Rich Table: “New American” is a term that sometimes describes overpriced faux-French places, and sometimes describes extremely creative izakaya-type places that blend traditional American fare with ingredients from Japan, Korea, and other countries in Asia. Rich Table is probably the best of these in the city, unless you count the Michelin-star places — and on “yums per dollar”, it beats most of those places hands down. In fact, Rich Table is probably the closest thing you can get in SF to a good modern/experimental Japanese izakaya — the two cuisines are basically converging. It also has a tasting menu.

Other great examples:

Routier (probably my favorite French restaurant in town, everything tastes interesting and a bit unusual, but almost always better than the standard version4)

Ernest (another “New American” place with lots of contributions from other cuisines, a bit like Rich Table; get the cheesecake)

The Morris (yet another great New American place, I recommend the crudite)

Four Kings (a Cantonese izakaya, everything very solid but could stand to be a little more experimental, I recommend the squab)

Happy Crane (another Cantonese izakaya, very new and experimental, still working out some kinks and can be hit or miss, but the hits are excellent)

Jilli (another Korean bar-style restaurant, the menu is much smaller than the others but everything is excellent and the flavors are extremely original; try the kimchi vodka rigatoni pasta with burrata)

Red Window (warning: order only the Spanish dishes, not the Italian ones!)

Phonobar (a vegan izakaya masquerading as a bar)

Type 2: Hipster fine dining

Most “fine dining” places in SF range from $200 to $600 per person, and come complete with Michelin stars and vigorous arguments over whether they’re “the best meal I’ve ever had” or “totally overrated”. But there are a few places you can go that offer a “fine dining” experience for a much more affordable price — usually around $90-$120 per person, but occasionally as low as $50.

There is some overlap with the “pseudo-izakaya” category here, because many of those places offer tasting menus. Rich Table, Maria Isabel, Aziza, Dalida, Ernest, and Happy Crane all count as “hipster fine dining” too. But here are a few others:

Mijoté: This might just be my favorite restaurant in the city. It’s officially a French restaurant — the “j” is pronounced like a “j”, not like an “h” — but it has very heavy Japanese influences. That’s probably because the chef, Kosuke Tada, is a Japanese guy who worked in France for a long time. Everything at Mijote is excellent; it’s only a prix fixe menu, and it changes frequently, but I’ve never once been disappointed with anything. Kosuke is a virtuoso; one time he catered deli sandwiches for a random party I went to in Oakland, and they were some of the best sandwiches I’ve ever had. And the price is very affordable as fine dining goes — just $84 per person.

7 Adams: This place does have one Michelin star (and richly deserves it), but it’s significantly cheaper than a lot of other Michelin places while being every bit as good. It’s “New American”, with somewhat fewer Asian ingredients than Rich Table.

Mr. Pollo: Still among the best-kept secrets in SF, even though it’s been around for many years. Two guys took over a space from a Colombian restaurant called Mr. Pollo, but kept the sign and the name. It’s not Colombian food (they used to do arepas every night, but stopped); it’s just whatever they feel like cooking this week. It’s always good, and it’s just $49 — maybe the best value for money in the entire city. Good luck getting a reservation, though.

Type 3: New hole in the wall

There are a number of cheap-ish restaurants that luckily manage to avoid the curse of SF’s astronomical rent. Most of these have just been sitting in the same spot since the 80s or 90s, and haven’t really updated their cuisines. However, there are some new ones that manage to open up, and some of these — through some mystical combination of skill, dedication, luck, or whatever — manage to be truly excellent. Here are a few examples:

Sichuan Tasty Restaurant: One of the most underrated restaurants in the city. I love Sichuan food, and though I tell myself I’ve been to one or two places in Flushing that were better than Sichuan Tasty Restaurant, I can’t quite remember if they really were better. Everything is delicious, the portions are huge, and somehow there are always seats open and you can just walk in. Remember to order everything with extra 花椒.

Chef Xiong Kitchen: Another incredible hidden gem. This place mainly serves Sichuan food, and its quality rivals Sichuan Tasty (which is just a few blocks away). But it’s incredibly versatile — it also serves Mongolian BBQ, and a bunch of other random Chinese dishes, including a few I’d never seen before. Everything is excellent, you can’t go wrong. They also stay open until 2 AM, in case you want to go out late at night!

Mini Potstickers: If you want great takeout dumplings in SF, this, in my opinion, is one of the two best places to get them. The boiled dumplings are as good as the nearby (and much more famous) Yuanbao Jiaozi, but Mini Potstickers has much more range — awesome mini fried dumplings (hence the name), and excellent jianbing.

Fuwa Dumpling: SF’s other great takeout dumpling place, also with good variety. Probably the best takeout boiled dumplings I’ve had in the city.

Gusto Pinsa Romana: Some restaurants have started aggressively marketing traditional Roman-style pizza as “pinsa”, and defined it as a whole new category of food. Well, more power to them. Of “pinsa” restaurants in SF, Pinsa Rossa is the most famous, but Gusto Pinsa Romana is — at least in my opinion — far and away the best. It’s cheap, it’s delicious, and it’s good to eat in or DoorDash.

Timur: As of this writing, Timur Indian and Nepalese Cuisine has a perfect 5.0 rating on Google with 688 reviews. That’s just incredible. It’s well-deserved — this small, out-of-the-way, affordable little restaurant is the best Nepalese food in a city filled with great Nepalese places. I especially recommend the korma.

Sherpa House: This one just barely gets edged out by Timur for “best Nepalese in town”, but honestly it’s very very close. Also a perfect 5.0 on Google as of this writing, with 202 reviews.

CQ Noodles: If you want some great spicy Sichuan noodles, this is your place. Conveniently located in the Outer Richmond, very close to Sichuan Tasty and Chef Xiong, the other two best Sichuan places in town.

Bo&Beurre: How would you like a huge, incredibly delicious banh mi sandwich in the middle of downtown San Francisco for only $6? Bo&Beurre is a little out-of-the-way Vietnamese place that will give you that sandwich. Somehow it’s much less well known than Saigon Sandwich, which is less than a block away and has lines down the street, but which is definitely not as good.

L&G: The other amazing super-cheap banh mi place in SF’s Little Saigon neighborhood is L&G — and it’s right next to Bo&Beurre. Actually, the two places are somewhat different — Bo’s sandwiches are drier and perfectly arranged, while L&G’s are a greasy street-food sauce-bomb. Both are amazing, both have no lines, and I’d recommend both over Saigon Sandwich.

Oodle Yunnan Rice Noodle: I am always looking for Chinese restaurants that are neither Cantonese, Sichuan, nor dumplings. This one fit the bill nicely. The food is delicious, the prices are affordable, the decor and ambiance are very cheerful and modern, the staff is super friendly, and you can just walk in.

El Gallo Giro Food Truck: I’m sort of stretching the boundaries of this category to include a food truck that has been parked in the Mission for over two decades, but it would be criminal not to. This little truck has some of the best tacos and burritos in the city. I don’t know how they do it so well. You need to pay with cash or Venmo, so be prepared.

Anyway, there’s a sampling of my three favorite types of restaurants in San Francisco. But there are a bunch more good ones, so let’s go through some of the standard categories.

Cuisine-specific recommendations

Japanese

I spent a lot of time raving about how great izakayas were, but ever since Izakaya Kou closed, there aren’t really any great ones in San Francisco. Kibatsu in Lower Haight and Fenikkusu in the Mission are the closest — both are very solid, but they ought to get more creative and update their menus more. Kibatsu has to be my favorite here, because they also have some great nigiri, including an umimasu that’s the best I’ve ever had.5 Noren also has good quality yakitori.

However, there are a number of very good Japanese restaurants of other, more standard types. Two ramen places — Taishoken and Mensho SF — really dominate the scene here; both have skilled Japanese chefs, and are similar in quality to what you’d get at very good ramen restaurants in Tokyo. I’d give the edge to Taishoken, since it takes reservations and doesn’t require you to wait an hour on the street in the middle of SF’s worst neighborhood like Mensho does.

There’s a kaitenzushi place called Izumi in Japantown mall that’s actually as good as any I’ve had in Japan (though much more expensive). Nabe is a Japanese hot pot place in the Sunset that also matches Japanese quality. There are also some “big roll” sushi places of the type you can’t get in Japan — my favorite ones like Arashi Sushi closed in the pandemic, but Rocket Sushi and DJ Sushi’s big rolls are criminally underrated and there are a bunch of new ones I haven’t tried yet. Potto and Shabu Club are also solid hot pot places with good-quality meat.

Omakase is difficult, because America has trouble getting fresh fish — both because of stricter health laws, and because of geographic distance from Japanese supply chains — and most of the famous places are just way overpriced. But there are a couple very solid ones, including Doma Sushi in Bernal, and Robin in Hayes Valley. Robin is a controversial pick here, because it’s non-traditional and experimental, but their experiments hit a lot more often than they miss, and if you’re in a country that has trouble getting fresh fish, experimentation is more important. An is also good, though for some reason it takes 3 hours to eat.

Vietnamese

San Francisco has a lot of great cheap Vietnamese food, especially in the “Little Saigon” area near the Tenderloin. I’ve already mentioned Bo&Beurre and L&G, but also check out Pho Tan Hoa and Larkin Restaurant for pho. Pho 2000 isn’t my favorite for Pho, ironically, but has very good appetizers and other dishes. But the best Vietnamese restaurant in the whole area is probably the amazing Hai Ky Mi Gia, a little Chinese-Vietnamese hole-in-the-wall that serves some of the best seafood egg noodles I’ve ever had. There are also plenty of good cheap and cheap-ish places scattered throughout the city, like Banh Mi Crunch in the Sunset, An Chi in the Richmond, and others. Banh Mi King is a good option if you’re downtown, though it’s a bit pricier.

As for upscale Vietnamese, that’s harder to find, partly because “upscale Vietnamese” is just a thing that not a lot of people try yet. My favorite is definitely Gao Viet in the Sunset, which serves absolutely huge portions of absolutely delicious food, including both traditional favorites like banh xeo and some truly wacky giant combos of stuff. Thanh Long makes a good crab, though it’s a little bit overpriced. Bodega SF is pretty good, though a bit uneven. There is also a place called Hem by LeQuy in the Mission that does traditional stuff, but slightly more expensive and higher-quality than most; it probably has the single best bowl of pho I’ve eaten in SF. Note that I still have a lot of Vietnamese places left to try in this city.

Nepalese

I’m not sure why, but there has been an explosion of good Nepalese places in San Francisco in recent years — most of them pretty affordable. Tyler Cowen has a rule that you should try the less-well-known ethnic cuisines, and that definitely applies to Nepalese in SF; the Nepalese places here tend to be much better than the similar Indian places. The three best I’ve found are the aforementioned Timur and Sherpa House, plus a third place called Base Camp that does excellent lunch and focuses more on noodle dishes than curries. But there are a lot of other very good ones, including Himalayan Cuisine, Dancing Yak, Nepa Indian Cuisine, and others.

Thai

For some reason, San Francisco has a lot of very good Thai restaurants. Unlike Vietnamese places, which tend to be cheap, Thai places tend to be mid-priced or even a little upscale. So they make good spots for date nights or big group outings.

The yummiest Thai place I’ve been to is the aforementioned Prik Hom, but in terms of yums per dollar, it — and every other Thai place — is absolutely mogged6 by Khao Tiew in West Portal. You can absolutely gorge yourself sick on amazing food at Khao Tiew for $30. But the downside is that the place doesn’t take reservations, so you have to wait two or three hours in line just to get in.7

Nor are those the only great places in town. Ka Kai in the Castro has an amazing sampler platter, and the best khao soi I’ve ever eaten. Nari has an absolutely incredible trout curry — probably the best Thai curry I’ve ever had — and is more affordable than Google says it is.8 There are a number of other solid ones as well.

Mexican

People from Los Angeles love to talk trash about how their Mexican food is so much better than San Francisco’s Mexican food. It isn’t. The taquerias in the Mission are as good as what you’ll get in LA, and the “Mission burrito” is a genuine local specialty.

What is the best Mission burrito in SF? This is a long-running debate, but my favorites are Taqueria Buen Sabor, El Metate, the venerable El Farolito, and the aforementioned El Gallo Giro Food Truck.9

As for tacos, for my top picks I of course recommend El Gallo Giro, and a nearby place called Taqueria Vallarta. Taqueria Los Altos is very good as well. There’s also a place called El Rey Taquiza Artesanal that’s excellent, though more expensive.

SF also has some good upscale Mexican places, especially the brilliant Maria Isabel and the delicious brunch spot el Mil Amores. There are a lot I haven’t been to yet, so I’ll report on those in Part 2. (SF also has the country’s only 3-Michelin-star Mexican place, Californios, though I haven’t been.)

New American

I’ve already covered most of my favorite New American places — Horsefeather, Rich Table, 7 Adams, Ernest, and The Morris. Another place I like a lot is Verjus, and side a can be a little uneven but has one of the best salads I’ve ever eaten (the Garbage Salad). Golden Eye Social, Arquet, and Plow are solid too, though Plow makes you stand in line. The chicken and fries at Zuni Cafe are as good as everyone says, so that’s worth a visit.

Italian

Like NYC and many other American cities, San Francisco has a lot of “red sauce” Italian places — traditional Sicilian-type places where you can get lasagna and spaghetti and infinite free bread. Those have never really been my thing, but I do have to say that Sotto Mare’s famous cioppino is really very good.10

But there are also some good mid-priced Italian places in town. Everyone says Cotogna is the best of these, and they’re correct; they had a corn pasta recently that’s the best pasta I’ve ever eaten. Seven Hills gets a lot of grief for being old-fashioned, but the quality is as good as what I’ve had in NYC, for a significantly more affordable price. Flour + Water is also quite solid, though I’d get the pasta instead of the pizza. Pearl 6101 is the most creative.

I’m not sure if pizza counts as Italian food at this point, but SF does have great pizza. Little Star is my favorite, with Chicago-quality deep dish and excellent Detroit-style pizza as well. Outta Sight is probably the best place to order for parties. Gusto Pinsa Romana is delicious, unusual, and surprisingly healthy. I also like Sforno and Back to Back a lot.

SF can’t hold its own against NYC (or Japan) in terms of Italian food, but it’s not hard to find good stuff here.

Chinese

Chinese food is undoubtedly San Francisco’s biggest culinary failing. Given the huge ethnic community — almost a quarter of the city’s population is of Chinese descent! — it should have a ton of great Chinese restaurants. Instead, it has endless mediocre Cantonese banquet places that look like they haven’t changed a thing (including their signage and decor) since 1982. When regular people (as opposed to snobby NYC chefs) say “food in SF sucks”, they often mean that the Chinese food sucks.

It does not all suck. There are some solid high-end places — Four Kings and Happy Crane are good “pseudo-izakaya” style places. Mister Jiu’s and Empress by Boon are solid Cantonese banquet places that aren’t stuck in 1982. There are also a few great hole-in-the-wall places, like Chef Xiong Kitchen, Sichuan Tasty Restaurant, Oodle Yunnan Rice Noodle, Mini Potstickers, Fuwa Dumpling, CQ Noodles, and Hai Ky Mi Gia.

But these are few and far between. And in terms of mid-price restaurants — the kind you’d be blown away by in Milpitas, Cupertino, or Daly City — there really just aren’t that many. Dumpling House in the Castro is a sleeper hit in my opinion, with the best green beans I’ve ever had, plus all-around great dumplings. Bao Bao in the Richmond is another underappreciated gem. Tasty Pot is a great individualized Taiwanese hot pot place, though some people (mistakenly) look down on it. There are a few solid hotpot places like IPOT, and a few solid dim sum places like Dragon Beaux and Palette Tea House. Z&Y is OK, I suppose. But honestly, it’s slim pickings, and if you want good Chinese food I’d usually just recommend taking a trip across the Bay Bridge or down to the south suburbs, which IMO rival Greater Los Angeles and surpass NYC in quality.

Others

There’s lots of other good stuff in San Francisco. SF has some great lunch sandwich places — the best is Tarragon Cafe, which is one of the city’s best-kept secrets, and where everything is delicious (I recommend ordering the salmon toast with avocado). Tartine Manufactory is great too. For Greek food, go to Kokkari for the higher end and The Parthenon for the lower end. For German, go to Suppenküche in Hayes. For brunch, Plain Jane and its sister restaurant Mymy are hard to beat, though there are lots of great places like Les Marais and Lapisara Eatery. Beit Rima is another solid Middle Eastern place. For pubs, I like Hazie’s and Stoa. Bar Jabroni is a wine bar that serves excellent food as well. Hook Fish has excellent fish sandwiches and such. Burma Love and its sister restaurant Teakwood are both yummy (get the duck curry at Teakwood).11 Brenda’s does have good cajun and soul food. For vegan food, I really love both Phonobar (which is basically a vegan izakaya), and Shizen, the vegan sushi spot. House of Prime Rib deserves its status as a local favorite, though it’s a LOT of food. JouJou is a solid (if a little uneven) new French spot. Piglet & Co. is a Taiwanese izakaya-style place that serves a lot of interesting pork dishes. Niku is an excellent steakhouse with great side dishes and a butcher shop next door that does great wagyu burgers.

I think that just about does it for my specific recommendations. That’s probably enough restaurants to keep you eating well for a couple of years!

Overrated restaurants?

When I give my recommendations, one thing some people ask me is to identify some restaurants that I think are overrated — they want to be warned away from bad restaurants, not just pointed toward good ones!

That’s a reasonable thing to want, but in general I’m kind of against this. As I said, tastes differ a lot — if you like a restaurant I don’t like, then good for you and good for the restaurant too! I don’t expect my preferences to be universally shared, and I certainly don’t regard them as the voice of authority.

But that said, knowing what I don’t like might help you to understand whether and how your preferences match or diverge from my own. Just remember that I’m not criticizing anyone for liking things I don’t like.

First, there are some restaurants I just really did not like at all. These include Lily, a Vietnamese place in the Richmond; La Ciccia, an Italian place in Bernal; and Bir Kitchen, a Nepalese place in the Mission. For whatever reason, the food at these places just really did not appeal to me. If it appeals to you, more power to you.

Second, there are a bunch of restaurants that form a sort of “standard canon” of what people in the tech scene — i.e., my own social scene — tend to recommend, but which I think are no better than similar places that receive much less attention. One of these is La Taqueria, which serves fine burritos but which in my opinion doesn’t justify the massive line around the block — especially when you have El Metate and El Gallo Giro so close by. Saigon Sandwich is the same thing for banh mi — Bo&Beurre and L&G are literally across the street with no line. I don’t understand why people form a massive line for pizza at Jules or dumplings at Dumpling Home, when Gioia and Dumpling House are similarly good and just a few minutes’ walk away.

Most of all, I can’t understand why everyone always wants to go to Izakaya Rintaro. I’ve been there five or six times now, but you can get pretty much the same sort of mid-level 1990s-era izakaya food at Nande-ya in Japantown, or Izakaya Sozai in the Sunset, or a number of other places — and honestly better, if you go to The Public Izakaya or Chome.

I once heard someone say “the line is the product,” and I guess maybe that’s right. Perhaps there is some level of endogenous coordination, where the fact that everyone else is excited for a particular place makes it more exciting to go there, even when it’s not very different from other places.

Anyway.

Neighborhoods

That about sums it up for my recommendations, but as I said, I’ve been to only 10% of the restaurants in San Francisco. If you want to come to SF and wander around and find places that you didn’t hear about in this guide, or the Michelin Guide, or on the standard “where to eat” websites, where do you go? I thought it would be good to offer some general thoughts on the various neighborhoods in the city.

The Mission: The Mission is, far and away, the best eating neighborhood in San Francisco. This is partly a function of size, since it’s larger than other centrally located neighborhoods. But it’s also very central, and it’s a place that lots of people come to hang out and drink and shop and party. That means it’s a little less burdened by SF’s geographic immobility — a lot of people are always in the Mission, and so restaurants are forced to compete for their business. That’s why the Mission has lots of good top-end places like Mijote, Flour + Water, Ernest, The Morris, Shizen, and so on, along with great mid-price offerings like Little Star, Base Camp, and Taishoken.

But the Mission also has a thriving Mexican community and lots of buildings with preferential rent arrangements or legacy leases, which is why you get lots of good cheap Mexican places too. Basically, if you walk around in the Mission, you’re going to find a lot of good stuff.

Japantown: Japantown has a lot of OK legacy Japanese places — izakayas and noodle spots and yakitori places that are still basically the same as they were in the 90s. Those are fine, I suppose. But the high level of foot traffic opens up room for a lot of genuinely great spots to open. It’s the city’s center for Korean food, with Bansang, Fermentation Lab, and Daeho Kalbijim, as well as a good Korean-run katsu spot called Ganji.

North Beach: Again, I recommend not going here for the standard local thing (red-sauce Italian). There are a lot of other random good places here, drawn by the heavy foot traffic — Noren for yakitori, Tasty Pot for hotpot, Red Window for tapas, Verjus for New American, and so on.

The Central/Outer Richmond: Probably the best place in the city to find good hole-in-the-wall spots, especially Chinese ones. I need to try a lot more places (and will report back in Part II).

The Sunset: Another good place to go hunting for hole-in-the-wall and mid-price spots. It’s very diverse in terms of cuisine. Both the Inner and Central Sunset are good.

Hayes Valley: My own neighborhood! It has a reputation for being full of yuppies, which is honestly pretty deserved, but it turns out that yuppies want to eat at good restaurants too. The good places here tend to be mid-priced spots like Teakwood, Phonobar, Hazie’s, Dumpling Home, and so on, making it one of the few neighborhoods in SF where you can consistently eat well for $40 a person.

Russian Hill/Polk St.: I haven’t been to enough places here, but it’s probably the best neighborhood for Italian food in the city. It also has a few Nepalese spots toward the south end of the neighborhood.

The Tenderloin: If you’re willing to brave the drug-induced disorder of the Loin, you can eat some extremely good cheap Vietnamese food in Little Saigon.

The Inner Richmond: This is a surprisingly overrated neighborhood. There are tons of restaurants, many with great reputations, but very few ones I’d put at the top of my list. I’m not sure why; maybe it’s central enough to have high rents, but peripheral enough not to have the heavy foot traffic that lures good places to the Mission, Japantown, Hayes, and North Beach.

Chinatown: Chinatown is mostly dominated by incredibly mediocre antiquated places that are still hanging on to their ancient cheap leases. There are a few good (expensive) Chinese places here, and a few good little bakeries like Yummy Bakery on Jackson, but in general this is SF’s most disappointing eating neighborhood.

The Embarcadero/FiDi: This is the neighborhood I’ve eaten in least, and I plan to review it in greater detail in Part 2. There are some solid famous places along the waterfront — Angler, Hog Island, La Mar, and so on. Kokkari is great. But overall I haven’t eaten here much.

SoMa: San Francisco’s great food desert. It was once the center of SF’s bustling daytime office culture, meaning that most of the restaurants were either lunch takeaway spots or happy hour bars. Then Covid hit, and everyone left downtown, and the takeaway spots and happy hour bars closed down. Now workers are slowly coming back to the area, but it’ll take a while for the neighborhood restaurant scene to recover, and there are only a couple of good places left.

Others: West Portal/Stonestown/Parkside, Dogpatch, Potrero, the Design District, and the Wharf area all have a few good spots, but aren’t really places to walk around and discover good new places. The Castro is OK, and is slowly getting some good new spots like Ka Kai and Potto. Noe Valley is surprisingly sparse, given the number of aging yuppies and well-to-do young families who live there. I haven’t been to Bernal enough. The Marina is dominated by bars, being more of a place to drink than a place to eat.

Anyway, that’s about all I have to say so far about eating out in San Francisco. In about a year, I’ll circle back with Part 2! Happy munching!

Subscribe now

Share

Zhu Rongji has passed away at the age of 97. He was not the only architect of the economic reforms that have made China a modern powerhouse — Deng Xiaoping, Zhou Enlai, Chen Yun, and others were all important — but Zhu was perhaps the most skilled, versatile, and far-sighted. He presided over China’s greatest era of reform and opening up — 1991 to 2003 — and handled both macroeconomic and microeconomic policy with a skill that perhaps no other Chinese leader has ever matched.

In many ways, Zhu was the man who really made China a capitalist country. The privatization of state-owned enterprises had begun earlier, but Zhu continued and accelerated it. He was the main champion of China’s successful drive to join the WTO, which supercharged China’s export industries and led to a massive wave of foreign investment; he also engaged in a deregulation program to lure foreign businesses to manufacture things in China.

When the Asian financial crisis of 1997 led to a wave of bad loans, Zhu bailed out the banks, and growth quickly resumed. And Zhu also devised the most important tax reform in China’s modern history, transferring large amounts of local government revenue to the central government, while simultaneously letting local governments do things like borrow more money and sell more leases on land to fund their own activities.

There is no question in my mind that Zhu ought to be remembered as one of history’s great heroes. When he entered office, East Asia had 1.22 billion people living on less than $3 a day (in 2021 dollars) — more than half the world’s total. By the time Xi Jinping became China’s top leader, that number was one-sixth of what it had been:

This is the biggest and fastest decrease in absolute poverty that the human race has ever seen. Zhu didn’t do it alone, of course, but the economic model he devised was so effective at generating rapid growth that he deserves more credit than perhaps any individual other than Deng Xiaoping. The list of people who have done more to decrease aggregate human suffering is short indeed.

China is no longer the growth powerhouse it was in those earlier years. According to official statistics, over the last 15 years, China’s growth rate has fallen to 5% — less than half of what it was in Zhu’s time:

Some independent estimates put the growth rate much lower — maybe 2-3%. China is going through a difficult economic time right now, and its leaders have a tradition of “smoothing” official numbers by understating growth during booms and overstating it during busts.

Some amount of slowdown is to be expected. China is a much richer country than it was when Zhu left office — maybe five times richer. When Zhu ended his historic tenure in 2003, China was about as rich as Tanzania or Zambia is now. Today, it’s closer to Argentina or the Dominican Republic. A country can’t keep growing at supercharged rates forever; eventually, technological catch-up and capital deepening run their course, and a country is forced to grow based more on innovation and improved efficiency.

To be fair, Xi Jinping has taken some serious steps toward making that transition. His government has designed a “new national system” of research and development that tries to coordinate the entire supply chain of innovation, all the way from government-funded research labs to private companies putting products on the shelves. He has unleashed the most expensive and extensive system of industrial subsidies in world history, promoting essentially every advanced manufacturing industry in existence, as well as AI. These policies may be unorthodox, but their goal is the right one — they’re aimed at pushing China past middle-income status by replacing catch-up growth with frontier innovation.

Whether Xi’s policies have been effective at this goal is less clear. There’s actually a big debate over whether Chinese total factor productivity — a measure of overall economic efficiency — grew in the 2010s. The Penn World Tables, which had until recently shown a severe slowdown in TFP growth, recently changed their assumptions and now report very healthy annualized growth of over 2% throughout the 2010s. Other sources, such as the Conference Board and Brandt et al. (2022), still think there was a slowdown, and peg the rate at around 1%. But everyone agrees that China’s TFP growth has been slowing over time, and that Xi’s “industrial policy for everything” hasn’t managed to re-accelerate it so far.1

In fact, a lot of people are noting that China now appears to have a “two-speed” or “k-shaped” economy — split between a high-tech manufacturing sector that’s increasingly the envy of the world, and a broader economy that’s still mired in the aftermath of the country’s epic real estate bust.

Read more

A few days ago, I published a guest post by Tim Fist and Saif Khan of the Institute for Progress, discussing the question of whether we should deliberately try to slow down the rate of AI progress:

The authors promised a raft of specific policy recommendations, and they didn’t disappoint. Here is part 2, with all of those recommendations. Did you know that 23 is my lucky number?

In our last post, we evaluated the claims of a recent open letter by AI company employees calling for governments to “pace” frontier AI development. To summarize:

• Rapid progress towards fully automated AI R&D has empirical support, but it’s less clear how much it will accelerate AI capabilities or pose severe risks.

• Despite substantial uncertainty, we believe some preparatory policy action is warranted. This follows both from how serious the possible direct risks are and the risk that political backlash to AI-driven disruptions results in poorly-reasoned policy measures, such as broad bans on new data centers.

• If “pacing” becomes necessary, we think it should consist of two parts: first, specifying thresholds for when automated AI R&D is likely to pose severe risks; and second, if a threshold is exceeded, incentivizing AI companies to reallocate resources away from the most risky research, and towards activities that make further automation safer, or diffuse the benefits of existing AI faster.

Without preparation now, however, our preferred pacing strategy will be impossible to implement. In this post, we’ll describe how the US can concretely prepare for the further automation of AI R&D and the risks it entails.

Still, we aren’t certain whether the benefits of pacing outweigh the downsides, especially given the risk that government regulation is implemented counterproductively. So to make policy preparation as targeted and low-regret as possible, we think any intervention should meet the following five criteria:

1. Target only AI development activities that could lead to serious and irreversible harms.

2. Minimize any slowdown in the diffusion of existing AI capabilities, and ideally accelerate it.

3. Impose low costs, or deliver clear benefits, even if automated AI R&D and its attendant risks prove unlikely.

4. Avoid systematically disadvantaging more cautious labs and countries.

5. Avoid establishing a new regulatory apparatus that is likely to be misused (e.g., by concentrating power in a small set of companies).

A surprisingly wide range of policy moves meet these criteria. We’ve identified 23 of them, and they span 7 areas:

• Transparency: giving the government and public more visibility into automated AI R&D

• State capacity: improving the government’s ability to understand and respond to automated AI R&D

• Risk management: developing a risk management strategy for automated AI R&D that accelerates defensive and commercial AI use

• Verification: accelerating the development of AI verification technologies to enable agreements between mutually distrustful parties

• Resilience: accelerating the development of technologies that improve society’s ability to withstand and recover from AI-driven disruptions

• Competition with China: extending the US AI lead over China to buy more time to manage risks and increase US leverage in international negotiations

• Diplomacy: creating option value for international cooperation to manage the risks of automated AI R&D

In the rest of this post, we’ll explain why we think policy action is justified across each area and give specific recommendations for each. For more details on this and the material from yesterday’s post, you can read our full report here.

Transparency

AI now regularly makes impressive breakthroughs in math and is superhuman at many aspects of software development and cybersecurity, with capabilities doubling every 7 and 5 months, respectively. But outside of frontier AI companies, how exactly the drivers of AI progress — e.g., curating more and better data, improving training algorithms, applying more reinforcement learning, simply making the model bigger — are unlocking AI capabilities is poorly understood.

The same is true for many of the crucial questions surrounding AI R&D automation. Much of the best information remains inside company walls.1 Given that automated R&D could rapidly accelerate AI progress with little warning, this dynamic could leave the government and public unprepared.

It doesn’t have to be this way. If we want society to respond well, we’ll need much more information about what’s going on at the frontier of AI.

More transparency could help us understand the science behind automated AI R&D as well as what it looks like within specific companies (e.g., how much they’re automating, what policies they use to manage the risks, and any R&D-related incidents).

A recent NVIDIA-led letter supported open-weight models from an open science perspective — a valuable goal. But public disclosure practices around the science of and common practices within AI development (for both closed- and open-weight models) are minimal.2 Building on the findings of CSET and the Elasticity Institute, we suggest transparency measures for information in five categories:

1. The science of general AI progress.3

2. AI’s ability to automate specific AI R&D tasks.4

3. Company progress toward automating AI R&D.5

4. Company AI R&D automation risk management practices and incidents.6

5. Company “model behavior specifications,” documents that describe the values and principles an AI model is trained to follow (e.g., OpenAI’s Model Spec for its GPT models and Anthropic’s Constitution for its Claude models).7

The vast majority of this information is not subject to disclosure requirements, and so it is either disclosed voluntarily in a limited way or not at all.8 Although some particularly sensitive information may only be suitable for disclosure to the US government, we generally recommend transparency measures that involve public disclosure.

As AI companies automate more of their R&D, this information would enhance public and policymaker understanding, improve policy responses, and enable the broader scientific community to do better work on alignment and security.

1. Frontier AI companies and relevant industry bodies should publicly share information relevant to trends and risks in AI R&D automation

The categories of information outlined above would improve policymakers’ and the public’s understanding of the extent and nature of AI R&D automation at frontier AI companies, enabling outside experts to model, project, and publish on associated trends and impacts. In turn, this would improve policy responses and bolster the broader scientific community’s work on alignment and security.

Disclosing information about the science of AI and general AI capabilities would be a return to the historical norms of open scientific publication in the US AI industry. As recently as 2020, OpenAI published detailed information on the architecture, training recipe, and data of GPT-3, while in 2022, Google published detailed scaling laws showing how training compute, data, and model architecture correlate with AI model capabilities.

To reestablish this norm, employees at frontier AI companies should encourage their leadership to publicly share information in the categories outlined above, and advocate for reestablishing broader industry norms, including through industry bodies such as the Frontier Model Forum.

We believe commercial and geopolitical concerns over the sensitivity of this information are manageable. Industry-level technical metrics related to the science of AI and general AI capabilities are likely well understood by most or all frontier AI companies in the US and China. They are therefore unlikely to alter the balance of AI capabilities between US AI companies and between the US and China. Additionally, specific company-level AI R&D automation activities are of extraordinary public interest, such that improving the US government’s and public’s ability to mount a policy response outweighs concerns over commercial sensitivity. Of course, some specific information on cutting-edge breakthroughs — particularly where non-US companies lack comparable knowledge — will be crucial to US strategic interests and AI leadership, and may thus be less suited to public disclosure.

However, we do not recommend legal mandates requiring disclosure of this information. Legal mandates for public disclosure should focus on information related to risk management, as follows.

2. Congress should legislate transparency about automated AI R&D risk management, incident reporting, whistleblower protections, and model behavior specifications

Several current federal bills include transparency-related requirements for internal deployment:

• The FRONTIER Act, introduced in July 2026, includes the following three types of transparency and reporting obligations directly relevant to internally-deployed models (which are often used for automated AI R&D): (1) The developer’s publicly available frontier AI framework must describe how the developer reviews risk assessments in deciding to internally deploy the model; (2) The developer must periodically submit confidential reports to the Department of Commerce containing “a summary of an assessment of catastrophic risks arising from internal use of its frontier models”; and (3) The developer must report, within 72 hours, critical safety incidents involving its frontier models, including whether an incident was associated with internal use.

• The AI Incident Reporting Act, introduced in June 2026, would require reports for internal-only AI models, including with respect to an internal model that “when unprompted, has demonstrated the ability to materially accelerate or automate the research, development, evaluation, engineering, or improvement of advanced artificial intelligence systems, including in ways that could significantly compress timelines for the development or deployment of more capable systems.”

• Finally, the AI Whistleblower Protection Act, introduced in May 2025, would strengthen protections for whistleblowers regarding risky activities involving AI security vulnerabilities, legal violations, or dangers to public safety, public health, or national security.

The bills described above would be effective in mandating public disclosure of frontier AI frameworks and government disclosure of incident reports for internally deployed models, as well as protecting whistleblowers. New legislation is necessary to mandate disclosure of model behavior specifications. Some bills, such as the FRONTIER Act, include much broader regulatory scope than public disclosure — but discussing those provisions is out of this report’s scope.

State capacity

Right now, the US government lacks the well-resourced institutions required to fully keep up with AI R&D automation, to inform politicians and frontier AI companies on how to respond, or to coordinate a pacing strategy.

However, absent a state authority trusted by society at large, managing the pace of rapid AI capability improvement will be impossible. Only democratically accountable leaders possess the legitimacy and power to coordinate the kind of mass reallocation toward defense and diffusion that may be needed.

The government should therefore build institutions capable of responding effectively to rapid AI progress.9

3. Congress should resource the Center for AI Standards and Innovation (CAISI) with a budget of at least $84 million per year and empower it to directly advise senior government officials and frontier AI companies

Congress should provide CAISI with the manpower, resources, and autonomy required to generate real situational awareness of AI progress for policymakers and shape frontier AI companies’ technology development and incentives. IFP previously recommended increasing CAISI staff to 184 personnel and giving it a budget of $84 million, and the America First Policy Institute recommended allocating $50–100 million to CAISI per year. This should be considered a minimum bar: developing a strategy for, recognizing the conditions justifying, and managing a coordinated effort to respond to the automation of AI R&D would likely require substantially greater capacity and resources.

More important than size, however, are CAISI’s autonomy and business model: CAISI should have a direct line of communication to a senior White House or cabinet-level official and be able to provide its expertise across the federal government. CAISI should also be empowered to:

• Forward deploy its staff into the frontier AI companies, thus increasing the government’s situational awareness of frontier AI companies’ major research projects, internal deployments, and risk management cultures,

• Organize a collection of 3rd-party AI evaluators and data producers,

• Establish direct contracts with the frontier AI companies to support priority research and security efforts (e.g., compute subsidies for alignment teams, implementation of control and verification technologies, and adoption of more robust model monitoring teams), and

• Inform government export control policies on frontier AI products and services.

4. The White House should set clear roles and responsibilities of US government agencies to increase specialization across AI policy

CAISI should be the government’s central hub for AI expertise and evaluation, working with the Department of War, Intelligence Community, and the Department of Energy. Accordingly, the Department of Commerce, where CAISI sits, should serve as the US government’s lead on frontier AI technology, with its primacy clarified in future executive orders.

CAISI would inform the White House’s development of overall frontier AI policy as well as other government agencies’ management of potential domain-specific AI risks. This would include generating evaluations, analyses, and recommendations to enable the CDC’s management of biological risks, the National Security Agency’s (NSA) and the Cybersecurity and Infrastructure Security Agency’s (CISA) management of cybersecurity risks, and Treasury’s mitigation of financial risks.

CAISI also benefits from being housed in the Department of Commerce, which has strong AI-relevant technical and policy capacity in National Institute of Standards and Technology (NIST) labs, the CHIPS Program Office (which has the government’s leading semiconductor expertise), the Bureau of Industry and Security (which handles AI-related export controls), and the International Trade Administration (which houses AI industry analysts).

The Department of War, in particular the NSA, is well-positioned to inform AI company security efforts, drive IT network security and product hardening, and build cross-company hunt infrastructure to track and disrupt AI-driven cyber threats. The NSA should also provide operational data on AI productivity in cryptologic operations to inform CAISI’s understanding of model performance in technical exploitation. Likewise, the Department of the Treasury is well positioned to evaluate and mitigate AI R&D automation risks in the financial services industry.

5. Intelligence agencies should improve their collection and analysis on foreign AI development and counter threats targeting US AI companies

Just as the US government needs visibility into US AI companies’ automated R&D, it has a vital interest in tracking AI R&D activities conducted by companies in the People’s Republic of China (PRC). The NSA, Central Intelligence Agency (CIA), and Federal Bureau of Investigation (FBI) should therefore dedicate collection resources to PRC regulators, lab executives, research leads, lab IT systems, IT service companies, and semiconductor companies. In particular, these organizations should establish an integrated, colocated mission organization focused on PRC technology targets, with common leadership, entitlements, and IT systems, thereby enabling integrated HUMINT-SIGINT-counterintelligence approaches to the PRC AI industry and associated strategic exploitation targets.

PRC technology analysts across these agencies, as well as the National Geospatial-Intelligence Agency and Defense Intelligence Enterprise, should be awarded entitlements to all raw SIGINT and HUMINT collection derived from PRC technology targets; this democratized access would enable the Intelligence Community’s development of indications and warning analysis as to when PRC AI development poses imminent risks, where semiconductor export controls are being circumvented, and where various risk management commitments are being flouted.

The US Intelligence Community should organize collection and analytic activities to provide this warning to senior government leaders, thereby enabling many of the export control and verification activities proposed below. These agencies should also contract with private-sector AI specialists to augment their analytic workforces, leveraging commercial researchers and technical expertise to improve their understanding of the collection.

As AI research and development is automated, frontier AI companies’ internal model weights and research advances will become increasingly valuable to intelligence services, rival AI companies, and agents that aim to improve their own capabilities or accomplish tasks beyond their current capabilities. The FBI, NSA, and CIA should therefore also partner to better track foreign espionage threats against US AI companies and to improve frontier AI companies’ counterintelligence programs. In doing so, these agencies should partner with the AI companies and their IT service providers to implement more effective application monitoring, information siloing, physical surveillance, and offensive counterintelligence operations to track and mislead industrial espionage campaigns, including those undertaken by AI agents.

The White House should direct these initiatives via a classified national security memorandum. This memorandum should outline key intelligence, operational, and security objectives vis-a-vis the PRC AI industry and protection of US AI technology, direct the establishment of a joint interagency mission organization focused on PRC technology targets, direct the establishment of relevant collection capabilities and IT systems to support related collection and analytic missions, direct the production of indications and warning-focused analysis, and expansion of relevant entitlements to specific populations of analysts.

Risk management

The AI policy community lacks a consensus strategy for managing the risks posed by rapidly improving AI capabilities.10 Given how fast automated AI R&D might improve AI capabilities if it happens, governments and AI companies should develop guidelines for risk management in advance.11

These guidelines would extend today’s efforts to evaluate and establish thresholds for AI safeguards, capabilities, and model releases, helping government and industry have a plan in place for the “day after” explosive growth in capabilities.

6. CAISI should develop guidelines for managing the risks of rapid AI capability improvement

In doing so, it should publicly take input from a range of stakeholders, including:

• Leading US AI companies, including OpenAI, Anthropic, Google DeepMind, Meta, SpaceXAI, SSI, and Thinking Machines;

• Foreign AI companies;

• Key companies across the AI supply chain, including semiconductor companies like NVIDIA; and

• Experts from third-party AI evaluation organizations such as METR and Apollo.

Related guidelines should include but not be limited to:

• Thresholds for unacceptable risk from an AI system when undertaking automated AI R&D;

• “If-then” commitments where reaching a certain capability or risk level triggers a mitigation or release decision;

• Specifications of required levels of human oversight during automated R&D as well as inference of continuously learning models;

• Preferred alignment and security techniques;

• The appropriate resource allocation between capability research and acceleration of techniques for monitoring, control, and alignment;

• Preferred capability research pathways that pose less risk; and

• Agent monitoring and control specifications, building on ongoing CAISI information gathering.12

Of particular importance are the risk thresholds: within our proposed implementation of “pacing,” their breach would trigger consideration of efforts to reallocate resources away from automated AI R&D, and towards societal resilience (see “Invest in AI resilience” section below), safety research, and AI diffusion.13 The guidelines should also describe how to carry out this reallocation, including plans to reallocate researchers and compute resources and to preserve the commercial viability of AI companies.

Verification

At some point, AI R&D automation might motivate the US to cooperate with other countries on managing its risks. Without an international agreement, regulations on US AI companies could put them at a disadvantage compared to their competitors, particularly Chinese AI companies.

But the economic or national security advantages of continuing to build more powerful AI may produce a strong incentive for each country to defect on any agreement. Preventing this kind of defection without resorting to aggressive measures like large-scale inspections or shutting down data centers would require something new: privacy-preserving verification allowing mutually distrustful parties — such as the US and — China to verify how their AI compute (the chips and data centers that are a necessary condition for developing and deploying frontier AI systems) is being used.

AI chips already include some physical mechanisms that allow them to securely report properties about the code they are running, but these mechanisms could be vastly improved.14 Other verification tech might involve doing “spot checks” on computation,15 using mutually trusted AI systems to inspect sensitive information in a counterparty’s systems,16 or leveraging power meters attached to AI data centers to infer what type of workload is running.17

Better verification technology would allow governments (or AI companies18) to agree to targeted, verifiable, privacy-preserving limits on the most dangerous activities while permitting beneficial AI applications and accelerating their diffusion.

However, no AI verification technology is yet mature. The field is still in its infancy. By some optimistic measures, barely 50 people are working on it worldwide. Making verification feasible therefore requires investment. If the nuclear arms control precedent is any indication, the ability to verify that agreements are being upheld is often necessary for parties to enter into them in the first place. Absent further progress in AI verification technology, “pacing” frontier AI development could become a euphemism for pausing it.

7. CAISI should co-lead an AI Verification Consortium (AIVEC) with industry to prototype and deploy verification technologies

There is currently no organization with a mandate to ensure that AI verification technologies are deployable if they are needed soon. AIVEC would fix that, bringing together the AI industry, philanthropy, and government in a public-private partnership to coordinate funding and drive faster progress on these technologies.

AIVEC would lead and coordinate several activities, including:

• Defining clear goals about what properties of AI workloads AIVEC and its constituents want to be able to verify and when those technologies need to be ready.

• Roadmapping the technologies and ensuring they are on track to reach those goals on the desired timelines.

• Disbursing direct grants to technical teams.

• Coordinating the creation of AI hardware testbeds to accelerate R&D cycles.

• Launching prize competitions.

• Funding and coordinating a buildout of data centers with the capacity for verifiable AI inference.

• Making further policy recommendations to the US government and serving as a forum for ongoing communication among frontier AI companies on verification efforts.

• Informing diplomatic discussions between the US and foreign governments about possible international agreements and verification measures to support them.

Some of these activities are described in greater detail below: AI hardware testbeds, prize competitions, and the construction of a verifiable AI inference data center.

AIVEC could be structured in a number of ways, with varying government participation and degrees of focus on the activities above. But no matter its structure, AIVEC should be empowered to conduct, fund, and direct the requisite R&D to make compute verification possible.

AIVEC could be set up as a philanthropic project without any government involvement, following the “general manager” model. However, US government involvement will likely be crucial to AIVEC’s success, for three reasons. First, US government leadership may be necessary to secure deep collaboration between the leading AI companies on this topic. Second, the US government can inform which verification targets are necessary for international agreements. Third, US government participation from the start would build confidence in the technology’s viability.19

8. AIVEC should coordinate the creation of AI hardware testbeds and make them available to government, industry, and nonprofit partners

Access to advanced AI hardware is a major bottleneck to rapidly prototyping verification technology. A single state-of-the-art AI server can cost hundreds of thousands of dollars, and rack-scale systems can cost over $3 million. An ideal prototyping cluster might cost tens of millions.

To accelerate testing cycles, AIVEC should play a coordinating role in setting up hardware testbeds, sourcing in-kind hardware transfers from AI companies and hyperscalers, ensuring alignment among AIVEC partners on testbed specifications, and providing R&D partners with direct access to the testbeds for unimpeded testing and experimentation.

These testbeds will support earlier development phases, enabling R&D teams in industry, government, and the nonprofit sector to prototype, tamper with, and even break verification components. These same testbeds could also eventually be used as small clusters in the adversarial grand challenge detailed in Recommendation 9 below.

9. AIVEC should launch philanthropically funded prize competitions for AI verification headed by CAISI

Prize competitions are effective at incentivizing innovation in technical domains where a clear goal can be specified, but there is no single team or best research approach to fund directly. Unlike direct funding of R&D in high-risk research directions, they enable funders to set ambitious goals without risking significant financial loss. Prizes also increase the prestige of working on specific problems, attract talent, and help build expertise in new and important technical fields.

The 2010 America COMPETES reauthorization Act authorizes the heads of federal agencies to create prize competitions, in which a reward (usually cash) is offered to participants to achieve a specific goal. These competitions can be administered by — and the rewards can be supplied by — private, nonprofit entities.20

Verification prizes may be run, administered, and adjudicated by AIVEC and CAISI, a nonprofit partner, or a CAISI-appointed panel of technical experts, and funded by philanthropic donations and/or funds from Federal agencies.

We recommend two different versions of prize competitions, which differ mainly in their complexity and funding requirements: lightweight, well-specified prizes that can be launched quickly and gather responses from a broad range of actors; and a grand challenge that involves fewer participants but provides them with R&D budgets, full access to hardware testbeds, and built-in adversarial dynamics.

Lightweight prize competitions

Not all AI verification problems have clear, fully specifiable goals. But some of them do. Potential initial verification targets include workload classification from telemetry21 and accelerated cryptographic proofs of AI workloads.22

AIVEC should coordinate the launch of a series of well-specified, philanthropically funded prize competitions for these domains of AI verification.

These prizes would not provide continuous hardware access or upfront funding, so they would require no participation downselection or in-kind transfers from industry. Funders can simply specify a problem and pay whoever credibly solves it.

An adversarial grand challenge

In addition to the lower-overhead prizes described above, AIVEC could coordinate a more ambitious adversarial grand challenge that provides real hardware to top teams, simulates agreements between them on how to use their compute, and incentivizes them to innovate in adversarially robust verification technology with a series of prizes.23

Any agreement that requires verification technology will, by definition, have strong incentives for both sides to defect, including by tampering with the verification technology. So just like software developers, compute verification developers will have to continually refine their safeguards against new exploits. Verification would therefore benefit from pilot deployments that allow for realistic stress tests.

The verification grand challenge could be structured as follows:24

• AIVEC gathers input from frontier AI companies and national governments on the kinds of agreements that may be useful in the future, if automated R&D leads to a pacing strategy being pursued.

• AIVEC opens a call for teams to participate in the competition. These could be national AI institutes (US CAISI, UK AISI, etc.), AI companies, or others.

• The teams submit their proposed verification systems, which are then evaluated, down-selected, and conceptually red-teamed by a panel of judges and expert evaluators.25 The top projects enter the competition.

• Build a series of small clusters with in-kind transfers from industry, potentially funded by philanthropy.

• Give each team full access to one of these small (e.g., <1 MW) clusters, where they are expected to run real AI workloads (training, reinforcement learning, or inference).

• Iterate on a series of mock agreements between the participants, based on the earlier input from frontier AI companies and national governments.26

• Set a series of prizes to incentivize teams to defect on the mock agreements without being caught by other participants or to provide proof that other participants defected.

Participants should have full access to their clusters, and be allowed to provide access to other groups as they see fit. Participants should also be encouraged to develop hardware or software verification technologies using a budget allocated to them and to deploy these technologies without restriction on their own clusters — or, if agreed by both parties, on others’ clusters.

This would essentially be a more structured version of the earlier recommendation of lightweight prize competitions, but with built-in simulated adversarial dynamics, upfront R&D budgets for the teams, and full hardware access.

The adversarial dynamic would be key to triggering a race to the top in AI verification, as well as simulating some of the real-life adversarial conditions that these verification measures may need to overcome.27

10. AIVEC should coordinate the construction of a fully verifiable data center

Hardware testbeds and prize competitions can produce verification components and reveal their weaknesses, but they will not necessarily demonstrate how these components integrate into a data center capable of running verifiable frontier workloads at acceptable cost.

AIVEC should coordinate industry, hyperscalers, and nonprofit partners to build a pilot verifiable data center. The goal would be to align AIVEC members on a complete reference architecture and demonstrate viability in the real world at a reasonable cost. Given the technology’s early stage, a complete prototype will likely require iteration across a series of pilot projects. For a vision of what a Focused Research Organization focused on verification might look like, see Faster AI Diffusion Through Hardware-Based Verification.

A key part of this project’s success will be incorporating red-teaming by the Intelligence Community and other top-tier operators to identify ways that these verification methods could be evaded.

11. DARPA and the NSF should set up AI verification R&D programs

Most promising verification technologies have not yet been turned into working prototypes. This is an urgent problem with national security implications, and is too far from commercialization for industry to solve on its own.28

DARPA should run a verification program to build end-to-end prototypes that enable AI companies to make verifiable claims about how their compute is being used, in a way that is privacy-preserving, IP-protecting, and secure against adversarial exploitation. This would follow the precedent of ARPA’s Project Vela, a project to verify Soviet Union compliance with the 1963 Partial Test Ban Treaty for nuclear weapons testing.

The NSF should fund the underlying science that will enable breakthroughs in verification. Two possible vehicles to achieve this include cooperative agreements with industry and NSF X-Labs. The X-Labs initiative is meant to support organizations built to “address technical challenges and bottlenecks that university and industry labs cannot easily solve through traditional methods.”29 Much of the required work in AI verification relies on infrastructure that is too expensive for individual university labs, and has no clear commercialization potential.

These programs would be complementary to the recommendations above. The urgency of advancing compute verification technology means we should be willing to make multiple parallel bets to achieve the goal of usable prototypes. While AIVEC may be best suited to leverage industry funding and frontier AI expertise, it may not be as well positioned to tap into DARPA’s deep national security expertise or the NSF’s interdisciplinary networks of scientific collaboration.

If and when the technologies do have commercialization potential, Small Business Innovation Research (SBIR) grants may be used by either agency to support startups building in the space before their products are commercially viable.30

12. Intelligence agencies should develop and operationalize unilateral means of AI compute monitoring

Technical means of AI verification can only work if the parties interested in engaging in verifiable agreements have high confidence that their counterparty does not have large data centers that have not been reported (and thus cannot be verified).

Moreover, rapid AI progress may mean that US AI companies or the US government will seek to strike deals to manage the risks of sudden AI R&D automation before the targeted verification technologies described above are developed. In that case, intelligence agencies will probably need to play a large role in being able to certify that their counterparty is complying with the agreement.31

US intelligence agencies should thus prioritize gaining visibility into where foreign competitors have large quantities of compute and how that compute is being used, similar to how the Intelligence Community tracks fissile material to support the International Atomic Energy Agency’s investigations. This could be achieved through SIGINT and HUMINT, satellite imagery and infrared sensing, and AI chip supply-chain intelligence and accounting.

AI resilience

A world more resilient to AI disruptions is possible. If, at the beginning of this year, codebases had been highly secure and cybersecurity experts across industry and government had been able to rapidly patch newly discovered vulnerabilities, the public release of Anthropic’s Claude Fable might not have prompted deployment restrictions by the US government.

We can make that hypothetical a reality by investing in AI resilience: the ability to withstand and recover from major AI-driven disruptions. If automated AI R&D presents serious risks, improving AI resilience could increase the point at which pacing the frontier would be required.

As recent incidents show, AI resilience is a worthwhile goal even at today’s levels of AI capability. Restricting access to closed US frontier models would only be a temporary solution. Based on historical trends, Chinese and open-weight AI models lag the frontier by around 8 months. That means a model with the capabilities of Claude Mythos will likely be widely deployable by Chinese organizations by the end of 2026, giving the US around 4 months to build resilience against the new risks this introduces.

13. NSA, CAISI, CISA, and ONCD should further invest in cybersecurity resilience

AI cyber capabilities are doubling roughly every 5 months, making cybersecurity an ongoing battle between offense and defense. This requires sustained investment in cyber resilience. The NSA, CAISI, CISA, and ONCD should plan for the moment when Mythos-class models are developed and deployed by Chinese organizations, as well as open-sourced.32

Measures to increase cybersecurity must span the relevant attack surface. Promising proposals include:

• “Operation Patchlight”: A more ambitious version of Anthropic’s Project Glasswing and OpenAI’s Daybreak, Operation Patchlight would use frontier AI to find and patch open-source code vulnerabilities before attackers exploit them by giving AI-powered security tools and free compute access to critical infrastructure defenders. A 3-year national effort coordinating AI companies, code maintainers, and industry could help shift the advantage to defense. This approach could also be extended to harden closed-source critical infrastructure code.

• “The Great Refactor”: AI could automatically translate hundreds of millions of lines of insecure critical open-source code into memory-safe Rust, which would eliminate whole classes of vulnerabilities. This project could be set up as a Focused Research Organization to systematically secure high-impact libraries before 2030 and save billions in cybersecurity costs.

• “Preventing AI Sleeper Agents”: The Departments of War and Commerce could create an AI Security Office to red-team and blue-team the full American AI tech stack against sleeper-agent risks.33 This effort could launch as a pilot to evaluate leading AI companies and develop prevention tools; if successful, it could grow into a large-scale national security program.

• Proactive forensic-depth AI cybersecurity: AI could automate digital forensics and threat hunting to turn forensic-depth cybersecurity from a post-breach service into an always-on detection method.34 More broadly, advanced AI systems can be used to identify and defeat tradecraft for, develop capabilities within, and operate cybersecurity monitoring systems. Philanthropists, cybersecurity companies, and the NSA can fund AI-driven improvements to these cybersecurity systems, including AI-developed YARA rules for firewalls, more realistic honeypotting systems, and network device monitoring systems.

• Cross-company internet hunt architecture: Internet, cloud, financial, and cybersecurity service providers can share telemetry and build federated analytics systems to track AI cyber campaigns across thousands of targets. The relevant data-sharing, collaborative analytic, and operational mitigation systems for this national hunt architecture do not yet exist, but IT service providers could work with NSA or CISA to build them.

• Hardening of strategic exploitation targets: Cloud, internet, and IT companies with large credential repositories, as well as managed service providers, are frequently compromised to enable downstream access to thousands of targets; advanced AI systems will enable these exploitation operations at enormous scale. AI and cybersecurity companies could partner with prioritized strategic exploitation targets to harden their core networking infrastructure (e.g., by implementing more secure hypervisors), thereby frustrating efforts to compromise downstream targets.

14. Congress, OSTP and the CDC should invest in biosecurity resilience

We should prepare for a future in which AI could lower the barriers to creating biological threats. Whether AI will meaningfully aid novices in creating biological weapons is hotly debated, and the measured effect was not significant as of mid-2025. But frontier models have continued to improve on every biology benchmark, including those measuring the tacit knowledge required for benchwork troubleshooting. Even if we are not certain, it is worth building resilience now for a world in which AI enhances bioweapon creation, since defenses will take time to build.

Biosecurity is uniquely difficult because bioattacks and biodefenses scale asymmetrically. There is a massive offensive advantage: compared with defenses such as pathogen-agnostic detection and medical countermeasures, attacks take less time to deploy, cost less, and benefit more from advances in AI. Broadly, attacks are relatively more bottlenecked by information and tacit knowledge, while defenses are relatively more bottlenecked by physical properties, such as manufacturing capacity and physical infrastructure. By removing friction in accessing information, including tacit knowledge, AI is poised to accelerate attackers’ progress more than defenders’.

The US should therefore build a strong defense to prevent, detect, and defend against biological threats. We recommend:

• Prevent dangerous biological threats from being created and released: Congress should secure the DNA supply chain, such as by passing the Biosecurity Modernization and Innovation amendment as part of the American Biotechnology Competitiveness Act, requiring DNA synthesis providers to screen customers and orders while minimizing burdens on legitimate research. Congress should also equip CAISI to evaluate new models for biological risks and to develop safeguards standards to reduce dangerous uplift without degrading legitimate use. Finally, OSTP should coordinate NIH and NIST pilots of tiered access controls for narrow subsets of pathogen data most likely to enable misuse and provide the gene synthesis screening framework requested in Executive Order 14292.

• Detect threats before they spread widely: The CDC should invest $80M annually in a federal pathogen early-warning system, adding metagenomic sequencing to existing traveler, wastewater, and clinical surveillance networks, so the US can rapidly detect novel biological threats. Using the resulting information, the Defense Innovation Unit should continue to fund the development of better AI tools capable of scanning these billions of data points for emerging threats, and explore new options for better biosurveillance.

• Defend the population and shorten the path out of the pandemic: Congress should sustainably fund the Strategic National Stockpile to grow and maintain day-zero reserves of respirators and medical countermeasures for future pandemics. Congress should also sustainably fund warm-base manufacturing capacity to ensure that we are prepared to rapidly manufacture countermeasures for novel threats. The Biomedical Advanced Research and Development Authority (BARDA) should provide advance purchase commitments for broad-spectrum antivirals. The Department of Health and Human Services should fund large-scale trials of air-cleaning technologies, such as Far-UVC and glycol vapors, and recommend evidence-based updates to ASHRAE Standard 241. Finally, Congress should provide incentives for buildings to procure these air-cleaning technologies.

Many more projects should be undertaken in biosecurity,35 as well as in other critical areas.36 This list is far from exhaustive. IFP is currently developing more concrete project proposals to increase AI resilience as part of The Launch Sequence.

Competition with China

AI chips are America’s asymmetric advantage. Unlike for algorithms and training data — the other key inputs to AI — policy can durably shape access to the large amounts of computing power needed to develop and deploy frontier systems.37 Due to compute export controls, US companies currently access approximately 10x more AI compute than Chinese companies,38 enabling US AI model capabilities to be approximately 8 months ahead of China’s. However, this lead could be much larger.

Increasing the lead would have multiple benefits. It would allow the US to simultaneously maintain long-term AI leadership while addressing the risks of automating AI R&D. It would buy more time to build AI resilience. And it would give the US greater leverage in any international negotiations about AI risks.

The US can increase its lead, potentially to over 2 years, by increasing its AI compute advantage, addressing adversarial distillation, preventing theft of US model weights, and making it easier to build AI data centers at home. If the US adopted the tighter AI chip and semiconductor manufacturing equipment (SME) export controls described below, its compute advantage could be well over 100x, adding several months to the lead. If the US countered adversarial distillation — the process by which Chinese AI companies train their models on the outputs of US models — it could add approximately another 7 months to the lead, per an industry estimate. And protecting US model weights from theft would ensure that China cannot quickly erase the lead.

15. Congress and the Bureau of Industry and Security (BIS) should strengthen controls on US and allied semiconductor manufacturing equipment

Existing SME controls have already been highly effective, establishing a 50–100x US advantage over China in the production of aggregate AI computing power. But gaps in SME controls could limit long-term effectiveness and prevent the US lead from being even larger, as even China’s limited existing production has relied on US and allied SME allowed into China.

The bipartisan MATCH Act, currently slated to be included in the must-pass yearly National Defense Authorization Act (NDAA), would require the Administration to negotiate with allied governments or otherwise implement unilateral controls to close key remaining gaps. Most importantly, the MATCH Act would mandate China-wide controls on all deep-ultraviolet (DUV) immersion equipment. Such equipment represents the most significant chokepoint in making AI chips and is still not restricted China-wide. BIS can also tighten these controls with its existing authorities.

16. Congress and BIS should close gaps in AI chip controls

The US compute production advantage of 50–100x has translated to only a 10x compute access advantage because the US has allowed China to access chips in four ways.

1. The US has approved sales of powerful US AI chips to China, including the NVIDIA H200.

2. The US has insufficiently restrained Chinese smuggling of restricted US AI chips.

3. The US is choosing not to enforce regulations that limit leading US and allied chip foundries, such as TSMC and Samsung, from producing AI chips for potential Chinese cutouts in third countries.

4. The US allows leading Chinese AI companies to legally access the most powerful US AI chips via cloud services. The White House has asserted the best current Chinese AI model — Moonshot AI’s Kimi K3 — was trained on cutting-edge NVIDIA GB300s accessed via cloud services in Thailand.

The Trump Administration, via BIS, should close gaps in these controls. Additionally, Congress should legislate to preserve the US compute advantage. Specifically, Congress should pass the AI OVERWATCH Act, which would codify restrictions on exports to China of the most advanced US AI chips, such as the NVIDIA GB300. Congress should also pass the Chip Security Act, which would mandate privacy-preserving location verification of AI chips to prevent smuggling. Stronger location verification can also enable more exports to ally and partner countries, accelerating faster AI diffusion while reducing smuggling risk. Due to supply constraints, every chip sold to China reduces the number of chips US organizations can deploy to maintain the US lead and deploy AI in defensive applications.

17. The Federal Trade Commission (FTC), Department of Justice (DOJ), BIS, CAISI, and Congress should help industry counter adversarial distillation of US AI model capabilities

Adversarial distillation enables Chinese AI companies to extract capabilities from frontier US AI models while using very little of their own compute, instead free-riding on the compute-intensive “reinforcement learning” steps required for training US frontier models. We recommend the following actions to help counter adversarial distillation:

First, BIS should prioritize export controls on AI chips and SME described in Recommendations 15 and 16. Beyond reducing China’s ability to train large, powerful AI models, compute controls also limit China’s ability to benefit from adversarially distilling US models. The smaller a Chinese AI model is relative to a large US teacher model, the less able it is to benefit from distillation of the US model.

Second, the FTC and DOJ should provide guidance providing antitrust safe harbors for industry threat intelligence sharing, building on 2014 guidance on safe harbors for cybersecurity threat information sharing. Additionally, Congress should codify such a safe harbor, such as by passing the Collaboration on Adversarial Threats and Security Risks Act. However, before passage, we recommend adding language to the bill confirming that any legal cause of action “shall be promptly dismissed” if the collaboration falls within the safe harbor. Currently, the bill provides a safe harbor only as an affirmative defense, which likely would be raised only after substantial litigation discovery. This could chill use of the safe harbor.

Third, CAISI should establish a public-private partnership to develop guidelines to raise the floor of industry-wide safeguards to combat adversarial distillation.

18. BIS should maintain visibility into sales of US chips

With tightened SME controls that ensure virtually all AI chips are produced by US chipmakers, BIS should ensure that the chipmakers maintain visibility into customers to confirm that chips do not fall into the hands of irresponsible or untrusted parties over which it is more difficult to practice oversight. BIS can then gather this information to maintain an AI chip registry, which also supports verification and intelligence collection activities described above.

19. DOW, the IC, CAISI, and relevant Federally Funded Research and Development Centers (FFRDCs) should establish consensus security guidelines for protecting model weights from theft and prototype them in a government facility

Consistent with the July 2025 AI Action Plan, CAISI and NSA should develop guidelines for industry to protect AI model weights from exfiltration by adversaries, including nation-state actors, particularly to reach Security Level 4 (SL-4) or SL-5. (Helpfully, RAND has already developed guidelines for SL-4 and SL-5 data centers.) Stealing AI model weights is currently possible for well-resourced state-backed actors and leading cyber-capable institutions. Compared to the model outputs obtainable via distillation, the model weights themselves are the crown jewels, and AI companies are not currently able to prevent their exfiltration. If American AI companies train an AI model capable of fully automated AI R&D, foreign adversaries will likely prioritize obtaining the model weights, as they could use them to automate training their own models.

The DOW, NSA, other IC agencies, CAISI, and relevant FFRDCs such as RAND should also prototype an SL-4 or SL-5 cluster for sensitive government applications. Having such a cluster available would allow highly sensitive model weights developed by industry to be transferred to a much more secure facility, if the situation demands it.39

A pilot phase could involve designing and building a secure inference data center through in-kind hardware transfers from leading American AI companies, chipmakers, or hyperscalers. RAND estimates the cost of such a data center to be $37–50 million for a proof-of-concept facility and $277–345 million for an enterprise-scale version.40

Eventually, this secure data center model should be scaled up to enable larger-scale secure deployments, e.g., to support preemptive cyberdefense programs that patch critical infrastructure such as Project Glasswing and Daybreak. For specific recommendations on how the US government can achieve this, see A Sprint Toward Security Level 5.

This cluster would also provide better protections against incidents such as an unreleased OpenAI model’s recent autonomous cyberattack on AI infrastructure company HuggingFace.41 Within a secure cluster, future models could be evaluated with less risk of them escaping containment, which is particularly useful if they are developed with minimal human supervision via automated AI R&D.42

20. Congress should ensure the US has sufficient electrical capacity to sustain AI leadership

To maintain and increase America’s AI compute lead over China, we need electricity to power data centers.

The Electric Power Research Institute now projects that data centers could consume 9 to 17 percent of US electricity by 2030, up from 4 to 5 percent in 2024 and that by 2030 a single model training run could require between 4 and 14 GW. After two decades of static electricity demand, utilities and regulators must suddenly accommodate enormous new loads.

Our regulatory and permitting systems are not equipped to handle this new growth. Connecting AI data centers to the grid will require both new power plants and new transmission lines. Generation and transmission projects take many years to permit and reach operation; for instance, a Department of Energy review found that new transmission projects take about ten years on average to plan, permit, and build. But these are policy choices, not physical limits.

By contrast, China has a much larger power system and is expanding it much faster. China generated more than twice as much electricity as the US did last year, up from parity in 2010.

In 2025, China added more than 430 gigawatts of wind and solar capacity, and is on track to add similar amounts in coming years, essentially adding the nameplate capacity of the entire US grid every few years.

Congress should treat electrical capacity as part of the country’s AI industrial base. It should pursue four related reforms.

First, Congress should fix the broken interconnection process for both generators and large electricity users. The Grid Connection and Congestion Management Act provides a model for generator interconnection. It would allow generators to accept curtailment during congestion rather than wait for every transmission upgrade required to guarantee firm delivery. Congress should also create a parallel national framework for large loads. In June 2026, the Federal Energy Regulatory Commission (FERC) ordered all six regional grid operators under its jurisdiction to justify or reform their rules governing service for flexible large loads. Congress should confirm FERC’s authority over large-load interconnection, codify minimum rules for study timelines, cost responsibility, co-location, and expedited service for curtailable loads, and extend equivalent requirements to regions without an independent grid operator.

Second, Congress should make commonsense reforms to the permitting system that governs infrastructure broadly. Laws like the National Environmental Policy Act, National Historic Preservation Act, and Clean Water Act could be streamlined without compromising their important substantive goals. Further, Congress should limit the executive branch’s power to disrupt already permitted projects through permitting certainty legislation like the bipartisan FREEDOM Act.

Third, Congress should reform transmission siting and funding. A high-voltage transmission line that serves several states should not be blocked because one state counts only the costs within its borders and ignores regional benefits. The Energy Permitting Reform Act of 2024 and the SPEED and Reliability Act provide models. Both would remove the Department of Energy corridor-designation process that now limits FERC’s backstop authority, allow developers to seek a FERC permit after states have had a year to act, and require the beneficiaries of federally permitted transmission to bear its costs. Several other pending bills supply provisions worth incorporating, including the FASTER Act, which would make FERC the lead permitting agency for qualifying projects, as FERC is for natural gas projects, and the BIG WIRES Act, which requires linking interregional grids.

Congress should also authorize a Grid Infrastructure Fund whereby participating data center developers would pay the cost of serving their loads and contribute to an insurance pool that protects ratepayers if a promised facility is not built after the infrastructure to connect it has been built. In exchange, the Fund would aggregate demand and equipment purchases, use federal credit to lower financing costs, and coordinate investment in transmission, generation, and domestic grid equipment manufacturing. A fund backed only by hyperscalers could mobilize an estimated $15 to 25 billion over five years, and federal lending authority and private co-investment could make it substantially larger.

Fourth, Congress should invest where private capital cannot absorb early-stage risk. The Department of Energy’s $2.5 billion Transmission Facilitation Program has supported nearly 1,000 miles and 7.1 gigawatts of new transmission capacity by acting as a guaranteed customer for projects that may otherwise not be economical to finance. Congress should expand this model and pair it with federal credit reconductoring. The bipartisan REWIRE Act of 2026 would speed permitting for advanced transmission upgrades and incentivize their deployment. Congress should also use Defense Production Act and Office of Energy Dominance Financing authorities to expand domestic production of transformers, switchgear, high-voltage cable, and turbines, building on prior proposals to fund grid equipment manufacturing under DPA Title III.

21. Congress should ensure data centers can be constructed in America

Continued AI leadership will require new data centers. Opposition to data centers has the potential to constrain US AI leadership and relatively advantage our strategic rivals, putting the US in a worse position to negotiate with them on managing automated AI R&D risks.

A supermajority of Americans, in a recent Gallup poll, report opposing an AI data center in their area. At least dozens of projects worth billions of dollars were blocked or delayed during the first quarter of 2026. Fourteen states and numerous localities are considering statewide moratorium proposals. New York has paused permits for new hyperscale data centers. Even Texas recently halted new grid approvals pending an audit. Multiple senators have proposed federal moratoria.

Instead of moratoria on new data centers, Congress should pass new legislation that balances the need to build strategically important infrastructure with the concerns of the host communities, ensuring a transparent and positive-sum result so that the country and localities both benefit from investment in AI infrastructure.

Specifically, Congress should pass legislation that provides for the creation of special industrial zones in communities that “opt in” and, in exchange, receive legally guaranteed community benefits.43 Such a bill should establish a transparent and public process through which a zone could be designated, following approval by the state and affected local governments. Federal agencies could then assess proposed zones for energy availability, security, and environmental suitability before selecting them. Selected zones would receive statutorily expedited, programmatic permitting.

Before receiving expedited treatment, developers would negotiate enforceable agreements with host communities. Benefits could include direct payments, guaranteed property-tax revenue, infrastructure improvements, apprenticeships, school funding, or compute access for local institutions like universities. Congress should also fund independent technical assistance so that small and rural jurisdictions can evaluate proposals and negotiate on equal footing with developers. Developers would be legally required to build, bring, or buy sufficient new generation; fund necessary grid upgrades; and make long-term financial commitments that protect customers from stranded costs, consistent with the White House’s Ratepayer Protection Pledge.

Diplomacy

Currently, the US AI capability lead is only about 8 months. Even if the US used all the policy tools at its disposal to extend its lead to over two years, that lead would still be fragile because the weights of the leading AI model could always be stolen. Ultimately, rapid progress in training compute efficiency — which may be increasing at a rate of 10x per year or more — suggests that new high-risk AI capabilities will eventually diffuse widely.44 At that point, Chinese AI companies may face the same risks as US developers.

If the risks of AI R&D automation materialize in either country, they would not be held back by national borders. At that point, it would be in the interest of both countries to cooperate.

It is therefore valuable to lay the foundation for future international cooperation on addressing the risks posed by AI R&D automation. While there will be challenges to US-China cooperation on AI, the US and the Soviet Union managed to cooperate on nuclear non-proliferation during the Cold War (in part thanks to zero-trust verification methods) even as they competed in other domains.

22. The US government should use its bilateral AI dialogue with China to jointly develop guidelines for managing risks from rapid AI capability growth and prepare verification measures

The Trump Administration’s September 2026 AI risk dialogue presents an opportunity for preparatory work in case a substantial increase in AI risks from increasing AI R&D automation ever occurs. At the dialogue, the US and China should discuss two areas particularly relevant to AI R&D automation risks.

First, the US and China can agree to develop and harmonize guidelines for automated AI R&D risk management. There is precedent for such an agreement; both US and Chinese AI companies signed the Frontier AI Safety Commitments at the AI Seoul Summit in 2024.

Second, to inform their respective domestic development of verification technologies (as discussed above), the US and China can begin discussions on which technologies they would use and in what contexts (e.g., bilateral or third-party auditing).

23. Countries with national AI institutes should collaborate on automated AI R&D risk management guidelines and technical capacity for AI verification

While the US and China are the most crucial countries in managing risks from AI R&D automation, other countries also have an important role to play. For one, AI capabilities will proliferate widely, meaning third-country AI developers will eventually be able to build increasingly advanced AI systems of their own. For another, both US and Chinese developers use large amounts of AI compute in third countries; these resources cannot be ignored in a verification arrangement.

The International Network for Advanced AI Measurement, Evaluation, and Science, whose members all have a national AI institute,45 should cooperate with other countries that have also established similar institutes, such as China, to harmonize risk management guidelines and verification techniques. Countries with a national AI institute are best positioned to contribute to these tasks with technical expertise. Once these countries set best practices, they can engage with broader multilateral groups to encourage wider adoption.

Subscribe now

Share