The economists' skepticism on AI growth might be the most interesting data point in this whole roundup. The gap between capability forecasts (where all groups agree) and economic impact forecasts (where they diverge) is basically telling us we're in the Solow Paradox again. Every GPT follows this pattern: impressive tech, bolted onto existing workflows, disappointing macro numbers, then decades of institutional redesign before the real payoff. I wrote a longer take here: https://noahpinionot.substack.com/p/you-can-see-ai-everywhere-except
Canada is usually the best comparator for the US, because its culture is the most similar by a fair margin. Ignoring its policy successes wastes a lot of time.
Noah: In general I'm a big fan. But here is a good example of why critics who accuse folks like you of hyping AI have a point:
You write "The “moderate” scenario here would have AI able to write high-quality novels, handle coding tasks that would take humans five days, create semi-autonomous labs, and use robots to perform basic household tasks. ... Why do economists think that even near-godlike AI wouldn’t translate into fast growth?"
The idea that the ability to write "high quality novels," create semi-autonomous labs, and perform basic household tasks is remotely "godlike" is absurd.
The AI effects on vulnerability finding are mostly a good thing. The AI being used to find vulnerabilities will result in fewer vulnerabilities being there to be found and hence fewer attacks happening. Vulnerabilities show up in a 1/x distribution over time so skipping ahead a couple decades is on net a good thing even though right now it's a bit of a python swallowing a horse.
As for AI in quantitative strategies, they're already dominating there and have been for years. You don't hear about it because quants are very secretive about their strategies.
Noah, I think your readers are saturated about punditry on AI. Most of us who have been trying to use AI for work don’t think it is such a big deal. There will be improvements and it was be another tool for marginal productivity gains.
Please write about something else. I haven’t been reading your blog for months because it just keeps talking about fantastical AI scenarios.
Thanks yea, I don’t code. I try to read and talk to people and understand things and make decisions and convince others. AI is marginally helpful in my job.
“The use of AI in the Iran war has come under scrutiny. The Pentagon has refused to say whether or not artificial intelligence was used in the Feb. 28 bombing of an elementary school in Iran that killed 175 people, many of them children.”
Garbage in garbage out. Outdated data used to train AI models was the direct cause of two missiles that targeted and destroyed two schools in Iran, killing hundreds of children and their teachers who made the fatal mistake of going to school.
What would be the reaction if terrorists bombed two elementary schools in the U.S., killing all the children and their teachers?
America is now a country with the Big Tech Bros Musk, Steve Davis, and DOGE High Tech babies killing hundreds of thousands of people, via preventable deaths, by dismantling USAID. Lest we forget, Marc Rubio signed off on this administrative genocide, which is projected to kill as many as 14 million people — one-third of them children — by the end of 2030. That this doesn’t even merit current mainstream news cycles speaks volumes about the moral rot of America.
Today, millions of Americans will attend Easter Sunday services at thousands of churches, blissfully unaware or uncaring about their country’s administrative genocide.
And the mainstream headlines will tout Elon Musk’s potential to become the world’s first trillionaire via the SpaceX IPO. Musk is the architect of the dismantling of USAID, and administrative genocide, which he found funny, posting: “We could have gone to parties this weekend. Instead we tossed USAID into the wood chipper.” According to witnesses, DOGE High Tech baby, Luke Ferritor, giggled over his keyboard as he dismantled USAID.
Most of these DOGE operatives are embedded in the federal government and on the payroll.
Rest assured that on Easter Sunday your tax dollars are at work killing hundreds of thousands of children. Apparently, Musk, Sam Altman, et alia are happy at their work, building “God in a box.”
Hitting a school by mistake when it's right next to a military target isn't particularly concerning. It's even less concerning when the country in question has been firing missiles at random civilian targets in countries that have nothing to do with the war.
The US probably isn't the only nation capable of distributing aid on the planet. If China is a bit smart, they will fill in the void, thus gaining some soft power in poorer countries. So will other countries.
That said, beware of labeling people as child killers just because they don't render international aid. How much do you spend on aid yourself, from your own pocket? Perhaps you could give 10-15 per cent of your personal income towards saving people in the underdeveloped world, instead of spending that money on, say, your personal vacations and other unnecessary expenses.
If you don't, are you a child killer? Then the entire neighbourhood you live in is a Child Killer City.
Let's assume for a minute that Mahajan's argument that it is very difficult for an advanced AI to design an effective bioweapon on the first shot is incorrect. You will then need to actually build the weapon, which means you have to either:
1. control your own lab capable of manufacturing it
2. trick another lab into manufacturing it
Are either of these feasible? Can the AI design a bioweapon such that the lab getting the order to manufacture it will not know? If #1 is not feasible, but #2 is, the obvious defense would be to run every incoming lab order through an AI. I doubt the AI can create a bioweapon that a similar AI would not recognize as dangerous.
And if the weapon is manufactured and released I see one more possible control point. The fact that the attacker possesses an AI that can one shot a bioweapon means that basically everyone in the world has access to the same AI to manufacture a vaccine or treatment. And the people on the defending side will have access to likely millions of times more compute than the attacker. Again, I don't know enough about the subject to say if this works. Does the ability to design a virus imply the ability to design a vaccine?
I know a lot more about cybersecurity than I do about biotech, and I'm not particularly concerned about the AI hacker scenario. A jailbroken Claude isn't particularly concerning because it just gives the attackers the same capability that the defenders already have, and the defenders will still have order of magnitude more compute power. This would only be concerning if the AI has an asymmetry where it is much better at attacking than defending. But in this case attacking and defending is basically the same thing. You tell the agent to find vulnerabilities. The only difference is whether you take those found vulnerabilities and patch them vs if you take advantage of them. So I don't really see why a malicious actor in possession of an agent could outpace the white hats using the same tools.
The quantum scenario is also not very worrying at the moment. According the abstract of that paper, the hypothetical 26,000 qubit machine that could crack the elliptic curve encryption is 4x larger than the largest existing quantum computer, and to crack RSA-2048, which it the most commonly used in the payments industry, would have to be 100x larger than that.
I think the AI cybersecurity nightmare is going to come from a very different direction. They won't be hacking computers, but rather hacking people. AI is capable of generating personalized phishing attacks that will be massively more convincing to the target than the spammed out garbage that they have today. AI can write an email that is relevant to a current conversation you are involved in. It can write the email in the same style as your boss. It can even generate a voice that sounds like you boss when you call the phone number in the email. The type of well prepared individually targeted phishing that we see now for high value targets will be able to go out to millions of people at a time.
On “being satisfied”, consider: my impression of the world five years ago was that pretty much every single decently useful software-related product idea got made. And often lavishly funded on the way. So…there did not seem to be problematic friction in conceiving, creating, and publishing great software ideas. Is it obviously a huge boost to spend trillions on something that makes water more wet? Likewise, we already published more novels per year than could be successful in the market. The supply of novels is fine. Do we want a server farm the size of Phoenix to make novels easier to write? One could respond that this is a general purpose technology - but most of the things it is really good at seem to be in the nature of making more abundant something that we already have enough of…. Not that it isn’t “cool” and “fun”! And I personally appreciate the improved voice control capacity that it has enabled…. Is there something that we truly do not have enough of that AI will supply? I mean, will it help fusion power happen?
Nonsense. Accurate data bases are updated for a reason: infrastructure changes with time. What was once a generic building was retrofitted for an elementary school. The U.S. has intentionally been bombing civilian infrastructure. What’s concerning that the U.S. has sunk to the moral level of Iranian leadership. I note how the systematic killing of “random” millions is ignored, set aside. The U.S. killing far more people at random — one-third children — via preventable deaths, than Iran could hope to kill. The millions of children who will die because of the dismantling of USAID, an administrative genocide, have nothing to do with any war. There is no excuse or argument for this atrocity. If one can’t draw a line here, one can’t draw a line anywhere. History will make this Elon Musk’s legacy: world’s richest man, working for the U.S. government, kills more people than the Nazi Death Camps system. Musk thinks this is funny and continues to be admired by millions of Americans. The current moral rot of America can’t be erased from history. I have no interest in catching a falling knife.
so is Tuesday the day the US commits a new set of war crimes and bombs civilian energy facilities in Iran for a war we are losing or is that gong to be for when he is even more desperate? Stay tuned.
As I understand it, Eric's obligations at a16z leave him too little time to podcast. There is likely also a fear of conflicting interests there. Would be cool to have Noah partner up with someone else that could take up the interviewing role. I really enjoy Econ102 and would regularly recommend it to my students as one of the best econ podcasts out there.
The two reforms we need to channel trading towards valuable activities (price discovery, and the hedging of risks), rather than pure zero-sum rent-seeking, are a small Tobin tax (say, 0.1% on each transaction) and quantization of trade times to something reasonably large compared to how fast computers can now operate. So, rather than everything operating in real-time, the market is a series of reverse auctions that end, say, every minute. There still might be SOME room for mischief with bids being slipped in during the last few milliseconds of an auction period, but if bids aren't public until the auction settles, this is probably OK.
You are assuming incorrectly that price discovery and hedging are the only valuable activities in the market. HFTs provide the market making function today. 30 years ago this same function existed, but it consisted of groups of human traders on the exchange floor screaming and shoving each other around.
And the benefit of the current system is quite tangible. In 1996 stocks didn't even trade in prices you would see anywhere else. Stocks had prices like $25 1/8, and therefore the bid/ask spread was at a minimum 12.5 cents. Now it's usually 1 cent or less, and a HFT skimming a penny off a trade is vastly superior to a human trader skimming dimes and quarters.
What would be the benefit to the reverse auction system vs buying and selling in the market we have now? I don't really see one, and I think if there were a benefit to trading in this way somebody would have already built a market that trades in that manner.
The point is to intentionally slow down the market, so we don’t have a Red Queen’s Race, with people investing billions of dollars in projects like drilling a precisely-straight tunnel from Chicago to New York, so they can run a fiber optic line that lets them gain a microsecond on the next guy.
If you're interested in conversations about AI biosecurity, Joe Betts-LaCroix of Retro Bio in Redwood City might be a good resource if he has the time. He has extensive experience in computer science, biology, and now AI, as Retro Bio has a special partnership with Open AI. I don't know if he has much experience in virology, but any conversation with him would be fascinating.
I will bet anything that developed country consumers are not satisfied. As a thought experiment, look at what people say they would do if they won the lottery and work backwards from that. The complication is that perhaps a lot of what people would demand is more leisure, which would have both positive and negative effects on GDP.
I think most of the forecasters are suffering from status quo bias. These exercises are better than nothing, but it’s like asking economists in 1995 to predict employment for social media managers and influencers in 2025. This is a shorter timespan, but I expect the changes to be at least as dramatic. If we don’t think plentiful cognition and robots can generate more growth than we see in regular countries right now (or Houston, growing at 10% annually), it seems like something is not computing.
AI is of greatest interest to wealthy people, especially corporate executives of tech firms. They control vast capital. They see their personal interests best served by putting that capital into AI. That same capital could instead go into affordable housing and public transit - which ordinary people need and which CEOs do not care about. Larger high end tax rates can enable such a change in the use of capital.
Not sure I'd be that worried about the quant trading concern: it only works if there is money to be made, and as they use more compute, it gets more expensive to squeeze dollars out (especially, as presumably the cost of electricity rises), thus there is a natural stopping function ...
I think you underestimate the size of the market, and the intense greed and competitiveness on Wall Street, where traders have laid their own fiberoptic line from New York to Chicago to shave milliseconds off the time it takes to get information, and where the big boys have been been honing their algorithms for automated trading for decades.
In a market which can lose half a trillion dollars in minutes from a tweet about a bogus attack on the Pentagon, the few million it takes to run their machines is hardly an obstacle. In fact, the new AI engines may be able to control both ends of such transactions, and the time-honored processes of "pump and dump" and of distributing false information to move stock prices may jump to new levels of market manipulation.
The economists' skepticism on AI growth might be the most interesting data point in this whole roundup. The gap between capability forecasts (where all groups agree) and economic impact forecasts (where they diverge) is basically telling us we're in the Solow Paradox again. Every GPT follows this pattern: impressive tech, bolted onto existing workflows, disappointing macro numbers, then decades of institutional redesign before the real payoff. I wrote a longer take here: https://noahpinionot.substack.com/p/you-can-see-ai-everywhere-except
I appreciated several of your points there.
Canada is usually the best comparator for the US, because its culture is the most similar by a fair margin. Ignoring its policy successes wastes a lot of time.
Noah: In general I'm a big fan. But here is a good example of why critics who accuse folks like you of hyping AI have a point:
You write "The “moderate” scenario here would have AI able to write high-quality novels, handle coding tasks that would take humans five days, create semi-autonomous labs, and use robots to perform basic household tasks. ... Why do economists think that even near-godlike AI wouldn’t translate into fast growth?"
The idea that the ability to write "high quality novels," create semi-autonomous labs, and perform basic household tasks is remotely "godlike" is absurd.
The AI effects on vulnerability finding are mostly a good thing. The AI being used to find vulnerabilities will result in fewer vulnerabilities being there to be found and hence fewer attacks happening. Vulnerabilities show up in a 1/x distribution over time so skipping ahead a couple decades is on net a good thing even though right now it's a bit of a python swallowing a horse.
As for AI in quantitative strategies, they're already dominating there and have been for years. You don't hear about it because quants are very secretive about their strategies.
Noah, I think your readers are saturated about punditry on AI. Most of us who have been trying to use AI for work don’t think it is such a big deal. There will be improvements and it was be another tool for marginal productivity gains.
Please write about something else. I haven’t been reading your blog for months because it just keeps talking about fantastical AI scenarios.
Speak for yourself. I don't write code anymore. I just supervise AI now. Calling it a big deal would be an understatement.
Thanks yea, I don’t code. I try to read and talk to people and understand things and make decisions and convince others. AI is marginally helpful in my job.
“The use of AI in the Iran war has come under scrutiny. The Pentagon has refused to say whether or not artificial intelligence was used in the Feb. 28 bombing of an elementary school in Iran that killed 175 people, many of them children.”
Garbage in garbage out. Outdated data used to train AI models was the direct cause of two missiles that targeted and destroyed two schools in Iran, killing hundreds of children and their teachers who made the fatal mistake of going to school.
What would be the reaction if terrorists bombed two elementary schools in the U.S., killing all the children and their teachers?
America is now a country with the Big Tech Bros Musk, Steve Davis, and DOGE High Tech babies killing hundreds of thousands of people, via preventable deaths, by dismantling USAID. Lest we forget, Marc Rubio signed off on this administrative genocide, which is projected to kill as many as 14 million people — one-third of them children — by the end of 2030. That this doesn’t even merit current mainstream news cycles speaks volumes about the moral rot of America.
Today, millions of Americans will attend Easter Sunday services at thousands of churches, blissfully unaware or uncaring about their country’s administrative genocide.
And the mainstream headlines will tout Elon Musk’s potential to become the world’s first trillionaire via the SpaceX IPO. Musk is the architect of the dismantling of USAID, and administrative genocide, which he found funny, posting: “We could have gone to parties this weekend. Instead we tossed USAID into the wood chipper.” According to witnesses, DOGE High Tech baby, Luke Ferritor, giggled over his keyboard as he dismantled USAID.
Most of these DOGE operatives are embedded in the federal government and on the payroll.
Rest assured that on Easter Sunday your tax dollars are at work killing hundreds of thousands of children. Apparently, Musk, Sam Altman, et alia are happy at their work, building “God in a box.”
Hitting a school by mistake when it's right next to a military target isn't particularly concerning. It's even less concerning when the country in question has been firing missiles at random civilian targets in countries that have nothing to do with the war.
People probably underestimate just how many such targets used to be hit in the 20th century.
It is a bit remarkable that since the early phase of the war, nothing of that sort happened again.
The US probably isn't the only nation capable of distributing aid on the planet. If China is a bit smart, they will fill in the void, thus gaining some soft power in poorer countries. So will other countries.
That said, beware of labeling people as child killers just because they don't render international aid. How much do you spend on aid yourself, from your own pocket? Perhaps you could give 10-15 per cent of your personal income towards saving people in the underdeveloped world, instead of spending that money on, say, your personal vacations and other unnecessary expenses.
If you don't, are you a child killer? Then the entire neighbourhood you live in is a Child Killer City.
Let's assume for a minute that Mahajan's argument that it is very difficult for an advanced AI to design an effective bioweapon on the first shot is incorrect. You will then need to actually build the weapon, which means you have to either:
1. control your own lab capable of manufacturing it
2. trick another lab into manufacturing it
Are either of these feasible? Can the AI design a bioweapon such that the lab getting the order to manufacture it will not know? If #1 is not feasible, but #2 is, the obvious defense would be to run every incoming lab order through an AI. I doubt the AI can create a bioweapon that a similar AI would not recognize as dangerous.
And if the weapon is manufactured and released I see one more possible control point. The fact that the attacker possesses an AI that can one shot a bioweapon means that basically everyone in the world has access to the same AI to manufacture a vaccine or treatment. And the people on the defending side will have access to likely millions of times more compute than the attacker. Again, I don't know enough about the subject to say if this works. Does the ability to design a virus imply the ability to design a vaccine?
I know a lot more about cybersecurity than I do about biotech, and I'm not particularly concerned about the AI hacker scenario. A jailbroken Claude isn't particularly concerning because it just gives the attackers the same capability that the defenders already have, and the defenders will still have order of magnitude more compute power. This would only be concerning if the AI has an asymmetry where it is much better at attacking than defending. But in this case attacking and defending is basically the same thing. You tell the agent to find vulnerabilities. The only difference is whether you take those found vulnerabilities and patch them vs if you take advantage of them. So I don't really see why a malicious actor in possession of an agent could outpace the white hats using the same tools.
The quantum scenario is also not very worrying at the moment. According the abstract of that paper, the hypothetical 26,000 qubit machine that could crack the elliptic curve encryption is 4x larger than the largest existing quantum computer, and to crack RSA-2048, which it the most commonly used in the payments industry, would have to be 100x larger than that.
I think the AI cybersecurity nightmare is going to come from a very different direction. They won't be hacking computers, but rather hacking people. AI is capable of generating personalized phishing attacks that will be massively more convincing to the target than the spammed out garbage that they have today. AI can write an email that is relevant to a current conversation you are involved in. It can write the email in the same style as your boss. It can even generate a voice that sounds like you boss when you call the phone number in the email. The type of well prepared individually targeted phishing that we see now for high value targets will be able to go out to millions of people at a time.
On “being satisfied”, consider: my impression of the world five years ago was that pretty much every single decently useful software-related product idea got made. And often lavishly funded on the way. So…there did not seem to be problematic friction in conceiving, creating, and publishing great software ideas. Is it obviously a huge boost to spend trillions on something that makes water more wet? Likewise, we already published more novels per year than could be successful in the market. The supply of novels is fine. Do we want a server farm the size of Phoenix to make novels easier to write? One could respond that this is a general purpose technology - but most of the things it is really good at seem to be in the nature of making more abundant something that we already have enough of…. Not that it isn’t “cool” and “fun”! And I personally appreciate the improved voice control capacity that it has enabled…. Is there something that we truly do not have enough of that AI will supply? I mean, will it help fusion power happen?
Nonsense. Accurate data bases are updated for a reason: infrastructure changes with time. What was once a generic building was retrofitted for an elementary school. The U.S. has intentionally been bombing civilian infrastructure. What’s concerning that the U.S. has sunk to the moral level of Iranian leadership. I note how the systematic killing of “random” millions is ignored, set aside. The U.S. killing far more people at random — one-third children — via preventable deaths, than Iran could hope to kill. The millions of children who will die because of the dismantling of USAID, an administrative genocide, have nothing to do with any war. There is no excuse or argument for this atrocity. If one can’t draw a line here, one can’t draw a line anywhere. History will make this Elon Musk’s legacy: world’s richest man, working for the U.S. government, kills more people than the Nazi Death Camps system. Musk thinks this is funny and continues to be admired by millions of Americans. The current moral rot of America can’t be erased from history. I have no interest in catching a falling knife.
so is Tuesday the day the US commits a new set of war crimes and bombs civilian energy facilities in Iran for a war we are losing or is that gong to be for when he is even more desperate? Stay tuned.
We're not "losing" the war. in Iraq. You can't lose a war that has no definition of "winning".
Why is the podcast over? Its the only reason i first found economics interesting / understood anything about it.
As I understand it, Eric's obligations at a16z leave him too little time to podcast. There is likely also a fear of conflicting interests there. Would be cool to have Noah partner up with someone else that could take up the interviewing role. I really enjoy Econ102 and would regularly recommend it to my students as one of the best econ podcasts out there.
The two reforms we need to channel trading towards valuable activities (price discovery, and the hedging of risks), rather than pure zero-sum rent-seeking, are a small Tobin tax (say, 0.1% on each transaction) and quantization of trade times to something reasonably large compared to how fast computers can now operate. So, rather than everything operating in real-time, the market is a series of reverse auctions that end, say, every minute. There still might be SOME room for mischief with bids being slipped in during the last few milliseconds of an auction period, but if bids aren't public until the auction settles, this is probably OK.
You are assuming incorrectly that price discovery and hedging are the only valuable activities in the market. HFTs provide the market making function today. 30 years ago this same function existed, but it consisted of groups of human traders on the exchange floor screaming and shoving each other around.
And the benefit of the current system is quite tangible. In 1996 stocks didn't even trade in prices you would see anywhere else. Stocks had prices like $25 1/8, and therefore the bid/ask spread was at a minimum 12.5 cents. Now it's usually 1 cent or less, and a HFT skimming a penny off a trade is vastly superior to a human trader skimming dimes and quarters.
What would be the benefit to the reverse auction system vs buying and selling in the market we have now? I don't really see one, and I think if there were a benefit to trading in this way somebody would have already built a market that trades in that manner.
Re: whether anyone has designed markets to use the reverse auction design, I don't know that anyone's done this _to the exclusion_ of the continuous order-book style, but plenty of markets have experimented with periodic auctions at various period lengths. See, for instance: https://www.cboe.com/insights/posts/how-periodic-auctions-enhance-trading-in-europe-and-the-u-s/
The concept of doing this with the intent of ending the arms race in high-speed trading is not original to me. Here's an academic treatment of the idea from a professor at Chicago Booth: https://ericbudish.org/publication/the-high-frequency-trading-arms-race-frequent-batch-auctions-as-a-market-design-response/
The point is to intentionally slow down the market, so we don’t have a Red Queen’s Race, with people investing billions of dollars in projects like drilling a precisely-straight tunnel from Chicago to New York, so they can run a fiber optic line that lets them gain a microsecond on the next guy.
In frank Herbert due novel's it is a sin to create a machine in imitation of a human soul.
With the result being 10,000 years of dystopian feudalism. It's not clear the right side won in the novels written by Frank Herbert himself.
One could make the case since the Arab oil embargo the US is a feudal society
Same in the Warhammer 40k universe. There the machines have their own spirits that you have to pray to. Blessings of the Omnissiah. 🙏
If you're interested in conversations about AI biosecurity, Joe Betts-LaCroix of Retro Bio in Redwood City might be a good resource if he has the time. He has extensive experience in computer science, biology, and now AI, as Retro Bio has a special partnership with Open AI. I don't know if he has much experience in virology, but any conversation with him would be fascinating.
I will bet anything that developed country consumers are not satisfied. As a thought experiment, look at what people say they would do if they won the lottery and work backwards from that. The complication is that perhaps a lot of what people would demand is more leisure, which would have both positive and negative effects on GDP.
I think most of the forecasters are suffering from status quo bias. These exercises are better than nothing, but it’s like asking economists in 1995 to predict employment for social media managers and influencers in 2025. This is a shorter timespan, but I expect the changes to be at least as dramatic. If we don’t think plentiful cognition and robots can generate more growth than we see in regular countries right now (or Houston, growing at 10% annually), it seems like something is not computing.
AI is of greatest interest to wealthy people, especially corporate executives of tech firms. They control vast capital. They see their personal interests best served by putting that capital into AI. That same capital could instead go into affordable housing and public transit - which ordinary people need and which CEOs do not care about. Larger high end tax rates can enable such a change in the use of capital.
Not sure I'd be that worried about the quant trading concern: it only works if there is money to be made, and as they use more compute, it gets more expensive to squeeze dollars out (especially, as presumably the cost of electricity rises), thus there is a natural stopping function ...
I think you underestimate the size of the market, and the intense greed and competitiveness on Wall Street, where traders have laid their own fiberoptic line from New York to Chicago to shave milliseconds off the time it takes to get information, and where the big boys have been been honing their algorithms for automated trading for decades.
In a market which can lose half a trillion dollars in minutes from a tweet about a bogus attack on the Pentagon, the few million it takes to run their machines is hardly an obstacle. In fact, the new AI engines may be able to control both ends of such transactions, and the time-honored processes of "pump and dump" and of distributing false information to move stock prices may jump to new levels of market manipulation.