Thank you so much for this article, Noah. I actually know a little about AI, having developed early Internet apps years ago and still playing around with basic machine learning systems today (i teach robotics among other things). Listening to the histrionics over the last few days has been hilarious but also disturbing, becasue emotional reactions yield poor policy.
Deepseek has simply managed to apply Moore's Law to a particular class of hardware and software everyone thought was immune to it. That's it. Why are we surprised that Moore's Law still applies?
It's not a revolution. I suspect DeepSeek is lying about the chips it used. I also suspect their "open source" released code may not be complete (they kept a few trade secrets; that would be the Chinese way.) But they have made a marginal improvement and others (incl US developers) will copy it to make better LLM systems. I also vaguely suspect the whole thing may be a psy-op from the CCP to see how the Trump admin reacts. (That would also be the Chinese way.)
But big picture, this is how CS has always worked. In my day, bored college students would write cool code and publish it on Usenet, and those of us working in the business would grab stuff that looked interesting and tweak it for our needs. That was where GPL came from. This is no different.
Thanks for rationality at a time when so many are lacking it, Noah.
Moore's Law is about putting more transistors on a chip. Deepseek has nothing to do with that. They're using commodity chips, and they came up with innovations in model training and piggybacked on other model outputs and concepts.
I don't know how to evaluate my hunch, but wouldn't OpenAI notice DeepSeek exfiltrating a large amount of data from its interface? Wouldn't they notice that extraordinary amount of activity?
I don't think it is quite so cut and dry that they noticed it. According to the FT
"The San Francisco-based ChatGPT maker told the Financial Times it had seen some evidence of “distillation”, which it suspects to be from DeepSeek. OpenAI declined to comment further or provide details of its evidence."
That's not exactly "we have concrete evidence it was Deep Seek". It is way too early for anyone to be speculating about any of this really. And is there even a benefit to race to speculate?
I wouldn't be surprised if they used a botnet, made up of hacked devices across the US and allies. Have each bot create an account, and break out your queries among them. Detecting it becomes tricky -- you're not getting a flood of queries from one source, you have to recognize a pattern of distributed queries that are trying to probe different aspects of the model.
And in a world of Internet-of-Things devices controlled by custom Linux builds that have been hacked together just to get them out to market, without worrying about security, things are only getting worse. And a ton of electronics is being bought _from China_, and G*d knows how many of those are basically already set up with a backdoor that lets China's state hackers use their spare cycles.
Apparently CloudFlare had a new IoT botnet DDoS just a few days ago.
If export controls would significantly slow down the ability of Chinese AI companies to stay competitive, might that make an invasion of Taiwan more likely - to take control of TSMC? The US oil embargo in 1941 pushed Japan into a reckless act of war. The stakes for China in this obviously aren't as high as they were for Japan in 1941, but then the risks of the military action are unlikely to be as high either - invading a small, nearby country that you have claimed is yours for 75 years vs. taking on an emerging superpower and its superpower allies with huge vested interests in the region. That might be a wager that they like the look of. Hope not.
I always assumed that in the event of a successful invasion of Taiwan somebody would ensure that the local TSMC was permanently incapacitated. It would be foolish to allow it to be captured.
Imagine instead of the recent low-Richter scale earthquakes on Taiwan a major +7 Richter earthquake that takes down TSMC and ASML manufacturing facilities and corporate campuses. U.S. car manufacturers will be parking a lot more new, unfinished (chipless) cars. Trailing-edge and leading-edge chips supply would be impacted, including smart phones to toasters to military weapons OEMs. Now imagine if Xi made the bad decision to invade/occupy Taiwan. The kill switches installed in chip fabs would be flipped. A significant stock market correction and/or crash would be triggered on multiple exchanges across the world. The most-valued chips would be trailing-edge 20nm, etc. chips, not 5nm, 3nm, and 2nm chips. Maybe an earthquake would beat Xi to the punch. Either way, the result would be the same.
TSMC manufactures chips for a huge range of applications. However, very few applications need the <20nm sizes that are produced by very few fabs (including TSMC). I would be shocked if anything in a car needs that kind of fine geometry (<20nm) technology. All of the dozen or so ICs that I have designed are on 180nm or larger technologies (2 at TSMC).
The chances of the TSMC fabs surviving a Chinese naval invasion seem pretty small. From the point of view of commerce a chip fab is a multi-billion dollar investment containing cutting edge technology. From the point of view of naval warfare a chip fab is an unprotected target that could be destroyed by a surprisingly small number of anti-ship missiles.
Japan could capture oil wells and make it work (even though they couldn't transport oil to their metropole later on). As for China? Well, if TSMC is attacked, then the Chinese wouldn't be able to use its factories since they would be blown up.
I believe that Xi knew about this already, so if he attacks Taiwan, he would base his casus belli on some other reasons.
So Taiwan have a sort of scorched earth contingency plan, if attacked? Basically don't leave anything standing that could be of value to the Chinese invader. And they'd need TSMC co-operation to replicate what TSMC does (as the US is currently doing). But they're not going to get that. So their AI challenge could be a flash in the pan.
I read the Team Vizuara explanation which is so technical I am not sure I got all of it but essentially Team DeepSeek had one or two brilliant ideas which could be legitimately invented by a few engineers at a hedge fund.
"The idea that the emergence of superintelligence would immediately lead to an apocalyptic cyberwar between the U.S. and China is, of course, pure science fiction. I see no reason to believe that a superintelligence breakthrough would cause war within a matter of months. Launching such a war would be too risky for everyone involved, so the U.S. would not do it. The clear precedent here is nuclear weapons, which the U.S. got in 1945 and the USSR only got in 1949. There were four years in which America could have preemptively nuked and destroyed the Soviet Union, but we didn’t. If the U.S. had a several-month lead or even a several-year lead over China in superintelligence, we would probably not try to attack and destroy China with it, because contrary to what you might read on X, we’re not insane warmongering psychos. And thus China would get superintelligence too, just as the USSR got the atom bomb."
This is misunderstanding superintelligence. If we get *actual* superintelligence, it will most likely take over the world on its own in order to achieve whatever goals it happens to have - and if humans don't like those goals, they'll have as much success opposing it as unassisted humans do at beating Stockfish at chess.
"If we get *actual* superintelligence, it will most likely take over the world on its own in order to achieve whatever goals it happens to have "
While I find the arguments for this to be compelling (instrumental convergence, reward hacking, etc), we should have a healthy level of uncertainty around any technology that doesn't exist yet: how it will be implemented, what its constraints will be, what the second-order effects will be, and so on.
Also worth noting that von Neumann was of the opinion that we *should* have used the bomb to destroy the USSR, and he’s probably among the better human analogues for AGI / ASI (although I should note that comparing humans to ASIs is an instance in which metaphors can easily do more harm than good.)
It's a fascinating point. Having said that, his ideas at that time seem radically underbaked. Nukes weren't that powerful or accurate then, so it's not clear how much damage the US could have done preemptively. The Soviet Union would then have had carte blanche to overrun Europe, and US soft power would have been disintegrated. If and when we get ASI, hopefully it comes up with better ideas than that.
"There were four years in which America could have preemptively nuked and destroyed the Soviet Union, but we didn’t. "
Actually nuclear bombs during that era were not that powerful - the main power that America and its allies have over the Soviets were their massive industrial capability (in WW2 it had higher GDP than all great powers in the Axis combined), as well as their logistical capability and their overwhelming airforce; compared with a nation that already lost 1/8 of its population and faced large-scale famine post-war.
However the Soviets at that time had probably the strongest army in Europe, not to mentioned the popularity of communist parties in Europe at that time (Czechoslovakia and France both had communist parties with the largest share of votes in parliaments), so attacking Soviet Union at that time could be successful, but you would see millions of body bags sent to grieving families in the US though.
You could check "Operation Unthinkable" to see what would happen if the Allies attacked the Soviets in 1945!
Noah Smith has mentioned in his podcast that a military alliance with superior manufacturing capability has "escalation dominance" over one that doesn't, and that during the Cold War the US would *eventually* win any non-nuclear war with the Soviet Union because it could keep building tanks, planes, etc. longer than the Soviet Union could.
This is basically not true. The Soviet Union had a big advantage over the United States in any war fought in Europe: it didn't have to send its military across the Atlantic Ocean to get to the battlefield. President Eisenhower concluded that maintaining a non-nuclear military force that could actually defend West Germany from a Soviet invasion would be so expensive (in terms of both money and physical resources) as to make it impractical, and instead chose to rely almost entirely on nuclear deterrence instead. Perhaps in the long run the United States could undergo a massive WWII style military buildup and have a second D-Day landing, but before that, West Germany and most of the rest of Western Europe would have fallen, just like it did when Hitler's army invaded.
Niall Ferguson, in his Civilization book, actually mentioned that if the Cold War went hot (not nuclear, of course), the Soviets would have an upper hand and pretty good chance to win the war; after all, they had larger armies, didn't need to ferry the army through Atlantic Ocean, had strong arms industries; but most importantly, the economic system and the people there were more resilient against wartime hardship.
You could even see it nowadays: Russia is still sanctions-resistant after 3 years of war against Ukraine, and in some types of weapons like artillery shells, they outproduce Europe!
I just don’t view China in the same vein as you and Matt Ygelsias. China is not some major threat or the enemy to me. And while I’m not a huge communist fan, democracy has major issues too, as we can see, and I’m generally indifferent to how they want to be governed. I want the US to be competitive, but I truly don’t view Chinese companies (and perhaps even the CCP) as much worse than US corporations. They all solely care about their own interests.
And with the above said, how can I support tariffs that result in deeper pockets for a handful of US tech bros and higher prices and shittier products for the rest of us. I have zero loyalty to OpenAI, Zuck, Google, etc
This isn’t intended to be anti-capitalist at all, but it is to make the point that I do not align with “they must be stopped” and “we cannot let them win”. Deepseek wasn’t part of some massive CCP funding scheme that gave them an unfair advantage, which may warrant action from the US. No, by all accounts they had way less resources but were simply better. And before the arguments of “US must control AI”, this was open source.. China / DeepSeek literally gave a gift to the rest of us without any clear means of domination of the race or hiding their IP
You make some valid points. But China *is* bulking up militarily at an amazing rate. And a nation doesn't sink such massive resources into their military unless they're prepared to use it for war. And Taiwan will be the flash point. Or possibly their aggressive territorialism in the South China Sea. Which will produce an incident, that will snowball into war.
It's not like we're white knights or anything. As an empire we've done some really shitty things since 1945. But China wants their turn in the sun.
I think the argument that's convincing to me is that China is the country in the next 20 years who is the most likely to disrupt global peace, and therefore we should take take that into account.
An invasion of Taiwan would be so incredibly disruptive globally that its hard to really predict all the potential fallout from it.
If I may engage in some "distillation" -- I think it's possible to believe that China is an evil behemoth intent on world domination AND want the US political system to exert some influence and/or control over the techbro free market feeding frenzy at the Trump smorgasbord and national defense raffle. You are reacting (rightly!) to the obscenity of these creeps securing their own futures at our expense, using the cover of "national security", while nobody is willing to act responsibly and in the actual national interest. Two generations ago we understood that nuclear technology posed both a mortal threat to humanity AND an unprecedented opportunity for humanity to contain traditional wars of attrition and to produce energy. But, importantly, we also understood that the power and the promise of the technology made it a natural subject for regulation and guidance under a politically cogent and responsible set of rules, restrictions and oversight. We did NOT say -- fuck it, let everybody play around and make money with nukes however they want, just as long as the NASDAQ goes up!
For me, the techbros simultaneously warning us that (1) AI is the most important tech in the history of humanity and the stakes of "winning" are nothing less than global domination, but (2) "we can not allow any puny, silly mortals limit us our self-directed goals of pursuing world domination for ourselves", is deeply discordant and profoundly at odds with all previous approaches to world-changing technologies with military and national security significance.
It's a new, dangerous path, and it's disheartening to see that Noah seems to be willing to verge down that path. We can be sure that the Chinese government will never make that mistake...
In one post (more related to Biden-Trump) Noah actually said that you should not expect this blog to be non-partisan, so your expectations should be that you are looking at Noah's points of view, with his own biases!
I think I said this at Slow Boring as well, but you seem to have confused the tariffs on automobiles and manufactured goods (which produce deeper pockets for auto executives and auto workers in the US, Europe, Japan, and Korea) with the export controls on semiconductors (which don’t obviously product deeper pockets for anyone). These are both controls on international trade, but they are in opposite directions.
Proposals from the Trump admin relate to marrying export controls with new tariff actions. This supposedly will bring manufacturing back to the United States, but this seems to make the development of AI models costlier, which could lead to the U.S. becoming less competitive in the AI race. Can the U.S. sustain an innovation edge without over-regulating itself?
Noah, I have subscribed in order to make one comment. You mention many true facts about the situation surrounding AI. But the bigger picture is that we are racing towards the creation of nonbiological intelligences smarter than any human, and the logical consequence of that, is that humans will no longer control their own destiny. We are creating beings that will replace us completely, and yet the people making this happen can barely even acknowledge this. You're just a commentator, but this applies to you too. The logical outcome of the AI race is not that China wins or that America wins, it is that AI wins and does as it pleases with the human race. I can't take any commentator very seriously, whose big picture doesn't include that aspect.
Great article, only thing I would point out, let's say it cost X billion for OpenAI to make the model with Y ability that Deepthink distilled from; that method (as it exists today) does not get us to Y+1 unless someone puts that money in to raise the high water mark. OpenAI used distilling to release it's own slight worse than their premiere one but cheaper. I'm a complete skeptic of LLM version of "AI"; until we can increase true capability for less money (rather than drafting off a more capable expensive model) the LLM future is looking pretty grim. Seeing the insane amount of money OpenAI et al. are investing in hardware they are still going scale scale scale, to push the envelop to Y+1, but it's not efficient in any form given the billions being put in.
I think a new AI company called 2nd best but 1st cheapest should do the DeepThink ditstillation approach to all the capable models to get whatever % the ability at orders of magnitude cheaper cost to build, and sell the API access for a small fraction less of what OpenAI is selling it for (which is still losing money on). That seems to be where the profit lies.
Innovation is an economic imperative. Given the amount of energy that will be required for the continued annual rollout of 40 new data centers, a newly designed ASICS transformer chip with 20x more processing power and the attendant energy savings will be a Godsend for AI on U.S. soil. And a silicon anode battery with 30% more energy density, half the charge time of conventional commercial Li-ion batteries, longer battery life, and with Brakeflow technology to mitigate thermal runways, will put the U.S. in a competitive position in the Battery War with China. Note the recent fire that destroyed a commercial aircraft in South Korea. In 2023, there were 163 occurrences of Li-ion battery fires on commercial aircraft, to say nothing of 142 Li-ion battery fires for the NYFD, involving loss of lives and property.
“The compute gap between US and China—further widened by export controls —remains DeepSeek’s primary constraint. DeepSeek’s leadership openly acknowledged a 4x compute disadvantage despite their efficiency gains.”
Indeed. A U.S. company has designed a purpose-built ASICS transformer chip with 20x the processing power of NVDA’s GPU chips. Not only does this significantly reduce the latency, it also represents an exponential decrease in the energy requirements. This chip is likely to begin production in late 2025/early 2026.
Of course, NVDA can design/manufacture a competing chip, but this is about a three-year process. All the money in the world can’t buy time, and three years is a significant first-mover advantage in a Silicon world. The real competitive threat to NVDA is a much better chip. The good news is that a better chip design is happening in the U.S. sure, it’s a startup company, but a silicon chip veteran with three decades of experience at a major Silicon Valley semiconductor company, who came out of retirement, is leading the design team. As with process-manufacturing engineers, there are master craftsmen teaching the chip designers of today and tomorrow — in Taiwan and Silicon Valley. I think hardware is the key to the AI and Battery Wars.
I've heard all the foundries have kill switches. I don't know the details, of course, but I assume they're designed to make them robustly unusable in order to deter that line of thinking.
I am so confused by people claiming the DeepSeek launch is what had the big hit on NVIDIA stock. DeepSeek announced R1 a full week before that bad market day! Do people think the market analysts for the tech or chip sector were all on vacation for a week?
Thank you so much for this article, Noah. I actually know a little about AI, having developed early Internet apps years ago and still playing around with basic machine learning systems today (i teach robotics among other things). Listening to the histrionics over the last few days has been hilarious but also disturbing, becasue emotional reactions yield poor policy.
Deepseek has simply managed to apply Moore's Law to a particular class of hardware and software everyone thought was immune to it. That's it. Why are we surprised that Moore's Law still applies?
It's not a revolution. I suspect DeepSeek is lying about the chips it used. I also suspect their "open source" released code may not be complete (they kept a few trade secrets; that would be the Chinese way.) But they have made a marginal improvement and others (incl US developers) will copy it to make better LLM systems. I also vaguely suspect the whole thing may be a psy-op from the CCP to see how the Trump admin reacts. (That would also be the Chinese way.)
But big picture, this is how CS has always worked. In my day, bored college students would write cool code and publish it on Usenet, and those of us working in the business would grab stuff that looked interesting and tweak it for our needs. That was where GPL came from. This is no different.
Thanks for rationality at a time when so many are lacking it, Noah.
Moore's Law is about putting more transistors on a chip. Deepseek has nothing to do with that. They're using commodity chips, and they came up with innovations in model training and piggybacked on other model outputs and concepts.
I don't know how to evaluate my hunch, but wouldn't OpenAI notice DeepSeek exfiltrating a large amount of data from its interface? Wouldn't they notice that extraordinary amount of activity?
They did notice it! The question I don't know the answer to is how they could prevent stuff like that.
I don't think it is quite so cut and dry that they noticed it. According to the FT
"The San Francisco-based ChatGPT maker told the Financial Times it had seen some evidence of “distillation”, which it suspects to be from DeepSeek. OpenAI declined to comment further or provide details of its evidence."
That's not exactly "we have concrete evidence it was Deep Seek". It is way too early for anyone to be speculating about any of this really. And is there even a benefit to race to speculate?
I wouldn't be surprised if they used a botnet, made up of hacked devices across the US and allies. Have each bot create an account, and break out your queries among them. Detecting it becomes tricky -- you're not getting a flood of queries from one source, you have to recognize a pattern of distributed queries that are trying to probe different aspects of the model.
Sounds like that would work.
It remains disturbingly easy to build a truly enormous botnet. Possibly the largest known / documented one was built by teenagers.
https://www.wired.com/story/mirai-untold-story-three-young-hackers-web-killing-monster/
And in a world of Internet-of-Things devices controlled by custom Linux builds that have been hacked together just to get them out to market, without worrying about security, things are only getting worse. And a ton of electronics is being bought _from China_, and G*d knows how many of those are basically already set up with a backdoor that lets China's state hackers use their spare cycles.
Apparently CloudFlare had a new IoT botnet DDoS just a few days ago.
https://cyberscoop.com/cloudflare-biggest-ddos-attack-mirai-variant-botnet/
If export controls would significantly slow down the ability of Chinese AI companies to stay competitive, might that make an invasion of Taiwan more likely - to take control of TSMC? The US oil embargo in 1941 pushed Japan into a reckless act of war. The stakes for China in this obviously aren't as high as they were for Japan in 1941, but then the risks of the military action are unlikely to be as high either - invading a small, nearby country that you have claimed is yours for 75 years vs. taking on an emerging superpower and its superpower allies with huge vested interests in the region. That might be a wager that they like the look of. Hope not.
I'm not sure! It's an interesting question.
I always assumed that in the event of a successful invasion of Taiwan somebody would ensure that the local TSMC was permanently incapacitated. It would be foolish to allow it to be captured.
Imagine instead of the recent low-Richter scale earthquakes on Taiwan a major +7 Richter earthquake that takes down TSMC and ASML manufacturing facilities and corporate campuses. U.S. car manufacturers will be parking a lot more new, unfinished (chipless) cars. Trailing-edge and leading-edge chips supply would be impacted, including smart phones to toasters to military weapons OEMs. Now imagine if Xi made the bad decision to invade/occupy Taiwan. The kill switches installed in chip fabs would be flipped. A significant stock market correction and/or crash would be triggered on multiple exchanges across the world. The most-valued chips would be trailing-edge 20nm, etc. chips, not 5nm, 3nm, and 2nm chips. Maybe an earthquake would beat Xi to the punch. Either way, the result would be the same.
ASML doesn't manufacture in Taiwan.
Samsung and Intel are manufacturing at scale in the 5nm to 10nm range.
Apart from self-driving, my understanding is that most cars do not use chips made by TSMC.
TSMC manufactures chips for a huge range of applications. However, very few applications need the <20nm sizes that are produced by very few fabs (including TSMC). I would be shocked if anything in a car needs that kind of fine geometry (<20nm) technology. All of the dozen or so ICs that I have designed are on 180nm or larger technologies (2 at TSMC).
The chances of the TSMC fabs surviving a Chinese naval invasion seem pretty small. From the point of view of commerce a chip fab is a multi-billion dollar investment containing cutting edge technology. From the point of view of naval warfare a chip fab is an unprotected target that could be destroyed by a surprisingly small number of anti-ship missiles.
Japan could capture oil wells and make it work (even though they couldn't transport oil to their metropole later on). As for China? Well, if TSMC is attacked, then the Chinese wouldn't be able to use its factories since they would be blown up.
I believe that Xi knew about this already, so if he attacks Taiwan, he would base his casus belli on some other reasons.
So Taiwan have a sort of scorched earth contingency plan, if attacked? Basically don't leave anything standing that could be of value to the Chinese invader. And they'd need TSMC co-operation to replicate what TSMC does (as the US is currently doing). But they're not going to get that. So their AI challenge could be a flash in the pan.
In Material World by Ed Conway, there are proposals for TSMC to blow up its factories in case China invades and occupies them!
I read the Team Vizuara explanation which is so technical I am not sure I got all of it but essentially Team DeepSeek had one or two brilliant ideas which could be legitimately invented by a few engineers at a hedge fund.
Qwen was the Chinese "national champion" LLM as far as anyone knew until last week.
"The idea that the emergence of superintelligence would immediately lead to an apocalyptic cyberwar between the U.S. and China is, of course, pure science fiction. I see no reason to believe that a superintelligence breakthrough would cause war within a matter of months. Launching such a war would be too risky for everyone involved, so the U.S. would not do it. The clear precedent here is nuclear weapons, which the U.S. got in 1945 and the USSR only got in 1949. There were four years in which America could have preemptively nuked and destroyed the Soviet Union, but we didn’t. If the U.S. had a several-month lead or even a several-year lead over China in superintelligence, we would probably not try to attack and destroy China with it, because contrary to what you might read on X, we’re not insane warmongering psychos. And thus China would get superintelligence too, just as the USSR got the atom bomb."
This is misunderstanding superintelligence. If we get *actual* superintelligence, it will most likely take over the world on its own in order to achieve whatever goals it happens to have - and if humans don't like those goals, they'll have as much success opposing it as unassisted humans do at beating Stockfish at chess.
"If we get *actual* superintelligence, it will most likely take over the world on its own in order to achieve whatever goals it happens to have "
While I find the arguments for this to be compelling (instrumental convergence, reward hacking, etc), we should have a healthy level of uncertainty around any technology that doesn't exist yet: how it will be implemented, what its constraints will be, what the second-order effects will be, and so on.
Uncertainty is a reason to be more concerned about this possibility, not less.
Not sure I agree. Concern over a negative outcome should rise as the probability of that outcome rises, and vice versa.
Also worth noting that von Neumann was of the opinion that we *should* have used the bomb to destroy the USSR, and he’s probably among the better human analogues for AGI / ASI (although I should note that comparing humans to ASIs is an instance in which metaphors can easily do more harm than good.)
It's a fascinating point. Having said that, his ideas at that time seem radically underbaked. Nukes weren't that powerful or accurate then, so it's not clear how much damage the US could have done preemptively. The Soviet Union would then have had carte blanche to overrun Europe, and US soft power would have been disintegrated. If and when we get ASI, hopefully it comes up with better ideas than that.
"There were four years in which America could have preemptively nuked and destroyed the Soviet Union, but we didn’t. "
Actually nuclear bombs during that era were not that powerful - the main power that America and its allies have over the Soviets were their massive industrial capability (in WW2 it had higher GDP than all great powers in the Axis combined), as well as their logistical capability and their overwhelming airforce; compared with a nation that already lost 1/8 of its population and faced large-scale famine post-war.
However the Soviets at that time had probably the strongest army in Europe, not to mentioned the popularity of communist parties in Europe at that time (Czechoslovakia and France both had communist parties with the largest share of votes in parliaments), so attacking Soviet Union at that time could be successful, but you would see millions of body bags sent to grieving families in the US though.
You could check "Operation Unthinkable" to see what would happen if the Allies attacked the Soviets in 1945!
Noah Smith has mentioned in his podcast that a military alliance with superior manufacturing capability has "escalation dominance" over one that doesn't, and that during the Cold War the US would *eventually* win any non-nuclear war with the Soviet Union because it could keep building tanks, planes, etc. longer than the Soviet Union could.
This is basically not true. The Soviet Union had a big advantage over the United States in any war fought in Europe: it didn't have to send its military across the Atlantic Ocean to get to the battlefield. President Eisenhower concluded that maintaining a non-nuclear military force that could actually defend West Germany from a Soviet invasion would be so expensive (in terms of both money and physical resources) as to make it impractical, and instead chose to rely almost entirely on nuclear deterrence instead. Perhaps in the long run the United States could undergo a massive WWII style military buildup and have a second D-Day landing, but before that, West Germany and most of the rest of Western Europe would have fallen, just like it did when Hitler's army invaded.
Niall Ferguson, in his Civilization book, actually mentioned that if the Cold War went hot (not nuclear, of course), the Soviets would have an upper hand and pretty good chance to win the war; after all, they had larger armies, didn't need to ferry the army through Atlantic Ocean, had strong arms industries; but most importantly, the economic system and the people there were more resilient against wartime hardship.
You could even see it nowadays: Russia is still sanctions-resistant after 3 years of war against Ukraine, and in some types of weapons like artillery shells, they outproduce Europe!
A good take from The Atlantic about this, even though it was in 2024: https://www.theatlantic.com/ideas/archive/2024/02/russia-sanctions-economy-putin/677555/
Another argument in favor of export controls is that high-end GPUs have a relatively short lifetime if they are run 24x7 training Ai systems. I've read estimates of 1-3 years (e.g., https://massedcompute.com/faq-answers/?question=What%20is%20the%20typical%20lifespan%20of%20an%20NVIDIA%20data%20center%20GPU?). So the high end chips that China has will already soon reach the end of their lifespans
I just don’t view China in the same vein as you and Matt Ygelsias. China is not some major threat or the enemy to me. And while I’m not a huge communist fan, democracy has major issues too, as we can see, and I’m generally indifferent to how they want to be governed. I want the US to be competitive, but I truly don’t view Chinese companies (and perhaps even the CCP) as much worse than US corporations. They all solely care about their own interests.
And with the above said, how can I support tariffs that result in deeper pockets for a handful of US tech bros and higher prices and shittier products for the rest of us. I have zero loyalty to OpenAI, Zuck, Google, etc
This isn’t intended to be anti-capitalist at all, but it is to make the point that I do not align with “they must be stopped” and “we cannot let them win”. Deepseek wasn’t part of some massive CCP funding scheme that gave them an unfair advantage, which may warrant action from the US. No, by all accounts they had way less resources but were simply better. And before the arguments of “US must control AI”, this was open source.. China / DeepSeek literally gave a gift to the rest of us without any clear means of domination of the race or hiding their IP
You make some valid points. But China *is* bulking up militarily at an amazing rate. And a nation doesn't sink such massive resources into their military unless they're prepared to use it for war. And Taiwan will be the flash point. Or possibly their aggressive territorialism in the South China Sea. Which will produce an incident, that will snowball into war.
It's not like we're white knights or anything. As an empire we've done some really shitty things since 1945. But China wants their turn in the sun.
Yes. And engaging in pretty widespread hacking in the US.
I think the argument that's convincing to me is that China is the country in the next 20 years who is the most likely to disrupt global peace, and therefore we should take take that into account.
An invasion of Taiwan would be so incredibly disruptive globally that its hard to really predict all the potential fallout from it.
If I may engage in some "distillation" -- I think it's possible to believe that China is an evil behemoth intent on world domination AND want the US political system to exert some influence and/or control over the techbro free market feeding frenzy at the Trump smorgasbord and national defense raffle. You are reacting (rightly!) to the obscenity of these creeps securing their own futures at our expense, using the cover of "national security", while nobody is willing to act responsibly and in the actual national interest. Two generations ago we understood that nuclear technology posed both a mortal threat to humanity AND an unprecedented opportunity for humanity to contain traditional wars of attrition and to produce energy. But, importantly, we also understood that the power and the promise of the technology made it a natural subject for regulation and guidance under a politically cogent and responsible set of rules, restrictions and oversight. We did NOT say -- fuck it, let everybody play around and make money with nukes however they want, just as long as the NASDAQ goes up!
For me, the techbros simultaneously warning us that (1) AI is the most important tech in the history of humanity and the stakes of "winning" are nothing less than global domination, but (2) "we can not allow any puny, silly mortals limit us our self-directed goals of pursuing world domination for ourselves", is deeply discordant and profoundly at odds with all previous approaches to world-changing technologies with military and national security significance.
It's a new, dangerous path, and it's disheartening to see that Noah seems to be willing to verge down that path. We can be sure that the Chinese government will never make that mistake...
In one post (more related to Biden-Trump) Noah actually said that you should not expect this blog to be non-partisan, so your expectations should be that you are looking at Noah's points of view, with his own biases!
I think I said this at Slow Boring as well, but you seem to have confused the tariffs on automobiles and manufactured goods (which produce deeper pockets for auto executives and auto workers in the US, Europe, Japan, and Korea) with the export controls on semiconductors (which don’t obviously product deeper pockets for anyone). These are both controls on international trade, but they are in opposite directions.
Well said. Thanks.
Proposals from the Trump admin relate to marrying export controls with new tariff actions. This supposedly will bring manufacturing back to the United States, but this seems to make the development of AI models costlier, which could lead to the U.S. becoming less competitive in the AI race. Can the U.S. sustain an innovation edge without over-regulating itself?
Noah, I have subscribed in order to make one comment. You mention many true facts about the situation surrounding AI. But the bigger picture is that we are racing towards the creation of nonbiological intelligences smarter than any human, and the logical consequence of that, is that humans will no longer control their own destiny. We are creating beings that will replace us completely, and yet the people making this happen can barely even acknowledge this. You're just a commentator, but this applies to you too. The logical outcome of the AI race is not that China wins or that America wins, it is that AI wins and does as it pleases with the human race. I can't take any commentator very seriously, whose big picture doesn't include that aspect.
Great article, only thing I would point out, let's say it cost X billion for OpenAI to make the model with Y ability that Deepthink distilled from; that method (as it exists today) does not get us to Y+1 unless someone puts that money in to raise the high water mark. OpenAI used distilling to release it's own slight worse than their premiere one but cheaper. I'm a complete skeptic of LLM version of "AI"; until we can increase true capability for less money (rather than drafting off a more capable expensive model) the LLM future is looking pretty grim. Seeing the insane amount of money OpenAI et al. are investing in hardware they are still going scale scale scale, to push the envelop to Y+1, but it's not efficient in any form given the billions being put in.
I think a new AI company called 2nd best but 1st cheapest should do the DeepThink ditstillation approach to all the capable models to get whatever % the ability at orders of magnitude cheaper cost to build, and sell the API access for a small fraction less of what OpenAI is selling it for (which is still losing money on). That seems to be where the profit lies.
Innovation is an economic imperative. Given the amount of energy that will be required for the continued annual rollout of 40 new data centers, a newly designed ASICS transformer chip with 20x more processing power and the attendant energy savings will be a Godsend for AI on U.S. soil. And a silicon anode battery with 30% more energy density, half the charge time of conventional commercial Li-ion batteries, longer battery life, and with Brakeflow technology to mitigate thermal runways, will put the U.S. in a competitive position in the Battery War with China. Note the recent fire that destroyed a commercial aircraft in South Korea. In 2023, there were 163 occurrences of Li-ion battery fires on commercial aircraft, to say nothing of 142 Li-ion battery fires for the NYFD, involving loss of lives and property.
“The compute gap between US and China—further widened by export controls —remains DeepSeek’s primary constraint. DeepSeek’s leadership openly acknowledged a 4x compute disadvantage despite their efficiency gains.”
Indeed. A U.S. company has designed a purpose-built ASICS transformer chip with 20x the processing power of NVDA’s GPU chips. Not only does this significantly reduce the latency, it also represents an exponential decrease in the energy requirements. This chip is likely to begin production in late 2025/early 2026.
Of course, NVDA can design/manufacture a competing chip, but this is about a three-year process. All the money in the world can’t buy time, and three years is a significant first-mover advantage in a Silicon world. The real competitive threat to NVDA is a much better chip. The good news is that a better chip design is happening in the U.S. sure, it’s a startup company, but a silicon chip veteran with three decades of experience at a major Silicon Valley semiconductor company, who came out of retirement, is leading the design team. As with process-manufacturing engineers, there are master craftsmen teaching the chip designers of today and tomorrow — in Taiwan and Silicon Valley. I think hardware is the key to the AI and Battery Wars.
Do export controls make it more likely China invades Taiwan? You won’t sell me the chips? Fine. I’ll just “nationalize” the foundries.
I've heard all the foundries have kill switches. I don't know the details, of course, but I assume they're designed to make them robustly unusable in order to deter that line of thinking.
That makes sense. Of course if you have already purloined the know how. And you have captured the people…. You can probably rebuild pretty quickly.
"much of it depends on tweaks, workarounds, and kludges that exist only in the collective memory of a huge number of human employees"
Well, and in the code itself, if you manage to steal that, which I don't doubt the Chinese are trying to do.
I am so confused by people claiming the DeepSeek launch is what had the big hit on NVIDIA stock. DeepSeek announced R1 a full week before that bad market day! Do people think the market analysts for the tech or chip sector were all on vacation for a week?
Yeah, maybe that wasn't it. https://www.understandingai.org/p/i-dont-believe-deepseek-crashed-nvidias