It really helps to think about the world as if people were rational working under full information.
It really makes sense for China to invade Taiwan as soon as this summer.
1- China won’t be able to win the long timelines world many policymakers are planning for. The David Sacks/Jensen Hwang argument is that by precluding China from the best AI accelerators, they’ll eventually have their own domestic supply the U.S. can’t control. When that happens, the U.S. will have a robot army they can cheaply deploy to defend the Taiwan strait.
2- Taiwan is really central for U.S. AI development. Really. Not only TSMC, but Micron, Mediatek, Alchip, all the ODMs like Foxconn.
3- China needs to really consider the possibility of recursive self-improvement. Waiting one year means another $500B in AI equipament for AI Labs.
4- The governance of the CCP over China is not guaranteed under a strong AI world.
5- Trumpism might be gone at the end of the decade and president Gavin Newsom might be much more willing to fight over Taiwan.
Seems unlikely to me this would be in their interest.
Regarding chips, US just started selling them H200s, and if they invade Taiwan they probably destroy TSMC and immediately lose their main source of compute. The majority of the semiconductor industry outside Taiwan would still not be friendly to China.
They also lose 50% of their trade even if the US continues trading with them, which tanks their GDP by something like 10% [1]. If the US sanctions them too, it would be 60-70% trade reduction.
Militarily, they are also building up faster than the US, and if they invade Japan/SK will militarize and might get nuclear weapons.
If they really believed in RSI they would do diplomacy to get more compute and just invest in their AI industry.
[1] Claude thinks the likely outcome is this:
Lost trade: $2.4-2.8 trillion (US, EU, Japan, UK, Canada, Australia, possibly South Korea/Taiwan)
Continuing trade: $2.0-2.4 trillion (Russia, most ASEAN, Brazil, India, Middle East, Africa)
Net trade: $2.0-2.4T vs previous $6.2T = 60-70% reduction
Trade and invest in the U.S. is giving resources to the U.S. now in exchange for pieces of paper that might or might not be valuable in 10 years in a radical y different world.
Money can be exchanged for goods and services in under a 10-year timeframe, including AI hardware and talent. To come out ahead from trade they just need to invest more of their surplus in AI than the US does.
The #1 issue for the Chinese public is the economy, just like in the US. This includes people with influence on CCP decision-making.
1- China won’t be able to win the long timelines world many policymakers are planning for. The David Sacks/Jensen Hwang argument is that by precluding China from the best AI accelerators, they’ll eventually have their own domestic supply the U.S. can’t control. When that happens, the U.S. will have a robot army they can cheaply deploy to defend the Taiwan strait.
This only tracks under the assumption that militarily useful AI’s development will be dependent upon compute, that it will arrive before China catches up in terms of semiconductor production, and that, once it comes to exist, it will be difficult to cheaply replicate. I think a majority of non-Western world leaders don’t expect any of those three things to be true. Whether or not you agree with them, that’s a useful point of information for modeling their behavior.
As best I can tell, China sees the massive investment in training frontier LLMs to essentially be a waste of money. They’ll replicate the models from six months to a year ago just in case, since it’s cheap enough to do so and having them around for prestige has some value, but they don’t expect them to be the wonder-weapon that some in the U.S. expect them to be. Likewise, Russia has far fewer people on hand, and seems to prefer allocating their best researchers towards more conventional work (e.g. hypersonic missiles).
Essentially, the East is gambling that AGI happens later or more unconventionally, and, in that scenario, the West is just helpfully providing them with free as-good-as-open-source R&D.
5- Trumpism might be gone at the end of the decade and president Gavin Newsom might be much more willing to fight over Taiwan.
This seems incredibly implausible on multiple fronts.
“Trumpism” is just the rise in nationalism in the West that’s been going on since 2012 at the latest and shows no signs of slowing down. Motivated people have been predicting that “Trumpism will disappear by next year” since Trump first walked down the elevator.
Right-wing politicians are almost always more China-hawkish than left-wing ones. In any case, in a major war, they enjoy a much greater degree of faith from the pipe-hitters, which imposes a practical constraint on hawkishness.
In my opinion, it doesn’t make rational sense for them to invade at all. Even in the best-case scenario for China, where they manage to pacify Taiwan after a tough fight, I would still expect the following:
1) They would be permanently shut out of all Western trade and technology sharing. 2) All critical semiconductor manufacturing in Taiwan will be destroyed by the US or the local Taiwanese military before China can get to it, and most of it is already in the process of being successfully transferred to the US. I also expect that most of the human talent would be taken to the US. 3) Even if the US did not directly intervene, the US and their allies would start massive rearmament and reindustrialisation programmes and maximally utilise their advantage in AI and other critical technologies in future. 4) Regarding point 4, if American AI victory is inevitable due to their computing advantage, China might still get a better deal in the current scenario, where it is perceived as merely an economic competitor and geopolitical challenger, rather than a direct adversary, as it would be in the event of an invasion of Taiwan.
There are also some indications that Taiwanese politics are slowly moving in a pro-China direction, with increased support for peaceful re-unification among younger KMT voters, which might also incentivise China to bide its time and avoid doing anything reckless.
They would be permanently shut out of all Western trade and technology sharing.
That’s not true. Many countries um the West were literally fascists during WWII. I can totally imagine worlds where China and the West get along after that.
All critical semiconductor manufacturing in Taiwan will be destroyed by the US or the local Taiwanese military before China can get to it, and most of it is already in the process of being successfully transferred to the US. I also expect that most of the human talent would be taken to the US.
Why should China care, if they are mostly cut from the output of that anyway?
3) Even if the US did not directly intervene, the US and their allies would start massive rearmament and reindustrialisation programmes and maximally utilise their advantage in AI and other critical technologies in future.
Trump just asked for a $1.5T military budget. It’s already happening.
4) Regarding point 4, if American AI victory is inevitable due to their computing advantage, China might still get a better deal in the current scenario, where it is perceived as merely an economic competitor and geopolitical challenger, rather than a direct adversary, as it would be in the event of an invasion of Taiwan.
American AI victory is not inevitable. And not taking this bet would relinquish China to the permanent undetclass of nations, which i’m sure Beijing doesn’t want.
“Culmination” suggests a subsequent decline. In 2025, scaling of RLVR delivered a lot of capabilities, and late 2025 was the first time since 2023 that 10x-20x scaling in pretraining compute (compared to original Mar 2023 GPT-4) has finally made a full appearance (probably in Gemini 3 Pro and Opus 4.5). There is 100x-400x more in scaling of compute by 2029-2031 (compared to the current models), and much more in the low-hanging fruit of doing things well rather than prototyping the first thing that sort of works. The only low-hanging fruit that likely mostly ran out in 2025 is raw scaling of RLVR (in proportion to pretraining), and even that probably still has a bit to go. Setting up better tasks and RL environments will plausibly be more impactful than some feasible amount of further scaling of RLVR relative to pretraining. Then there’s continual learning that might be quite impactful in 2026-2027.
I expect some sort of culmination in 2027-2032 (assuming no AGI), when scaling of compute slows down and there have been at least 1-2 years to learn to make better use of it. The first stage of compute slowdown probably follows 2026, if 2028 won’t see 5 GW training systems (which would be on trend for training system compute growth in 2022-2026, but currently doesn’t seem to be happening). The second probably follows 2028-2030, when funding to secure ever more training compute mostly stops growing, and so compute mostly falls back to growth according to its price-performance.
Kinda, but there won’t be enough natural text data at the higher end of this range (using 2028-2030 compute) to just scale pretraining on text (you’d need more than 1,000 trillion tokens with repetition, maybe 200-300 trillion unique tokens), something else would need to happen instead or you start losing efficiency and compute ends up being less useful than it would be if there was enough text.
The steps of scaling take a long time, so only late 2025 models get to be shaped compute optimally for 2024 levels of pretraining compute, and run on hardware announced and first available in the cloud in 2024. This is just 2 years from 2022, when GPT-4 was trained, and the first of two 10x-20x steps at the 2022-2026 pace of scaling, with a third step remaining somewhere beyond 2026 if we assume $100bn per year revenue for an AI company (at that time). With 2026 compute, there just might be enough text data (with repetition) to say that scaling of pretraining is still happening in a straightforward sense, which brings the change from original Mar 2023 GPT-4 to 100x-400x (for models that might come out in 2027).
But this 100x-400x is also a confusing point of comparison, since between 2023 and 2027 there was the introduction of RLVR scaling (and test time reasoning), and also all the improvements that come from working on a product (as opposed to a research prototype) for 4 years. Continual learning might be another change complicating this comparison that happens before 2027 (meaning it might be a significant change, which remains uncertain; that it’s coming in some form, at least as effective context extension, seems quite clear at this point).
Thank you! This makes me wonder if one can predict whether CoT-based AGI will be reached at all. Setting aside any forecaster’s nightmares like a time horizon growing exponentially until the very last couple of doublings or a potentiallyinflated time horizon of Claude Opus 4.5, one might try to predict the influence of the 100x-400x increase in compute on the METR-like[1] time horizon. And how much do you expect “setting up better tasks and RL environments” to increase the logarithm of the time horizon?
I also had this very impression looking at the METR graph since the post-o3 growth returned to the old trend. Alas, there is Claude Opus 4.5 with its 4hr49 min time horizon, which is on the pre-o3 faster trend (see, however, twocomments pointing out that the METR benchmark is no longer as trustworthy as it once was and my potential explanation of Claude’s abnormally high 50%/80% time horizon ratio). I just can’t wait for METR to evaluate Gemini 3 Pro and/or GPT-5.2 (and GPT-5.2 Codex Max when it is released?) and see if the new crop of models has a high 50% time horizon without Claude’s issues...
ASI is often assumed as the all knowing first-principles thinker.
Sam Altman once said: “We don’t know how we’ll make money. Maybe we’ll create the AGI and then ask it to make us one billion dollars”. Despite that, my read of the Bitter Lesson is that empiricism, not cleverness, is what have been pushing the AI endeavor forward. I have seen my fair share of researchers during the 2010s creating ontologies and programming in Lisp, only to get obliterated by deep learning and brute force.
Machine learning has obliterated humans in fields such as chess, computer vision, and is on track to do so in math and programming. These are themes where if you are given 1,000,000x more compute, you can do 1,000,000x more experiments.
I don’t see how this translates to the rest of the human endeavor.
Medical research will be extremely bound by 1- the amount of humanoid robots doing experiments in the lab and 2- the amount of human trials you can do. You can’t expect to ask the ASI “cure cancer from first principles” and only because the ASI is a super human coder, it’ll have the emerging capability of curing cancer.
If my framing is right, parts of the economy that will benefit from increasing compute budgets will be radically transformed. Parts of the economy where the “real time empiricism” is important will “only get better”.
It really helps to think about the world as if people were rational working under full information.
It really makes sense for China to invade Taiwan as soon as this summer.
1- China won’t be able to win the long timelines world many policymakers are planning for. The David Sacks/Jensen Hwang argument is that by precluding China from the best AI accelerators, they’ll eventually have their own domestic supply the U.S. can’t control. When that happens, the U.S. will have a robot army they can cheaply deploy to defend the Taiwan strait.
2- Taiwan is really central for U.S. AI development. Really. Not only TSMC, but Micron, Mediatek, Alchip, all the ODMs like Foxconn.
3- China needs to really consider the possibility of recursive self-improvement. Waiting one year means another $500B in AI equipament for AI Labs.
4- The governance of the CCP over China is not guaranteed under a strong AI world.
5- Trumpism might be gone at the end of the decade and president Gavin Newsom might be much more willing to fight over Taiwan.
Seems unlikely to me this would be in their interest.
Regarding chips, US just started selling them H200s, and if they invade Taiwan they probably destroy TSMC and immediately lose their main source of compute. The majority of the semiconductor industry outside Taiwan would still not be friendly to China.
They also lose 50% of their trade even if the US continues trading with them, which tanks their GDP by something like 10% [1]. If the US sanctions them too, it would be 60-70% trade reduction.
Militarily, they are also building up faster than the US, and if they invade Japan/SK will militarize and might get nuclear weapons.
If they really believed in RSI they would do diplomacy to get more compute and just invest in their AI industry.
[1] Claude thinks the likely outcome is this:
Lost trade: $2.4-2.8 trillion (US, EU, Japan, UK, Canada, Australia, possibly South Korea/Taiwan)
Continuing trade: $2.0-2.4 trillion (Russia, most ASEAN, Brazil, India, Middle East, Africa)
Net trade: $2.0-2.4T vs previous $6.2T = 60-70% reduction
Trade and invest in the U.S. is giving resources to the U.S. now in exchange for pieces of paper that might or might not be valuable in 10 years in a radical y different world.
Money can be exchanged for goods and services in under a 10-year timeframe, including AI hardware and talent. To come out ahead from trade they just need to invest more of their surplus in AI than the US does.
The #1 issue for the Chinese public is the economy, just like in the US. This includes people with influence on CCP decision-making.
This only tracks under the assumption that militarily useful AI’s development will be dependent upon compute, that it will arrive before China catches up in terms of semiconductor production, and that, once it comes to exist, it will be difficult to cheaply replicate. I think a majority of non-Western world leaders don’t expect any of those three things to be true. Whether or not you agree with them, that’s a useful point of information for modeling their behavior.
As best I can tell, China sees the massive investment in training frontier LLMs to essentially be a waste of money. They’ll replicate the models from six months to a year ago just in case, since it’s cheap enough to do so and having them around for prestige has some value, but they don’t expect them to be the wonder-weapon that some in the U.S. expect them to be. Likewise, Russia has far fewer people on hand, and seems to prefer allocating their best researchers towards more conventional work (e.g. hypersonic missiles).
Essentially, the East is gambling that AGI happens later or more unconventionally, and, in that scenario, the West is just helpfully providing them with free as-good-as-open-source R&D.
This seems incredibly implausible on multiple fronts.
“Trumpism” is just the rise in nationalism in the West that’s been going on since 2012 at the latest and shows no signs of slowing down. Motivated people have been predicting that “Trumpism will disappear by next year” since Trump first walked down the elevator.
Right-wing politicians are almost always more China-hawkish than left-wing ones. In any case, in a major war, they enjoy a much greater degree of faith from the pipe-hitters, which imposes a practical constraint on hawkishness.
In my opinion, it doesn’t make rational sense for them to invade at all. Even in the best-case scenario for China, where they manage to pacify Taiwan after a tough fight, I would still expect the following:
1) They would be permanently shut out of all Western trade and technology sharing.
2) All critical semiconductor manufacturing in Taiwan will be destroyed by the US or the local Taiwanese military before China can get to it, and most of it is already in the process of being successfully transferred to the US. I also expect that most of the human talent would be taken to the US.
3) Even if the US did not directly intervene, the US and their allies would start massive rearmament and reindustrialisation programmes and maximally utilise their advantage in AI and other critical technologies in future.
4) Regarding point 4, if American AI victory is inevitable due to their computing advantage, China might still get a better deal in the current scenario, where it is perceived as merely an economic competitor and geopolitical challenger, rather than a direct adversary, as it would be in the event of an invasion of Taiwan.
There are also some indications that Taiwanese politics are slowly moving in a pro-China direction, with increased support for peaceful re-unification among younger KMT voters, which might also incentivise China to bide its time and avoid doing anything reckless.
Thank you for replying.
That’s not true. Many countries um the West were literally fascists during WWII. I can totally imagine worlds where China and the West get along after that.
Why should China care, if they are mostly cut from the output of that anyway?
Trump just asked for a $1.5T military budget. It’s already happening.
American AI victory is not inevitable. And not taking this bet would relinquish China to the permanent undetclass of nations, which i’m sure Beijing doesn’t want.
It might be the case that 2025 was the culmination of many low-hanging fruits driving AI development than an actual take-off.
“Culmination” suggests a subsequent decline. In 2025, scaling of RLVR delivered a lot of capabilities, and late 2025 was the first time since 2023 that 10x-20x scaling in pretraining compute (compared to original Mar 2023 GPT-4) has finally made a full appearance (probably in Gemini 3 Pro and Opus 4.5). There is 100x-400x more in scaling of compute by 2029-2031 (compared to the current models), and much more in the low-hanging fruit of doing things well rather than prototyping the first thing that sort of works. The only low-hanging fruit that likely mostly ran out in 2025 is raw scaling of RLVR (in proportion to pretraining), and even that probably still has a bit to go. Setting up better tasks and RL environments will plausibly be more impactful than some feasible amount of further scaling of RLVR relative to pretraining. Then there’s continual learning that might be quite impactful in 2026-2027.
I expect some sort of culmination in 2027-2032 (assuming no AGI), when scaling of compute slows down and there have been at least 1-2 years to learn to make better use of it. The first stage of compute slowdown probably follows 2026, if 2028 won’t see 5 GW training systems (which would be on trend for training system compute growth in 2022-2026, but currently doesn’t seem to be happening). The second probably follows 2028-2030, when funding to secure ever more training compute mostly stops growing, and so compute mostly falls back to growth according to its price-performance.
100x more compute means the leap from GPT-3 to GPT-4.
Kinda, but there won’t be enough natural text data at the higher end of this range (using 2028-2030 compute) to just scale pretraining on text (you’d need more than 1,000 trillion tokens with repetition, maybe 200-300 trillion unique tokens), something else would need to happen instead or you start losing efficiency and compute ends up being less useful than it would be if there was enough text.
The steps of scaling take a long time, so only late 2025 models get to be shaped compute optimally for 2024 levels of pretraining compute, and run on hardware announced and first available in the cloud in 2024. This is just 2 years from 2022, when GPT-4 was trained, and the first of two 10x-20x steps at the 2022-2026 pace of scaling, with a third step remaining somewhere beyond 2026 if we assume $100bn per year revenue for an AI company (at that time). With 2026 compute, there just might be enough text data (with repetition) to say that scaling of pretraining is still happening in a straightforward sense, which brings the change from original Mar 2023 GPT-4 to 100x-400x (for models that might come out in 2027).
But this 100x-400x is also a confusing point of comparison, since between 2023 and 2027 there was the introduction of RLVR scaling (and test time reasoning), and also all the improvements that come from working on a product (as opposed to a research prototype) for 4 years. Continual learning might be another change complicating this comparison that happens before 2027 (meaning it might be a significant change, which remains uncertain; that it’s coming in some form, at least as effective context extension, seems quite clear at this point).
Thank you! This makes me wonder if one can predict whether CoT-based AGI will be reached at all. Setting aside any forecaster’s nightmares like a time horizon growing exponentially until the very last couple of doublings or a potentially inflated time horizon of Claude Opus 4.5, one might try to predict the influence of the 100x-400x increase in compute on the METR-like[1] time horizon. And how much do you expect “setting up better tasks and RL environments” to increase the logarithm of the time horizon?
As for compute mostly falling back to growth according to its price-performance, the AI-2027 compute forecast doesn’t mention anything better than 3nm chips, and TSMC’s 2nm chips would be significantly more expensive than those of previous generations.
The METR benchmark itself has yet to include tasks requiring more than 16 hrs of work.
I also had this very impression looking at the METR graph since the post-o3 growth returned to the old trend. Alas, there is Claude Opus 4.5 with its 4hr49 min time horizon, which is on the pre-o3 faster trend (see, however, two comments pointing out that the METR benchmark is no longer as trustworthy as it once was and my potential explanation of Claude’s abnormally high 50%/80% time horizon ratio). I just can’t wait for METR to evaluate Gemini 3 Pro and/or GPT-5.2 (and GPT-5.2 Codex Max when it is released?) and see if the new crop of models has a high 50% time horizon without Claude’s issues...
See my comment trying to pushback on Daniel and Eli. I feel we both are on similar conclusions.
ASI is often assumed as the all knowing first-principles thinker.
Sam Altman once said: “We don’t know how we’ll make money. Maybe we’ll create the AGI and then ask it to make us one billion dollars”. Despite that, my read of the Bitter Lesson is that empiricism, not cleverness, is what have been pushing the AI endeavor forward. I have seen my fair share of researchers during the 2010s creating ontologies and programming in Lisp, only to get obliterated by deep learning and brute force.
Machine learning has obliterated humans in fields such as chess, computer vision, and is on track to do so in math and programming. These are themes where if you are given 1,000,000x more compute, you can do 1,000,000x more experiments.
I don’t see how this translates to the rest of the human endeavor.
Medical research will be extremely bound by 1- the amount of humanoid robots doing experiments in the lab and 2- the amount of human trials you can do. You can’t expect to ask the ASI “cure cancer from first principles” and only because the ASI is a super human coder, it’ll have the emerging capability of curing cancer.
If my framing is right, parts of the economy that will benefit from increasing compute budgets will be radically transformed. Parts of the economy where the “real time empiricism” is important will “only get better”.