On the AGI-thing: there recent have been suggestions to extend the 2-way split of looking at AI into 'Narrow' and 'General' (such as Google DeepMind have promoted) into a 3-way split: 'Narrow', 'General', and in between: 'Broad&Shallow' (Marcus), GenAI is in that category and has no real route to 'General' (e.g. not via scaling). See https://garymarcus.substack.com/p/agi-versus-broad-shallow-intelligence
Or https://ea.rna.nl/2025/01/08/lets-call-gpt-and-friends-wide-ai-and-not-agi/ (where 'Broad&Shallow' is called 'Wide'). Marcus links to that story as well. Broad&Shallow is more precise, as the 'shallowness' is the essential issue that makes evolving into 'General' hard if not impossible. The 'Wide' article contains some more background, e.g. GPT-o3 ARC-AGI results and the importance of 'imagination' when looking at intelligence).
Regarding the critical role BIS plays in tech export controls, as well as the resource challenges they face, I recommend watching this excellent conversation between Greg Allen of CSIS (a DoD Joint AI Center plank holder) and Alan Estevez, Under Secretary of Commerce for Industry and Security.
It is always helpful to hear directly from the source, so to speak, on the motivations behind and the challenges of implementing technology export controls. It's a thoughtful, even introspective discussion. With Alan acknowledging at the end that he has no idea how the next administration will approach these issues.
I can't really comment on the political/organizational bets but the two scientific/technological ones have already been lost:
1) the scaling hypothesis has always looked dubious and in the last months even "AI" hypsters have come around to admitting that this won't work.
2) generative models won't get us to AGI. In fact, we have no idea whether it's attainable at all, but generative models (w/o the capacity for reasoning, and trained on and for specialized tasks) can't.
So what's the most plausible outcome if there's all this investment into something that doesn't bring the envisioned technological payoff? Especially since on-shoring data centers also means on-shoring the immense electricity production and water consumption that come with it, and that already put a strain on US capacities?
The other big question in US trade policy is how third countries will be affected and how they will respond. At least rhetorically, the Biden Administration policy was about "friendshoring". But that won't be the case under Trump. Not only will tariffs be applied to (putative) friend and foe alike, but Musk is doing his best to alienate every country in the world. It's certain that other countries won't be highly co-operative with any Trump effort to control AI, possible that there will be an active effort to undermine it, for example, by breaking with strong IP (Cory Doctorow has just written about this in relation to Canada)
I don't have anything to say about the substance of your post because I agree with it. However, I would like to mention a charming example of a learned Roman slipping into Greek that I ran across in Robert Kaster's book Emotion, Restraint, and Community in Ancient Rome.
It comes from a letter from Cicero to Atticus upon learning that Marcus Calidius lost his bid for a consulship; Cicero writes "... frankly do rejoice, since feeling nemesis is different from feeling phthonos."
Kaster explains that the difference is between feeling pain at another's undeserved success (nemesis) versus feeling pain because the other is your peer (phthonos). But he goes on to say that "Cicero's use of the Greek terms here is not just mannered but a means of achieving clarity as well. For if we ask how Cicero would have expressed the same idea using the most commonly deployed Latin ... he would have said 'plane gaudeo, quoniam invidia ab invidia interest' - 'frankly, I do rejoice, since invidia is one thing, invidia quite another'".
> A couple of weeks ago, a Chinese company, DeepSeek announced results that suggest that it has trained a frontier AI model without access to the most advanced semiconductors. DeepSeek’s model seems to achieve equivalent results to powerful US models with far fewer weights. If this works as it appears to, it may be that the semiconductor chokehold is less decisive than the U.S. hopes.
This seems wrong, DeepSeek used NVIDIA H800s which are essentially as good as H100s, pretty much near SOTA. These chips could be imported legally in 2023 because BIS initially set the thresholds wrong, but that was rectified one year ago so these types of (US-designed) chips are now banned, and can only be smuggled into China.
It's true that DeepSeek used relatively few GPUs to get relatively impressive results, but that's more a signal that they're good at compute efficiency, not that they're not compute-constrained. Like, they'd still benefit from having access to lots more GPUs! And the export controls are still significantly limiting their ability to get more GPUs.
> Initially, many national security people saw export controls as financial sanctions but for physical products.
Citation needed? The US has been doing export controls for a century. There's a very long history of the US using export controls as a tool against China in particular, sometimes successfully and other times less so.
> Doing export controls well is hard.
Agreed, but so far the controls have had an effect. Cf. DeepSeek's founder saying last year that its main constraint is the ban on GPU imports.
> If this plan gets implemented, the U.S. is going to have to do it at much broader scale than before, for much more ambitious objectives with the grudging cooperation of surly and truculent chip companies (at best), who have made it clear how much they hate the rules, and complications in getting the ‘end users’ who buy semiconductors to comply with the requirements that the U.S. wants to impose.
In many ways, the rule makes it easier for the US to enforce the chip controls. Where before smugglers could get some substantial amount of GPUs into China via third countries like Singapore and Malaysia, that's now harder to do at scale due to country caps on GPU imports.
> It is certainly not inconceivable that the BIS could beef up its capacities, as Treasury’s Office of Foreign Assets Control has beefed up, building up the internal structures to analyze, to think strategically, and to revise strategy based on results. But it is a real challenge - and will be much harder to pull off in a new administration that is apparently quite hostile to technical expertise and the “deep state.”
I agree that this is a challenge, but (1) it does seem feasible to improve BIS's enforcement, and (2) even with its current capabilities, the export controls so far have definitely meaningfully limited the amount of compute owned by Chinese actors, and the new rules will likely make that effect larger. Like, AI chip companies/distributors like NVIDIA, Supermicro, Ingram Micro etc. do do some amount of due diligence, and they're not going to just ignore BIS's rules. The hardest bit before was probably for BIS to get AI chip resellers in third countries to comply, or to get those AI chip companies/distributors to monitor resales in third countries, but that's probably less of an issue now with country caps on GPU imports.
Putting the squeeze on the needed chips will just push China to find workarounds. They can brute force their AI development using lower technology chips, and/or make their resource demands smaller (as some AI companies are doing).
The assumption for doing this "now, now, now" is based on the AI scaling belief. We already know that human-based training resources are effectively exhausted, removing one piece of the scaling ladder. Would ever-larger models improve capabilities? Not exponentially it appears so far. Therefore if the hoped-for path to AGI and superintelligence slows down and becomes incremental, then the immediacy of staying ahead starts to evaporate, giving competitors more time to catch up, probably with new approaches. The US in trying to win the race may just close itself off from actually winning.
Get it that the conquest of Gaul is a well known book but as a policy we should translate foreign language quotes, especially those in dead languages. It’s a little obnoxious not to. My two cents
Interesting.
On the AGI-thing: there recent have been suggestions to extend the 2-way split of looking at AI into 'Narrow' and 'General' (such as Google DeepMind have promoted) into a 3-way split: 'Narrow', 'General', and in between: 'Broad&Shallow' (Marcus), GenAI is in that category and has no real route to 'General' (e.g. not via scaling). See https://garymarcus.substack.com/p/agi-versus-broad-shallow-intelligence
Or https://ea.rna.nl/2025/01/08/lets-call-gpt-and-friends-wide-ai-and-not-agi/ (where 'Broad&Shallow' is called 'Wide'). Marcus links to that story as well. Broad&Shallow is more precise, as the 'shallowness' is the essential issue that makes evolving into 'General' hard if not impossible. The 'Wide' article contains some more background, e.g. GPT-o3 ARC-AGI results and the importance of 'imagination' when looking at intelligence).
Regarding the critical role BIS plays in tech export controls, as well as the resource challenges they face, I recommend watching this excellent conversation between Greg Allen of CSIS (a DoD Joint AI Center plank holder) and Alan Estevez, Under Secretary of Commerce for Industry and Security.
It is always helpful to hear directly from the source, so to speak, on the motivations behind and the challenges of implementing technology export controls. It's a thoughtful, even introspective discussion. With Alan acknowledging at the end that he has no idea how the next administration will approach these issues.
https://www.csis.org/events/reflecting-commerce-departments-role-protecting-critical-technology-under-secretary-commerce
I can't really comment on the political/organizational bets but the two scientific/technological ones have already been lost:
1) the scaling hypothesis has always looked dubious and in the last months even "AI" hypsters have come around to admitting that this won't work.
2) generative models won't get us to AGI. In fact, we have no idea whether it's attainable at all, but generative models (w/o the capacity for reasoning, and trained on and for specialized tasks) can't.
So what's the most plausible outcome if there's all this investment into something that doesn't bring the envisioned technological payoff? Especially since on-shoring data centers also means on-shoring the immense electricity production and water consumption that come with it, and that already put a strain on US capacities?
The other big question in US trade policy is how third countries will be affected and how they will respond. At least rhetorically, the Biden Administration policy was about "friendshoring". But that won't be the case under Trump. Not only will tariffs be applied to (putative) friend and foe alike, but Musk is doing his best to alienate every country in the world. It's certain that other countries won't be highly co-operative with any Trump effort to control AI, possible that there will be an active effort to undermine it, for example, by breaking with strong IP (Cory Doctorow has just written about this in relation to Canada)
And let's not forget that the current NVIDIA GPUs are made by TSMC in Taiwan, which may induce China to grab that island as quickly as possible.
I don't have anything to say about the substance of your post because I agree with it. However, I would like to mention a charming example of a learned Roman slipping into Greek that I ran across in Robert Kaster's book Emotion, Restraint, and Community in Ancient Rome.
It comes from a letter from Cicero to Atticus upon learning that Marcus Calidius lost his bid for a consulship; Cicero writes "... frankly do rejoice, since feeling nemesis is different from feeling phthonos."
Kaster explains that the difference is between feeling pain at another's undeserved success (nemesis) versus feeling pain because the other is your peer (phthonos). But he goes on to say that "Cicero's use of the Greek terms here is not just mannered but a means of achieving clarity as well. For if we ask how Cicero would have expressed the same idea using the most commonly deployed Latin ... he would have said 'plane gaudeo, quoniam invidia ab invidia interest' - 'frankly, I do rejoice, since invidia is one thing, invidia quite another'".
Scaling hypothesis is already off. According to Ilya Sutskever, scaling has plateaued aroung the gpt-4 time.
> A couple of weeks ago, a Chinese company, DeepSeek announced results that suggest that it has trained a frontier AI model without access to the most advanced semiconductors. DeepSeek’s model seems to achieve equivalent results to powerful US models with far fewer weights. If this works as it appears to, it may be that the semiconductor chokehold is less decisive than the U.S. hopes.
This seems wrong, DeepSeek used NVIDIA H800s which are essentially as good as H100s, pretty much near SOTA. These chips could be imported legally in 2023 because BIS initially set the thresholds wrong, but that was rectified one year ago so these types of (US-designed) chips are now banned, and can only be smuggled into China.
It's true that DeepSeek used relatively few GPUs to get relatively impressive results, but that's more a signal that they're good at compute efficiency, not that they're not compute-constrained. Like, they'd still benefit from having access to lots more GPUs! And the export controls are still significantly limiting their ability to get more GPUs.
> Initially, many national security people saw export controls as financial sanctions but for physical products.
Citation needed? The US has been doing export controls for a century. There's a very long history of the US using export controls as a tool against China in particular, sometimes successfully and other times less so.
> Doing export controls well is hard.
Agreed, but so far the controls have had an effect. Cf. DeepSeek's founder saying last year that its main constraint is the ban on GPU imports.
> If this plan gets implemented, the U.S. is going to have to do it at much broader scale than before, for much more ambitious objectives with the grudging cooperation of surly and truculent chip companies (at best), who have made it clear how much they hate the rules, and complications in getting the ‘end users’ who buy semiconductors to comply with the requirements that the U.S. wants to impose.
In many ways, the rule makes it easier for the US to enforce the chip controls. Where before smugglers could get some substantial amount of GPUs into China via third countries like Singapore and Malaysia, that's now harder to do at scale due to country caps on GPU imports.
> It is certainly not inconceivable that the BIS could beef up its capacities, as Treasury’s Office of Foreign Assets Control has beefed up, building up the internal structures to analyze, to think strategically, and to revise strategy based on results. But it is a real challenge - and will be much harder to pull off in a new administration that is apparently quite hostile to technical expertise and the “deep state.”
I agree that this is a challenge, but (1) it does seem feasible to improve BIS's enforcement, and (2) even with its current capabilities, the export controls so far have definitely meaningfully limited the amount of compute owned by Chinese actors, and the new rules will likely make that effect larger. Like, AI chip companies/distributors like NVIDIA, Supermicro, Ingram Micro etc. do do some amount of due diligence, and they're not going to just ignore BIS's rules. The hardest bit before was probably for BIS to get AI chip resellers in third countries to comply, or to get those AI chip companies/distributors to monitor resales in third countries, but that's probably less of an issue now with country caps on GPU imports.
Putting the squeeze on the needed chips will just push China to find workarounds. They can brute force their AI development using lower technology chips, and/or make their resource demands smaller (as some AI companies are doing).
The assumption for doing this "now, now, now" is based on the AI scaling belief. We already know that human-based training resources are effectively exhausted, removing one piece of the scaling ladder. Would ever-larger models improve capabilities? Not exponentially it appears so far. Therefore if the hoped-for path to AGI and superintelligence slows down and becomes incremental, then the immediacy of staying ahead starts to evaporate, giving competitors more time to catch up, probably with new approaches. The US in trying to win the race may just close itself off from actually winning.
Get it that the conquest of Gaul is a well known book but as a policy we should translate foreign language quotes, especially those in dead languages. It’s a little obnoxious not to. My two cents