
Nvidia and Anthropic Publicly Clash Over AI Chip Export Controls (cnbc.com) 19
Nvidia publicly criticized AI startup Anthropic on Thursday over claims about Chinese smuggling tactics, just days before the Biden-era "AI Diffusion Rule" takes effect on May 15. The confrontation highlights growing tensions between AI hardware providers and model developers over export controls.
"American firms should focus on innovation and rise to the challenge, rather than tell tall tales that large, heavy, and sensitive electronics are somehow smuggled in 'baby bumps' or 'alongside live lobsters,'" an Nvidia spokesperson said, responding to Anthropic's Wednesday blog post.
The Amazon and Google-backed AI startup had called for tighter restrictions and enforcement, arguing that "maintaining America's compute advantage through export controls is essential for national security." Anthropic specifically proposed lowering export thresholds for Tier 2 countries to prevent China from gaining ground in AI development.
Nvidia countered that policy shouldn't be used to limit competitiveness: "China, with half of the world's AI researchers, has highly capable AI experts at every layer of the AI stack. America cannot manipulate regulators to capture victory in AI."
"American firms should focus on innovation and rise to the challenge, rather than tell tall tales that large, heavy, and sensitive electronics are somehow smuggled in 'baby bumps' or 'alongside live lobsters,'" an Nvidia spokesperson said, responding to Anthropic's Wednesday blog post.
The Amazon and Google-backed AI startup had called for tighter restrictions and enforcement, arguing that "maintaining America's compute advantage through export controls is essential for national security." Anthropic specifically proposed lowering export thresholds for Tier 2 countries to prevent China from gaining ground in AI development.
Nvidia countered that policy shouldn't be used to limit competitiveness: "China, with half of the world's AI researchers, has highly capable AI experts at every layer of the AI stack. America cannot manipulate regulators to capture victory in AI."
Color me SHOCKED! (Score:3)
So, the company that makes money exporting their chips, is in favor of more exporting
And the company that makes money building proprietary models, is in favor of less exporting to countries who build competing models.
News at 11 folks!
Globalization happened. And you can't walk it back (Score:2)
Following the money can show motive. But where is the evidence in either case?
The problem with the US tightening up its technology and refusing to export means they will be competing against the entire world, and are going to lose. Because you get out of technology only as much as you put into it. If you put a little bit of one nation's economy into it, then expect your technology to advance slowly. If dozens of nations and hundreds of international corporations put money into technology, that's a much deep
Re: Globalization happened. And you can't walk it (Score:2)
I think there is no real evidence in any of these cases.
Export controls on base technology are a dumb waste of time. The idea that the US needs to restrict GPU exports to China rests on a very misplaced assumption that they can't just make their own - they can, and they can also do it better and cheaper. There are still too many people that think China is the China of the 90s, unable to do anything innovative. That day is long, long gone.
If you want to hand the global GPU market over to China, forcing them
Re: (Score:2)
And a laissez-faire free market capitalism can't compete with a centrally planned economy that has the will and expertise to throw their economic might behind a problem.
Imagine if every start up that wanted to make a CPU or AI chip suddenly got a loan from government-controlled banks and grants that almost completely cover payroll? That's free labor and cheap loans to purchase equipment.
The US is at a serious disadvantage compared to China, and we would be fools to take them head on, economy versus economy.
I find it quite shocking (Score:3)
that their positions correlate with what they perceive might be better for themselves from a business standpoint
Well.... (Score:3)
It you take the folks who wrote https://ai-2027.com/ [ai-2027.com] seriously, then the export controls are good for the US *IF AND ONLY IF* there's a short timeline to AGI., If AGI is delayed beyond around 2030 they will cause China to become more advanced in chip development than the US & friends.
Since I still estimate AGI to be around 2035 (plus or minus 5 years) that seems to make the export controls a bad idea.
Re: (Score:1)
2027 timeline isn't happening, already saturation is happening for reasoning models. Limitation is not compute, it's input training data.
AGI will take time if this path can ever get to it (looking at the structure of our brains, it's not as simple as just growing and retraining the model).
Thus this is completely idiotic. They will annoy everyone. And then why should America control who buys GPU, why shouldn't it be e.g. Netherlands? They are more critical in the semiconductor supply chain than USA.
Re: (Score:2)
The limitation is obviously not training data. Human brains can master subjects with a comparably tiny training set.
You are right that it will require architectural/topological improvements rather than simply scaling more.
"Algorithmic progress", as the ai-2027 people call it.
Re: (Score:1)
You can expose a dog to the same amount of training and he won't reach the same level, not with all data in the world. Our evolutionary algorithms (encoded in DNA) are just better at developing intelligence.
Re: (Score:2)
The limitation is obviously not training data. Human brains can master subjects with a comparably tiny training set.
LLMs are not human brains, so the amount of data needed to train the human brain is irrelevant to the discussion. Also, the irrelevant data doesn't count. Also, every single human experience both internal and external is part of the training set. How much data is that? How are you measuring it?
Re: (Score:2)
LLMs are not synonymous with or the end of the line for AI; you're the first one to even mention LLMs in this thread.
Sikiriki made a very definitive but false claim as to why "2027 timeline isn't happening"; the level a human brain can reach with far less data than AI is currently fed is very, very relevant. Clearly intelligence doesn't _require_ that much data.
You are right that a human processes a lot of visual and spatiotemporal input, which a lot of AI currently does not do. It is very much the question
Re: (Score:2)
LLMs are not synonymous with or the end of the line for AI; you're the first one to even mention LLMs in this thread.
I don't think that I am, but I don't remember my skimming it so well that I could say that for sure, and I'm sure as hell not going to go look. However...
I believe we are a few algorithmic improvements away from AGI.
Maybe. It's conceivable. But it took a long time to get where we are (with software that hallucinates and can credibly be said to pass a Turing test, if the tester is not too smart and educated anyway) compared to how long people thought it would take to get here. People have been saying what you've been saying for a long time. That was before we got to w
Re: (Score:1)
Currently, models don't improve much with more compute, Grok is good example. DeepSeek is another example where training cost was modest.
And the last big step up was using distilled reasoning traces of other models (i.e. a lot of reasoning data).
Re: (Score:2)
>> China to become more advanced in chip development than the US
That sure won't happen by 2030, or probably any time within the next 10 years. They don't have the manufacturing hardware and it will take quite some time even to reach parity.
Re: (Score:2)
>> China to become more advanced in chip development than the US
That sure won't happen by 2030, or probably any time within the next 10 years. They don't have the manufacturing hardware and it will take quite some time even to reach parity.
I've heard this prediction many times. I wonder what the basis behind it is. Is it that China has more people and engineers, and that population size advantage results in faster advancements? Or maybe the Chinese are smarter, more ambitious, more motivated, more methodical, more dedicated, etc. than the Americans? Or China has a better system for research based on the government structure, corporate structure, etc.
I can believe that there's a possibility that China may become more advanced than the US i
Re: (Score:2)
Basically, China is not a unitary entity. Some portions of it are quite advanced. Also, if they care to, they can afford to purchase things.
FWIW, there are reasonable arguments that it will take longer...but I don't think they are strong enough to be relied on.
Note that IF there are no fundamental technical advances, they the progress will be limited by two things:
1) The ability to make large scale wafers of sufficient purity/quality.
2) Access to the machinery for laying out the circuits. So far there'
What lunacy (Score:3)
The idea that restricting silicon exports can meaningfully constrain Chinese activity is so ridiculous, I don't know where to start. It might very very slightly delay some of their projects - it's absolutely not going to stop their progress. There may have been a chance to do that 25 years ago (maybe) - we are WELL passed that... and there's a real question about the wisdom of any of this paranoia. We are governed by idiots.
Anthropic's concern should be US competitors (Score:2)
Anthropic's worried about the Chinese getting Nvidia GPUs? To me, that's a misplaced concern. First, Chinese models aren't going to gain market share in the US, either due to anti-Chinese sentiment among consumers or due to explicit governmental restrictions. Second, Anthropic needs to worry much more about getting beat by US competitors. If Anthropic has trouble getting or affording AWS and Google Cloud hardware, the problem is too much US competition. Third, one disadvantage Anthropic has is that it