Nathan Lambert: belief

Transcript context

…Can we take this actual tangent and we’ll return back to the hardware, is the philosophy, the motivation, the case for export controls? What is it? Dario Amodei just published a blog post about export controls. The case he makes is that if AI becomes super powerful and he says by 2026, we’ll have AGI or super powerful AI and that’s going to give a significant … Whoever builds that will have a significant military advantage. And so because The United States is a democracy and as he says, China is authoritarian or has authoritarian elements, you want a unipolar world where the super powerful military, because of the AI is one that’s a democracy. It’s a much more complicated world geopolitically when you have two superpowers with super powerful AI and one is authoritarian. So, that’s the case he makes. And so the United States wants to use export controls to slow down, to make sure that China can’t do these gigantic training runs that will be presumably required to build the AGI. This is very abstract. I think this can be the goal of how some people describe export controls, is this super powerful AI. And you touched on the training run idea. There’s not many worlds where China cannot train AI models. I think export controls are decapping the amount of compute or the density of compute that China can have. And if you think about the AI ecosystem right now, as all of these AI companies, revenue numbers are up and to the right. Their AI usage is just continuing to grow, more GPUs are going to inference. A large part of export controls, if they work is just that the amount of AI that can be run in China is going to be much lower. So on the training side, DeepSeek V3 is a great example, which you have a very focused team that can still get to the frontier of AI on … This 2,000 GPUs is not that hard to get all considering in the world. They’re still going to have those GPUs. They’re still going to be able to train models. But if there’s going to be a huge market for AI, if you have strong export controls and you want to have 100,000 GPUs just serving the equivalent of ChatGPT clusters with good export controls, it also just makes it so that AI can be used much less. And I think that is a much easier goal to achieve than trying to debate on what AGI is. And if you have these extremely intelligent autonomous AIs and data centers, those are the things that could be running in these GPU clusters in the United States, but not in China. To some extent, training a model does effectively nothing. They have a model. The thing that Dario is sort of speaking to is the implementation of that model, once trained to then create huge economic growth, huge increases in military capabilities, huge increases in productivity of people, betterment of lives. Whatever you want to direct super powerful AI towards, you can, but that requires a significant amounts of compute. And so the U.S. government has effectively said … And forever, training will always be a portion of the total compute. We mentioned Meta’s 400,000 GPUs. Only 16,000 made Llama. Right? So the percentage that Meta’s dedicating to inference, now this might be for recommendation systems that are trying to hack our mind into spending more time and watching more ads, or if it’s for a super powerful AI that’s doing productive things, it doesn’t matter about the exact use that our economic system decides. It’s that, that can be delivered in whatever way we want. Whereas with China, you know, your expert restrictions, great. You’re never going to be able to cut everything off. And I think that’s quite a well-understood by the U.S. government is that you can’t cut everything off.…

Stored transcript either side of the excerpt. The highlighted words are the published quote; the surrounding text is unedited source, never generated.