I too "am also perplexed that the fact everyone in the AI industry is saying this deeply worrying and perturbing stuff isn't drawing more attention." Even more so considering that the limited guardrails from the prior administration have been removed. What is your sense of how the upcoming "AI Action Plan" might impact AI development? Do you think there will some effort to provide basic regulation in an attempt to prevent worst-case scenarios, or unfettered development?
I've found that the issue of AI models having potential national security capabilities is one that draws bipartisan concern and attention. My sense of this administration is that they want clear evidence of risks to guide their policy. I'm hopeful we'll get some form of 'national security testing of AI models' though the exact way we achieve that is unclear, and it may be that it's more of a norm among companies than a hard regulation (though in fairness, the Biden admin stuff mostly worked on the basis of voluntary commitments the companies made, so there's precedent here). I don't have a great sense of the AI Action Plan - though I expect it'll help further motivate investment in things like electricity infrastructure and chips.
Thanks Jack for highlighting the Bytedance paper on inference cluster. I suspect they have 10s of thousands of H20 and L40 GPUs. I am working on detailed look at each major Chinese AI developer and what their business model is, and how they are using heterogeneous clusters to do training and inference, and doing their own optimizations across a range of pieces of the hardware stack....
I too "am also perplexed that the fact everyone in the AI industry is saying this deeply worrying and perturbing stuff isn't drawing more attention." Even more so considering that the limited guardrails from the prior administration have been removed. What is your sense of how the upcoming "AI Action Plan" might impact AI development? Do you think there will some effort to provide basic regulation in an attempt to prevent worst-case scenarios, or unfettered development?
I've found that the issue of AI models having potential national security capabilities is one that draws bipartisan concern and attention. My sense of this administration is that they want clear evidence of risks to guide their policy. I'm hopeful we'll get some form of 'national security testing of AI models' though the exact way we achieve that is unclear, and it may be that it's more of a norm among companies than a hard regulation (though in fairness, the Biden admin stuff mostly worked on the basis of voluntary commitments the companies made, so there's precedent here). I don't have a great sense of the AI Action Plan - though I expect it'll help further motivate investment in things like electricity infrastructure and chips.
Thanks for a detailed question!
Surely a 100+ page pdf is no longer marketing
One would hope!
Thanks Jack for highlighting the Bytedance paper on inference cluster. I suspect they have 10s of thousands of H20 and L40 GPUs. I am working on detailed look at each major Chinese AI developer and what their business model is, and how they are using heterogeneous clusters to do training and inference, and doing their own optimizations across a range of pieces of the hardware stack....
I would love to read this analysis (and likely highlight it here) - please email me!