A few thoughts on these remarks by Secretary Bessent on AI dialogues with China:

(1) Secretary Bessent is correct that the US has a lead in AI, and that the most recent US models represent step-changes in capabilities. Future model releases will be even bigger step-changes, which is why we need to maximize the US lead over China--not sell it the tools to catch up.

(2) China sees these step-changes in capabilities happening too, which is making China feel greater urgency than ever to catch up to the United States. This is China's top priority in AI.

(3) It's true that the United States and China will face shared threats from non-state actors obtaining advanced AI models. It is also clear that China is more concerned about catching up to the United States than mitigating threats from non-state actors. And Chinese AI-enabled cyberattacks/military capabilities will also soon pose a massive threat to U.S. national security.

(4) So long as this remains China's perspective, it will be impossible to reach an effective agreement with China on AI guardrails (that China will comply with). China will not constrain its own AI development in any way, unless it has given up hope on catching up with the United States and fears that if it does not impose such constraints the United States would use its more powerful AI capabilities to cause significant damage to China. This is not the world we are in right now.

(5) Talking about AI guardrails with China is logical, provided the United States is clear that those conversations will not include export controls. When the Biden administration attempted these discussions with Beijing, China did not engage on safety risks and instead requested the United States relax export controls - which we told them was off the table. It would also be a lot easier to have those discussions with China if we had any formal AI guardrails in place domestically, which we could insist that China implement too. Right now, we have no legal requirements.

(6) Secretary Bessent's comments do not seem to bode well for the future of open-source AI in the US. If the United States is (correctly) concerned about non-state actors obtaining powerful models, at some point the United States must restrict what AI capabilities can and cannot be open-sourced. Worth watching how this evolves.