I have been somewhat more skeptical of US->China chip export controls, at least if done under the classical justification of "the US needs to stay ahead in the race", than most people in AI Safety. Some quick thoughts I wrote up in a group chat about that are cheap to share:

> Yeah oli do you have a favorite reference for your argument there? Seems wrong to me but I’m not super confident

My rough reasoning is that it's part of a racing strategy by the USG which is bad for the same reason as any other nation trying to concentrate all the GPUs under their control with the explicit goal of building powerful AI systems as quickly as possible would clearly be bad. Then beyond that, I think the specific justification for export controls has been particularly zero sum and racing focused and has substantially hurt international diplomacy.

There are also some secondary effects I am uncertain of. It seems like naively export controls put substantial pressure on the semiconductor supply chain to bifurcate and become more robust to shocks, but the current semiconductor manufacturing bottleneck is among the very best levers for many nation states in the world to slow down the AI race. In general companies trying to diversify the semiconductor supply chain seem very bad to me, and the export controls presumably have a large similar effect (but here I am reasoning more from first principles and am more uncertain)

I find arguments of the form "export controls put the US into a position of strength with regards to China and so make it easier to lead on a treaty" real but weak. Clearly only a very small fraction of the US's negotiation leverage is downstream of having a lead specifically in AI, and most of it is downstream of the general geopolitical power balance, and so I don't find it that compelling that the US's export controls will actually make a difference in what the US could negotiate for.

I think conventional escalation mechanics are for the near future going to be the primary driver that determines success of a treaty, not who is leading in AI (i.e. I think the U.S. would be in a similarly good position to ask China to halt AI progress if it had less of a lead, because the ask would be mostly based in economic sanctions and threats of conventional and very indirectly nuclear escalation, and that of course isn't changed very much based on the current state of the AI race. If anything the U.S. doing all of these export controls of course makes it much more suss for them to then turn around and ask everyone to stop, given that they've been taking all of those costly actions on the explicit rationale of wanting to go as fast as possible and using AI to dominate other countries and so asking other people to slow down will now require some very costly commitment from the US that they are actually stopping themselves)