This is a special post for quick takes by Morgan S. Only they can create top-level comments. Comments here also appear on the Quick Takes page and All Posts page.
It seems like OpenAI and Anthropic have a substantial lead over everyone else, so if they want to make pacing happen they have to work something out between themselves first. One way to start might be a 1-day pause.
The simple version would be to shut everything off, including inference (which would certainly get everyone's attention as a costly signal). IIUC it wouldn't be that hard to check that all their data centers aren't using any power.
If they want to serve inference that day, it becomes harder to verify that they're not doing a big training run, although with 3rd parties involved this doesn't seem crazy.
A variation on this is to license one team on each side to covertly do a training run, and see if the other side can detect it. This would serve to get real evidence on how hard this is to do and what tools are missing on the verification side.