Claim 1: Limiting algorithmic progress is the best way to “pace” the frontier 

I think the right mechanism in the near term involves limiting inputs to algorithmic progress, for two reasons:

First, AI companies have not secured their algorithmic progress, and so any marginal algorithmic progress will probably be stolen and then used to accelerate competitors. Therefore, we should pause (or strongly limit) algorithmic progress until it can be secured.

One objection to this is that securing the algorithms is probably incredibly difficult. So won't China overtake us? My counterarguments are: (1) we could limit algorithmic progress enough that we slow down a lot but still outrace China, (2) doing an intelligence explosion without the ability to secure the algorithms seems like an incredibly bad plan (or that ability coming only very late, e.g. after full automation of AI R&D), (3) I'm in favor of incremental de-escalation of the AI race, even if it narrows the US/China lead, because I think a full on race runs such a high chance of AI takeover, and (4) If the US adopts this policy, then we will be strongly incentivized to pursue additional policies that would prevent China from building unsafe AGI/ASI (such as Plan A or Plan S). 

Second, it's (basically) the only way to slow without algorithms overhang. I’ve argued for this previously here. 

Insofar as we aren't able to make good safety cases now for the next generation of model capabilities (which imo we can't; I think we're entering the regime where each step of development entails non-negligible existential risk), then this implies we should limit final capabilities in the immediate future, which implies that AI companies & governments should immediately take steps to limit algorithmic progress. 

Claim 2: Verifying limitations on algorithmic progress will centrally involve compute accounting.

Compute accounting would involve defining some categories with how an AI company spends its compute (e.g., internal inference, external inference, running monitors, capabilities R&D, and non-capabilities R&D) and then publicly reporting: the percentages on each category. As an additional measure, AI companies could publish a high level description of N (e.g. 30) projects in the non-capabilities R&D bucket sampled weighted by compute use, on the theory that if a project isn’t capabilities, AI companies should be OK describing what it is to their competitors, and that this helps avoiding this category from being abused by fake safety research). 

The best proposals for limiting algorithmic progress that I’m aware of are: compute reallocation, internal deployment lags, and, a bit more speculatively, “mandatory vacation”, where AI companies would take a mandatory whole-company break from everything capabilities related for N days a month. Verifying all of these involves compute accounting:

• compute reallocation and mandatory vacation involve directly accounting the compute, either to a percentage, or to zero. 
• Internal deployment lags involves arguing that there aren't extra workloads running the most capable models somewhere you don’t know about.

We’ve seen relevant initial steps towards compute accounting out of OpenAI and Anthropic. I think it should be a top priority for AI companies to (1) give a more complete accounting of how compute is being used and (2) allow competent 3rd parties to verify the claimed compute accounting is correct.