Research into relative effects of different pacing interventions

Project summary

I will do research into the question under the title

What are this project's goals? How will you achieve them?

To produce research outputs answering the question under the title.

I would turn the appendix into a table with one row for each intervention, ordered by how long it would delay deployment of AIs that rival current leading systems. The first column in this table is yes or no: can the leading lab substitute around the rule? If yes, that rule doesn’t slow them down - so it goes to the bottom of my list immediately, with no delay estimate needed. (I expect about a third of the list will fall here.) The remaining rows would be sorted into tiers, corresponding to how many months of delay they would buy. These need to be wide enough to be honest, but narrow enough to make a real distinction between something taking a month versus a year. The only place that the value of delaying progress in AI by a month enters is at the end, when we decide how far down this list we should go before stopping; it’s just some number that we apply equally to each row in the table.

There would also be an incidence column, i.e. for each regulation does it help or hurt the leader in the race? This is really important, because what distinguishes safety measures from cartels is the sign. So this column would be published right next to every row.

This entire task should take a few weeks to a couple of month if I use public information on the progress of the leading labs and keep a detailed spreadsheet. However, to get to the point of having a table with a credible amount of accuracy on each rule, I would need a collaborator who can put in the work to estimate the effect of each rule on the leading labs, row by row. Once this table existed, I would expect both sides of the debate to be forced to argue about the actual details in the table, rather than general slogans and heuristics. I’d love to get started on that first cut.