It seems like an even more useful thing to do would be to come up with clear conditions under which they would support a pause (so not just RSI but a definition of RSI). Obviously this is really hard! But it seems like an obvious point of failure here is that the labs all agree on a conditional pause but the condition is something unverifiable, vague or very low probability (e.g. we’ll pause if there’s an AI disaster that kills thousands).
It seems like an even more useful thing to do would be to come up with clear conditions under which they would support a pause (so not just RSI but a definition of RSI). Obviously this is really hard! But it seems like an obvious point of failure here is that the labs all agree on a conditional pause but the condition is something unverifiable, vague or very low probability (e.g. we’ll pause if there’s an AI disaster that kills thousands).