Although it’s not a plan to be chosen by US and therefore out of reach for AI2040, do you think it plausible that some nuclear armed country losing the race(ex Russia) threatens US and China to halt with the threat of nuclear war and it results in mutual slowdown? It’s definitely unpreferable but seems somewhat likely.
We need Training-Run Assessments!
AGI is extremely consequential. There should be independent assessments not just for the final checkpoint, but for the full process.
This matters most for scheming where the whole problem is that it might be hard to detect in the final ckpt.