We should also consider pacing based on limiting the ingredients that go into frontier models, such as training compute, the nature of training runs, or internal use of AI to improve AI. I do worry that some of these measures may be more “gameable” than external behavior, but this is the kind of topic worth discussing with embedded evaluators.Pacing within democracies will be limited by the lead that US companies have over authoritarian regimes, chiefly the Chinese Communist Party. If we slow down by more than this amount, then (unpaced) CCP-associated projects will pull ahead, creating significant national security risk.
Applying the Utilitarian Lens, Amodei points out that they should consider the pacing based on limiting the ingredients that go into frontier models, and he worries that some of the measures may be manipulated than external behavior. So what choice produces the most good and the least harm for everyone affected? Amodei points out how pacing within democracies will be limited by the lead that US companies have over authoritarian regimes, and that if they slow down, CCP-associated projects will pull ahead and create national security risk. Not slowing down produces the most good and least harm.