LucaRighetti comments on Dangerous capability tests should be harder

[LucaRighetti]

Then that seems bad, but also that AI is not counterfactual—so adding safeguards to models is probably not the way to get the risk down.