34 points | by alvis 2 days ago ago
1 comments
Useful direction, but the hard part seems to be measuring novelty after each fix. Are they reporting whether later red-team cases are genuinely distinct, or mostly variants of the same failure mode?
Useful direction, but the hard part seems to be measuring novelty after each fix. Are they reporting whether later red-team cases are genuinely distinct, or mostly variants of the same failure mode?