Austin
Building things at the intersection of AI research and engineering.
Writing
Poirot Confronts Goodhart
Using RL to train a text compressor — five reward formulations, five failure modes, and why the best evaluator makes the worst training signal.
Poirot's Judgement: Building an Automatic Scorer for Text Compression
An information-theoretic scorer for text compression — and why the language model matters more than the formula.