Work

I have a theory that with enough compute, pure policy neural networks can beat SOTA chess engines. I work on building quality datasets, pretraining on Stockfish soft targets, and methods of RL.

Instruct model: think, then move. Stockfish for labels and eval.

Opponent-adjusted college football ratings: how many points a team would beat an average FBS team by on a neutral field. I’m competing in a college football model pick ’em competition.

Pre-snap play inference for NFL / FBS.

Live fantasy draft assistant for a league's scoring.

Small LM trained from scratch without encyclopedic facts.

Distill a teacher into a smaller student via activation match.

I’m interested in primes and their distribution.