Vibeleaderboard
← All Intel
Intel / repo

nanoAlphaZero – Train a grandmaster-level chess model in 24h with TPUs

Source
github.com
Author
tdoubleu
Date

Context

nanoAlphaZero reports training a chess model for 24 hours on a TPU v4-32 pod, then scoring 54.1% of available points in 512 games against the Searchless Chess 270M model. That used 400 search simulations, distinct from the roughly 12 used during training.

The project's grandmaster-level description rests on that opponent having previously achieved a 2895 human Lichess Blitz rating. The direct result is a match against another model under specified search conditions. It does not assign the new model a directly measured human rating.

Terms in this piece · Glossary
  • eval — A repeatable test for AI quality — a set of tasks plus scoring — used the way software teams use test suites, because model output is too variable to judge by eyeballing.
Recommended reads
Comments

Checking sign-in…

Loading comments…