Fable 5.2 chess 24hrs was designed, written, trained and tuned by Anthropic's Fable 5.2 model in a single autonomous 24-hour session. It started with an empty folder and a C++ compiler. Nobody answered questions, and no existing engine code, network or book was allowed. It did not even look anything up online. It wrote its own NNUE trainer and trained the network on its own self-play games along the way.
Rating: 3542 ± 15 on the CCRL Blitz scale (2,220 games at 10s+0.1s, single thread, against ten CCRL-rated engines and three of the other AI engines).
Why I'm excited about this one:
- +81 Elo over Opus 5.5 (3461), which had only just set the record with a 207 Elo jump over Fable 5.1.
- +288 Elo over Fable 5.1 (3254). That is the gain from one point release of the same model line, under the same rules and the same 24 hours.
- It is the first engine in this series that I'd say has an attractive playing style. It doesn't just win on tactics; the games are a pleasure to watch, and it feels like a properly strong engine.
- It is about 260 Elo behind Stockfish 19, which is 3803 on the current CCRL Blitz list (and that is Stockfish on 8 CPUs against this engine's single thread). For one day's unassisted work, I find that remarkable.
What's inside:
- PEXT bitboards, about 2,200 lines of C++
- PVS with null move, LMR, singular extensions, ProbCut, and history and correction-history tables
- NNUE: (768x5 -> 256)x2 -> 1x8, SCReLU, horizontally mirrored, trained from zero on about 175M self-play positions (CPU only, no ML framework)
- 55 search parameters tuned by SPSA during the session
https://github.com/stevemaughan/fable-5 ... s/tag/v1.0
Source, the model's own write-up and its hour-by-hour log:
https://github.com/stevemaughan/fable-5.2-chess-24hrs
The benchmark rules and template, if you want to try another model:
https://github.com/stevemaughan/chess-engine-benchmark
I'd love to hear what you think of its play, and independent test results are very welcome.
Steve