Rexigon chess - new engine built with Kiro AI / Claude 3.5

[silentshark]

I imagine before long there will be thousands of these GenAI built engines appearing. So I had a go, using Kiro which uses Claude 3.5 as the brains. Been playing round with this since the summer.

I called the engine Rexigon, for no apparent reason. Built doing this 'vibe coding' thing where AI does most of the heavy lifting. Some bits and bobs

- currently incorporating the Pesto PST's (only) for eval.
- playing around 2400 ELO (CCRL scale)
- nothing innovative, uses all the usual techniques - I added one thing at a time, then tested, looking for obvious regressions. Here's the order of stuff I added:
- alpha-beta search with basic PST's
- quiescence
- basic move ordering
- LMR and NMP
- hash tables
- draw scoring
- better PST's and scoring interpolation between mid and endgame
- pesto's PST's
- static NMP
- killer heursitic
- SEE and drawscore fixing
- history heuristic
- more aggressive LMR
- time handling improvments
- improved SEE
- better time usage
- more LMR tweaks
- incremental PST's to speed things up
- singular extensions and multicut - then removed it
- NMP improvement
- LMP

I found that often the vibe coded stuff is not great, it missed stuff or just gets things wrong.

Anyway, I can pop onto github if anyone wants a peruse. I have a hunch that with a bit more work (eval is just Pesto PST right now so could be improved), it will likely catch my engine, Francesca, which took many years of blood, sweat and tears to develop! That was back when source code for chess programs wasn't really available - think there was an early version of GnuChess then Crafty, but nothing much!

Cheers, all,
Tom

[smatovic]

2400 CCRL Elo with PeSTO eval, nice.

What kind of board representation does the engine use?

I would add it to the YATT thread, but the engine did not develop the evaluation on its own.

Giovanni Lavorgna showed a way how to use a driver script to automate the vipe coding process:

GeminiChess, an LLM built engine
viewtopic.php?t=85490

Jost Triller showed how to generate Stockfish patches:

Re: ChatGPT usage in computer chess?
viewtopic.php?p=981272#p981272

I think this is where it starts to get interesting, a test-framework feedback loop with generative AI for chess engine development.

--
Srdja

[glav]

I imagine before long there will be thousands of these GenAI built engines appearing. So I had a go, using Kiro which uses Claude 3.5 as the brains. Been playing round with this since the summer.

I called the engine Rexigon, for no apparent reason. Built doing this 'vibe coding' thing where AI does most of the heavy lifting. Some bits and bobs

- currently incorporating the Pesto PST's (only) for eval.
- playing around 2400 ELO (CCRL scale)
- nothing innovative, uses all the usual techniques - I added one thing at a time, then tested, looking for obvious regressions. Here's the order of stuff I added:
- alpha-beta search with basic PST's
- quiescence
- basic move ordering
- LMR and NMP
- hash tables
- draw scoring
- better PST's and scoring interpolation between mid and endgame
- pesto's PST's
- static NMP
- killer heursitic
- SEE and drawscore fixing
- history heuristic
- more aggressive LMR
- time handling improvments
- improved SEE
- better time usage
- more LMR tweaks
- incremental PST's to speed things up
- singular extensions and multicut - then removed it
- NMP improvement
- LMP

I found that often the vibe coded stuff is not great, it missed stuff or just gets things wrong.

Anyway, I can pop onto github if anyone wants a peruse. I have a hunch that with a bit more work (eval is just Pesto PST right now so could be improved), it will likely catch my engine, Francesca, which took many years of blood, sweat and tears to develop! That was back when source code for chess programs wasn't really available - think there was an early version of GnuChess then Crafty, but nothing much!

Cheers, all,
Tom

Congratulations! I would be very interested too in the source code and in any info you feel to share about the project.