AI has moved so fast that five years ago I never could have imagined where we’d be today. It honestly feels like we’re living in a different era.

Every now and then I take on projects to really dig into what new models can do. Tony is one of them. We already know AI can write chess engines, so I wanted to make it harder: can it design its own programming language and then write a genuinely strong engine on top of it? The answer is yes, as long as you guide it well :)

Tony is written in T, a language with Turkish syntax. Its compiler needs nothing but Lua. It targets both x86 and ARM, and before you picture some toy project, this is close to a fully fledged compiler, AVX acceleration included. It produces native builds for Windows, Linux, and macOS directly.

The engine has its own NNUE architecture. I took the data generation process seriously, maybe a bit too seriously, and the result ended up only about 70 Elo behind top engines like Stockfish. That gap might look small, but it’s exactly why Stockfish is Stockfish: those last few Elo points take a huge amount of resources. I got this far mostly thanks to the experience I’d built up working on Turkish draughts for a while.

I’m not sure we’re ready for AI progress this sharp. But I can say I had a lot of fun.

Code: https://github.com/cturan/Tony

NNUE Data: https://huggingface.co/datasets/cturan/tony_nnue