[P] Vibecoded on a home PC: building a ~2700 Elo browser-playable neural chess engine with a Karpathy-inspired AI-assisted research loop

Developer Adam Jesion has open-sourced Autochess NN, a neural chess engine built using what he calls 'vibecoding'—an intensive, AI-assisted research and development loop inspired by Andrej Karpathy's workflow. Using a single consumer-grade RTX 4090 GPU, Jesion employed AI tools to read papers, prototype ideas, and iteratively optimize the system from supervised learning to self-play reinforcement learning. The current public version (V3) is a 16-million parameter model combining a residual CNN with a transformer that uses 'learned thought tokens.' It was trained on over 100 million chess positions, starting with a 2200+ Elo Lichess dataset, fine-tuned with Syzygy endgame tables, and refined through self-play RL with search distillation. Remarkably, the final model runs via CPU inference with a shallow 1-ply lookahead, delivering move calculations in under 2 milliseconds.

Jesion wrapped the engine in a fully-featured, free browser application that makes the AI's decision-making transparent. Users can play against the AI, use a board editor, import PGN files, replay games, solve puzzles, and—crucially—see a move analysis panel that displays the engine's top-move probabilities and how its 'thinking' step shifts them. The project demonstrates an unusually compute-efficient path to strong performance, with Jesion estimating the engine's strength at approximately 2700 Elo. He is now developing V4 with a 50M-parameter architecture featuring Dynamic Attention Bias (DAB) and speculating on a V5 concept called 'Temporal Look-Ahead,' where the network internally represents future moves to inform current decisions. The work challenges assumptions about the resources needed for advanced AI system development and provides a blueprint for iterative, tool-augmented research.