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

2026-03-21 · Source: Machine Learning · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Gaming & Interactive Media · Depth: Advanced, short

Summary

Adam Jesion built Autochess NN, a browser-playable neural chess engine, as a personal experiment to understand AlphaZero-style systems. The project, developed using an AI-assisted research loop on an RTX 4090, features a residual CNN + transformer architecture with ~16M parameters and learned thought tokens. It uses a 19-plane 8x8 input and a 4672-move policy head plus a value head, trained on over 100M positions through supervised pretraining, Syzygy endgame fine-tuning, and self-play RL with search distillation. The engine, which performs CPU inference with shallow 1-ply lookahead under 2ms, is wrapped in a free browser app for interactive play, board editing, PGN import, puzzles, and move analysis. Jesion hypothesizes it is unusually compute-efficient for its strength, potentially exceeding 2500 Elo.

Key takeaway

For AI Scientists developing neural game engines, your focus should be on integrating AI-assisted research loops to accelerate development and achieve high Elo ratings with compute efficiency. Consider combining supervised pretraining with self-play reinforcement learning and explore novel architectural elements like "thought tokens" or "Temporal Look-Ahead" to push performance boundaries. Your next step could be to pressure-test compute-efficiency claims against established benchmarks.

Key insights

An AI-assisted research loop can rapidly develop high-performance neural chess engines on consumer hardware.

Principles

• Combine supervised pretraining with self-play RL.
• Integrate CNNs and Transformers for chess AI.
• Optimize for compute-efficiency at high Elo.

Method

The pipeline involves 2200+ Lichess supervised pretraining, followed by Syzygy endgame fine-tuning, and then self-play reinforcement learning with search distillation.

In practice

• Utilize an RTX 4090 for rapid AI model iteration.
• Implement a browser interface for model inspection.
• Explore "Temporal Look-Ahead" for future move representation.

Topics

• Neural Chess Engines
• AlphaZero
• Self-Play Reinforcement Learning
• Transformer Architecture
• AI-Assisted Development

Best for: AI Scientist, Research Scientist, AI Engineer, Machine Learning Engineer, AI Researcher

Related on AIssential

• Trained transformer-based chess models to play like humans (including thinking time) [P] — Transformer models can emulate human chess play and thinking times using large-scale Lichess data.
• AIエージェントが最適化プログラミングコンテストで初優勝 — AI agents can now win complex, real-time optimization contests by scaling LLM inference and self-learning.
• AI is rewiring how the world’s best Go players think — AI has transformed professional Go, shifting training paradigms and democratizing access while raising questions about creativity and human understanding.
• Building a Chess Coach — Anant Dole and Asbjorn Steinskog, Take Take Take — Combining classical chess engines with LLMs and human-like move prediction enhances AI chess coaching.
• Partnering with Ineffable Intelligence: A Superlearner for the Era of Experience — Superintelligence may emerge from AI systems learning purely through self-experience, unconstrained by human data.
• Augmenting Game AI with Deep Reinforcement Learning — This paper proposes a framework for training reinforcement learning models to overcome current limitations and accelerate machine learning adoption in game AI.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.