Chess AI: From Brute Force to Neural Networks

▶ Watch (3:00)

Claude Shannon proposed two chess engine types in 1949. Type A used brute force. Type B selected moves intuitively. Computers grew fast. Type A engines scaled until Deep Blue beat Kasparov in 1997. Later AlphaZero used neural networks to play at superhuman level. LLMs today fail at chess because they predict language. DeepMind trained a transformer to predict Stockfish evaluations from millions of positions, playing at grandmaster strength, but it cannot explain its moves.

Grounded Chess Explanations

▶ Watch (7:00)

After a game, Stockfish evaluates every move. Custom detectors identify forks, pins, skewers, and doubled pawns. Maya, a neural network from the University of Toronto, predicts human move probabilities given a rating. All this context feeds Gemini 2 Flash. The LLM translates structured data into natural language. This prevents hallucination. The commentary explains why a move is brilliant or bad, referencing the threats and opportunities. End-to-end latency stays under 3 seconds.

Autonomous Agent for Feedback

▶ Watch (9:58)

Users can report bad commentary. The app posts it to Slack and injects it into a Cloud Code channel via MCP. Cloud Code runs a triage skill. It investigates the position, modifies prompts or detectors, regenerates commentary, and verifies its work. It then asks a developer for confirmation on Slack. If approved, it submits a pull request. The developer merges from a phone. This loop reduces fix time from days to minutes.

Latency, Evals, and Domain Expertise

▶ Watch (12:44)

Gemini 2 Flash delivers 3-second end-to-end latency. Time to first token is about one second. Reasoning models are too unpredictable. The team built 16 scenarios covering tactical patterns, blunders, and hallucination limits. They use LLM as a judge and compare models via OpenRouter. Gemini scores 75%. Claude thinking scores 60%. GPT-5 mini scores lower. Domain experts verify output. Separating the data pipeline from language generation is the key lesson.

Notable Quotes

LLMs often hallucinate because obviously they’re trained on language, they can’t calculate. Asbjørn Steinskog · ▶ Watch (6:01)

the LLM’s job is only to translate this information into English Asbjørn Steinskog · ▶ Watch (8:58)

Typically when you finish a chess game, you want to get the analysis and the results pretty quick. Anant Dole · ▶ Watch (12:51)

Key Takeaways

  • Separate data pipeline from language generation for low-latency AI features.
  • Use autonomous agents with human confirmation to close the feedback loop.
  • Build evaluation scenarios with domain experts to maintain quality.