01INTERPRETABILITY / JUN 2026
Searchless Chess
A look inside DeepMind’s 270M-parameter chess transformer. Probing what it knows about the board—and where that knowledge lives.
Python · JAX · Scikit-learnprobe_results.txtREAD ONLY
Material balance, recovered.
0.999R²
Strongest interpretable structure: layers 1–3.
Research notes
THE METHOD
Trained 136 linear probes across all 17 layers, testing eight chess concepts on 5,488 positions. Recovered material balance with R² = 0.999 and center control with R² = 0.989.
THE FINDING
Interpretable structure peaked in layers 1–3, then degraded as later layers shifted toward action values. A Chrome extension and WebSocket pipeline visualized move preferences during post-game analysis.