Eon BPU Chess Lab

biological processing unit Drosophila larval connectome

source connectome Frozen Drosophila wiring, trained only through thin input/output projections.

The live BPU panel above the chessboard now carries the simultaneous activity visualization.

neurons2,952

synapses~63K

corefrozen recurrent

Greedy puzzle accuracy 27.4% BPU vs Transformer 16.7%

Search-augmented BPU 49.2% alpha-beta depth 6, full 10K eval

Relative lift over AI +64% parameter-matched greedy eval

modern AI baseline Transformer 1.46M params

connectomics BPU Drosophila Fly 1.68M trainable params

current position Configure the lab match

0 / 18 waiting

match configuration Parameter-matched proof

AI decision time 118 ms

Fly BPU decision time 104 ms

AI playing strength AI 1270 Elo

Fly BPU playing strength BPU 1710 Elo

live transcript BPU advantage

match summary

CNN-BPU beat a parameter-matched Transformer on the paper's ChessBench evaluation: 27.4% vs 16.7% greedy puzzle accuracy. Start the match to watch the inference stream.

What is real in this demo?

The live match runs exported BPU and Transformer checkpoints directly in the browser, then masks each policy to legal chess moves. The measured result is the repo reproduction: a CNN-BPU with a frozen Drosophila larval connectome core beats a parameter-matched Transformer on ChessBench greedy puzzle accuracy, and alpha-beta search nearly doubles the BPU score.

Dataset

ChessBench, 527M Stockfish-labeled positions

Eval

10K Lichess puzzles, legal-move masked predictions

BPU

2,952-neuron frozen recurrent Drosophila connectome

Baseline

1.46M-parameter Transformer