Real wiring.
Honest results.
A fly brain supplies the connections. Scrabble games supply the lessons.
Here’s what was measured, what learned, and how well it worked.
Loading the experiment…
Yes, this is still machine learning.
The model was trained on expert Scrabble games using frozen fly-brain connections. Only the KC → MBON output weights and output bias learned. The current game chooses the highest-scoring move, then uses the trained model to break score ties. Every candidate still runs through the circuit.
Does the wiring matter?
| Model | Accuracy | 95% interval |
|---|
A single seed and one split. These are predictions of the eventual winner on expert-played board states, not the win rate of the playable opponent.
The frozen circuit
Left mushroom body, as annotated in this source snapshot. Neuron counts are retained from the data. The in-game dots are schematic positions; the connections and synapse counts are measured.
What the model learned from
A proposed board, the known rack leave, and tiles remaining are encoded before drawing replacement tiles. Entire games stay in one split, so another turn from a test game cannot appear in training. Tied games are excluded.
Input encoding and modeling choices
The input has binary dimensions: 225 occupied flags, 450 tile-value bucket bits, 244 available premium-square channels, 189 rack-slot channels, and 7 bits for the bag count.
A fixed random projection drives PNs. Measured, signed PN → KC weights expand this input. APL feedback inhibits activity, and a top-5% cap makes the code sparse. The output coefficients are fitted on measured KC → MBON contacts.
The shuffled control rewires PN → KC edges while preserving every PN out-degree and KC in-degree, then refits its output weights. The raw logistic baseline bypasses the circuit entirely.
Sources & downloads
Original source files & SHA-256 hashes
Raw source data and training run offline. The app loads only the exported model and word list; gameplay runs in your browser.