flyword.A SMALL BRAIN AT PLAY
CONNECTOMICS, MEET WORDPLAY

A tiny brain.
A different kind of player.

Can you out-spell a brain the size of a poppy seed? Every candidate board runs through
measured fruit-fly wiring. Does the real wiring beat a shuffled one? Not in this run.

Play the fly ↓
  • 173,587measured synapses
  • 168,550playable words
  • 1learned output layer
01

The board

You0
VS
The fly0
YOUR TURN
Triple wordDouble wordTriple letterDouble letter
YOUR TILES86 tiles in the bag
POINTS—

Loading the measured circuit and word list…

How to play

Click a square or enter a coordinate (H8 is the center), type a word, then choose across or down. Clicking rack tiles adds letters to your word. The first move must cover the star; later moves must connect. Blanks score zero and are assigned automatically. Find a word suggests the highest-scoring legal move. The fly replies on its own a moment after you play. Six consecutive scoreless turns end the game.

02

Inside the mushroom body

LIVE CIRCUIT
A little wiring. A lot of possibility.WAITING FOR INPUT
…PROJECTION NEURONSReceive the board
…KENYON CELLSExpand & sparsify
…OUTPUT NEURONSApproach / avoid
0Kenyon cells active
~5% SPARSITY
↳

The wiring was measured from an actual fly brain.
Nothing in the expansion layer was trained.

03

A decision, in the making

MOVE EXPLORER

Every row is a legal move. The fly plays the one its circuit prefers most; that preference is near chance, so it chooses valid words but rarely strong ones. Score only breaks ties.

✳
Every word starts a different thought.

Play your opening word to watch the fly consider its reply.

Fly preference · candidate afterstate
THE CONTROL EXPERIMENT

Does the wiring matter?

Same inputs. Same learning. Different connections.

Degree-preserving shuffle · only the output learns
REAL…
SHUFFLED…
RAW LOGISTIC…
Held-out game accuracy
FREQUENTLY ASKED

What is Flyword?

Is Flyword a free online Scrabble game?

Yes. Flyword is a free, browser-based Scrabble-style word game with no download, no account and no ads. You play against an AI opponent on a standard 15×15 board with double/triple letter and word squares, using the open ENABLE word list. It is an independent research toy, not an official Scrabble product.

What does it mean to play against a fruit fly brain?

The opponent's decision circuit is built from the real FlyWire connectome of a Drosophila melanogaster mushroom body: 173,587 measured synapses between projection neurons, Kenyon cells and mushroom body output neurons. Every candidate move is fed through that fixed biological wiring; only a small output layer is trained. When the fly picks a word, you watch the sparse neural code that produced it.

Is this real neuroscience or just a neural network?

Both. It is a biologically constrained neural network: the connections come from electron-microscopy reconstructions of an actual fly brain (FlyWire FAFB v783), not from random initialisation. A control experiment with degree-preserving shuffled wiring shows how much the real anatomy matters for this task. Full methodology, evaluation metrics and downloadable data are on the methodology and data pages.

Who is Flyword for?

Word-game fans who want a new kind of Scrabble opponent, students and teachers exploring connectomics or computational neuroscience, and anyone curious how a brain the size of a poppy seed makes decisions. Works on phone, tablet and desktop.

NOTES FROM THE LAB

A fly-inspired player,
with measured wiring.

This experiment uses the left mushroom body from FlyWire FAFB v783. Anatomical PN → KC synapse counts stay frozen. A fixed random projection maps a 1,115-dimensional binary board-and-rack encoding into PN activity. APL feedback suppresses KCs, then at most 5% fire.

Only KC → MBON task weights are fitted. Weights are supported on observed KC → MBON contacts; summing MBON contributions produces an eventual-winner prediction. The approach/avoid labels describe this learned task signal, not a biological claim about Scrabble.

The control uses bipartite edge swaps preserving each PN’s outgoing degree and each KC’s incoming degree. A raw-input logistic regression provides a third baseline. All three use the same split by whole game. Hovering evaluates the proposed board and remaining rack before drawing unseen tiles.

Transmitter predictions use ACh as positive and GABA/glutamate as negative. These signs, normalization, the 0.5 APL feedback gain, input projection and firing cap are explicit modeling choices. Winner prediction on expert games is not a measurement of playing strength.

The playable lexicon is ENABLE, an open word-game list, rather than a licensed tournament dictionary. This is an independent research toy, not an official Scrabble product.

Full evaluation reportSource hashes & binary manifest ↗Attribution & licenses
A FRESH PERSPECTIVE

Swap a few letters.

Enter the tiles you want to return (use ? for a blank). New tiles are drawn before yours go back into the bag. Swapping uses your turn and requires at least seven tiles in the bag.
GAME OVER

You win!

YOU0
VS
THE FLY0