A fruit fly's brain is playing
a game I wrote for it.

Watch it play

This is not a video. Every time you reload, a fresh episode is generated on a two-core VPS: the connectome runs step by step and streams each frame here as it is computed, so you are watching the fly decide, roughly live. The right panel is what its giant-fibre escape neurons were doing at that instant. Pick a different decoder, or ask for another run — the brain never remembers the last one.

connecting to the connectome…

Who is actually playing?

Same brain, same game, four different ways of turning its spikes into moves. Every run is a fresh episode; hitting an obstacle ends it.

policy	mean	best	runs	full 18 s
random	93 m	98 m	2	0
giant-fibre reflex	88 m	171 m	6	0
linear readout on the brain	164 m	216 m	6	2
oracle (knows the future)	216 m	216 m	2	2

How the loop closes

1. The game writes a scene

   Three lanes, obstacles sweeping in at 12 units per second, tall blocks and low barriers. The game knows nothing about the fly.

2. The eye turns it into optic-lobe drive

   Angular growth feeds LPLC1/LPLC2, close-range threat feeds LC4, on the side the obstacle is really on. This encoder ships with the simulation package — it is the part that is mine, and it is the part to be suspicious of.

3. The connectome runs

   20 ms of simulated time per step, every neuron and synapse accounted for. Measured on this box: about 14 ms of CPU per step, roughly 70 steps a second — faster than the 50 a second the game needs, which is why the run can stream at full speed. No GPU, no training, no backpropagation.

4. A readout reads the descending neurons

   One linear model predicts a target lane, a second predicts a jump. That prediction is the only thing that ever touches the keyboard.

Stated plainly: the eye is mine, the wiring is the fly's, and the readout is a two-layer linear model. "The fly plays" means the frozen connectome carries a decodable signal for what to do — not that it learned the game.

First, the control experiment

Before trusting any of the above: does a single spike from the connectome actually move a body? One threat, one correct answer, and a lesion to prove causation.

The lesion is the point. Silence that one neuron and the fly is hit every single time, at exactly zero clearance. Not a correlation — a cause.

What this does not show

• It cannot jump. With three lanes and at most two blocked, dodging always works, so the oracle never jumped once. The jump decoder was trained on zero positives and is at chance level. That pathway is untested, and I left it out of the claims rather than dress it up.
• The lane signal is coarse. Regression error of 2.06 on a target spanning three lanes. Descending neurons carry a crisp categorical "danger, that side" and only weak graded information — which is exactly why the learnt player still gives up a quarter of the oracle's distance.
• Six runs per policy. Enough to rank them, not enough for tight error bars.
• This is not the fly's own eye. The encoder is a four-channel model of the optic lobe. Change the encoder and the numbers move — the wiring is the fly's, the question is mine.