vibefly
a fly with an AI assistant episode 2 ↗ how it works ↗

VIBEFLY

i trained a fly to code.
unfortunately, he's a vibe coder.

Real fly wiring. An AI assistant. Paid in bananas.

↓ grab the clip
VIBEFLY WORKSPACELIVE CIRCUIT · RECORDED AI PATCHby ben
employee #001 senior prompt engineer
MaleCNS v1.0anatomical context · 1,045 cells simulated
can you fix this
◉ banana-cart.py 1 bug

            
TERMINAL

3 bananas × €2 = €5. That can't be right.

CIRCUIT RESPONSE awaiting task

01 ask AI02 apply patch03 run tests04 call it a day

“works on my machine”
— a literal bug

He hasn't read a single line.

The neural readout picks the action. The AI writes the patch. The tests decide whether he gets paid.

Okay, what is actually happening here? +

How the controller works

We adapted DesktopFly's 1,045-neuron MaleCNS locomotor circuit, keeping its 17,224 measured connections fixed. An artificial encoder turns software status into sensory currents. A reward-trained action policy is distilled into a linear readout of the circuit's simulated spikes.

This page reruns the circuit locally, chooses actions from the trained readout, applies a previously recorded Codex patch, and evaluates the arithmetic tests in your browser. The fly's typing is animation. The large brain is measured anatomical context, not a claim that the whole brain is running.

—trained workflow success
11%random-action baseline
0%disconnected frozen readout

Synthetic workflow trials: 500 episodes per training seed, three seeds. The simulator supplies patches with an 82% repair probability. These are workflow results, not coding-benchmark scores. A handwritten workflow performs comparably; this does not establish a biological or native-wiring advantage. The three recorded code repairs are separate evidence.