Fly Brain: a Telegram bot on a real connectome

The problem

The FlyWire team has mapped the entire brain of a fruit fly: 138,639 neurons and 54.5 million synapses. My goal was to make this science understandable to someone with no biology background, and to make it obvious that the answers come from a brain model, not from a chatbot making things up.

The bot itself speaks Ukrainian, so the screenshots below are in Ukrainian.

The solution

A Telegram bot where this wiring diagram works as a model. The user taps a button, for example “swat the fly with a newspaper” or “give it sugar”, and the bot:

1. runs a spiking simulation of the whole brain;
2. shows what the fly did: escaped, started feeding or cleaned its antennae;
3. sends a video with a schematic fly on the left and the signal running through its brain on the right;
4. adds a neuron report with a brain map and explains why it happened.

Code decides the verdict, not the LLM. What the fly did is determined by three output neurons validated in the Shiu et al. paper (Nature, 2024). The LLM only explains the result and cannot invent neurons.

“Break the fly”. The same stimulus goes to a normal brain and a modified one. Switch off 2 neurons and the fly stops feeding. Switch off 4 and it fails to take off, so the newspaper lands on it. Cause and effect are visible.

Three modes:

• “Guess what the fly does” runs with no LLM: you guess first, then the simulation runs and code checks your answer.
• “Break the fly” shows two flies side by side, before and after.
• “Experiments”: an LLM agent (OpenAI or Claude) with seven tools answers questions using simulation, path search in the synapse graph, comparison and images.

Speed. In the standard Brian2 simulator a whole-brain run takes minutes, which is too slow for a chat. I rewrote the model in numba with exact integration, and neurons that are guaranteed not to fire are not stepped at all. A test confirms the spikes match the full calculation exactly.

The result

Brain in the simulation	138,639 neurons, 54.5M synapses
Simulation	0.3–2.5 s (10 runs × 1 s of brain time)
Video rendering	3–4 s
Bot reply with video	5–17 s
Automated tests	86

The paper’s results are reproduced: sugar triggers the proboscis with a sigmoid dose-response, bitter suppresses feeding, and touching the antenna triggers antennal grooming.

Honest limitations

This is not an “uploaded mind” or a copy of a fly. It is a wiring diagram plus very simple neurons, with no neuromodulation, learning or body. The fly in the videos is a schematic drawing; the simulation decides what it does and when. The bot says so in every answer.

FlyWire data is licensed under CC BY-NC 4.0, so the project is non-commercial.