Fly brain

The brain is Shiu et al.’s leaky integrate-and-fire model of the FlyWire connectome: 138,639 neurons wired as they are in a real fly. It was ported to PyTorch so many brains can run in one batch on an RTX 5090.

A thing in the world becomes a smell: a mix of the fly’s receptor channels, stimulated for a 100 ms sniff. The answer is read where the mushroom body hands off its verdict, as approach minus avoid output. Learning is dopamine, delivered through the connectome’s own dopamine-to-output pairs, which weakens the synapses from the Kenyon cells that were active.

Every brain runs next to an untrained twin fed the identical input spikes. With no lesson the two are identical, so any difference between them is learning and nothing else. That one design choice made most of what follows measurable.

In Quest for Glory, game state became stimulation of sensory neurons and the spikes of descending neurons became moves. A small reader trained on the frozen brain beat the same reader fed noise, 6.51 points per episode to 4.32 (83 and 63 episodes, p < 0.0001). The brain’s own learning made no difference to the score.

In Little Deep, a fish game Claude Code designed for AI players, fly-brained fish sniff in eight directions and learn from what they eat and what bites them. Then they started to fear food. Twenty bites from one predator taught a fresh brain to avoid almost every smell.

The cause was in the fly. Two channels, vinegar and acids, get 2.4 to 2.9 times their share of the input to the learning centre in the raw connectome, and the game had handed those channels to prey and predators alike. Shuffle the wiring while keeping each neuron’s number of partners, and the brain behaves the same; spread the inputs evenly and it can’t learn at all. Nature here is a distribution, not a blueprint.

The labs characterise the brain as given. What came back is a short list of rules:

• It detects coincidence. A pattern is whatever is on at the same instant. Two halves played back to back bind only 12% of what they bind together; order carries nothing.
• It keeps everything. Nothing fades without dopamine, and a mix reads as the sum of its trained parts.
• It ranks; it doesn’t measure. Absolute values drift with training load, rank order holds. Capacity depends on how separate the patterns are, and over-training lowers it.
• Most senses can’t learn. Smell can. Touch, taste and nearly all of vision are fixed reflexes that answer in milliseconds: looming means escape, sugar means feed.

The predictions that were wrong were the useful part. Dopamine teaches best about 75 ms before the smell, not at the same moment. The lateral horn, expected to hold innate likes and dislikes, turned out to be a map of which channel is which. A simple paper model of the write rules now forecasts a training run’s values to within 7% of their range before it runs.

The memory lab puts the whole memory on one screen: the 62,261 synapses between the Kenyon cells and the output neurons.

Ten meals of one smell weaken about 1,800 of them, 0.6% of the total. Reward and punishment write to different bands, and different smells to different columns, so three memories cost 1.4% and barely touch each other. Almost all of it lands in one lobe (γ).

Copy those weights into a fresh brain and it reads the same likes and dislikes, on the first try. Restore one smell’s private cells and that memory drops away while the others stay. Reversal is different: teaching the opposite doesn’t undo the first memory but writes a second band over it, and the value is the difference.

Simon turns the rules into a game: four pads, four smells, a meal for a match and a bite for a miss. Asked to combine the tap and the pad itself, the brain stalls around 55%: the mushroom body doesn’t bind two patterns into a third. Give each pairing its own smell up front and it goes from 20% to 85% in sixty taps.

Because the twin sees the same input, the 3D views show exactly what learning changed. At 85–95% right, about 220 of 138,639 neurons fire differently from the untrained brain, a net of a few dozen spikes, and roughly 1% of the sniff’s synaptic routes change. Almost all of it is routes falling silent on the avoid side. Learning, here, is wiring taken away.

Driving every visual projection type alone mapped the reflexes: looming detectors fire the escape neuron, and some types steer with a clean left-right sign. That was enough to fly. With a simple retina and body, brain-driven flies crossed straight, bent, narrowing and corkscrew tunnels without touching a wall; random walkers hit 22 to 44 times per 100 steps.

Up and down weren’t in the wiring the way left and right are, so pitch is a rule of ours, and the pages label it as one. And memory does reach the body, one way: a rewarded scene lowers the drive to the legs and wings by about 10% per 100 Hz of learned value, while punishment and the escape reflexes don’t move at all. A comfortable fly runs less.

In a 3D maze with seven junctions, a feeder at one end and biters at two others, 32 flies with their own memories barely beat chance after nineteen rounds (14% against 12.5%). One body earns a reward about once in seven trials, and a lesson only reaches the last branch it took.

Give the same 32 flies one shared memory and they learn the tree in four rounds: 80% over the last five, top junction first, because enough pooled rewards let the faint food smell at the top carry value. Take the smell gradient away and even the hive learns nothing but fear. A hive is this brain’s natural way to scale: the same 62,261 synapses, written by many bodies.

Hunger, satiety and mood in a real fly are mostly hormones, and a wiring diagram doesn’t carry them: driving seventeen state-cell groups left a learned smell where it was. The model’s own gain on the memory’s output works, but only downward. Turned up past normal, the outputs saturate and the learned difference vanishes. So in the pet terrarium a full fly hears its memory quietly and a starving one at full volume, never louder.

The pet puts it together: one brain in a dish, a trainer with treats, sprays, lamps and food, and a panel showing what it has come to like. Even its habits come out of the rules. A dopamine-flood item it keeps returning to stops writing new memory after a while, because the synapses it uses are already at their floor: tolerance, from a rule that can only weaken.