Teach a real connectome a task
The wiring stays exactly as it was mapped. A small readout learns to turn the circuit's activity into movement, and you watch the learning curve, the neurons and the body while it happens. Then you train a rewired control and find out whether the real wiring made a difference.
Teach a real fly circuit to drive
1,325 neurons from the FlyWire connectome, fixed. Train the readout from 20 descending neurons until the car stays on the track.
Teach the worm to find food
The complete 302 neuron worm connectome, fixed. Train a readout from 14 interneurons to turn towards the smell of food.
The fly plays poker
Kuhn poker, the smallest real poker, played by 1,846 FlyWire neurons and a trained readout of 21 output neurons.
The fly plays Pong
A ball bounces towards the fly. Its collision detectors see which side the ball is on; a trained readout moves the paddle.
Odour T-maze with a delay
Smell odour A, turn left; smell odour B, turn right. Then the same after a pause, which needs memory.
Two flowers: learn which one pays
One flower gives sugar 80% of the time, the other 20%, and they swap halfway. Chance is 50%.
Hold a course in gusty flight
Gusts knock the flying fly off course. Its motion sensing HS cells and a bar ahead guide a trained steering readout.
Chase a moving target
Follow another fly around an arena, as a courting male does. Two readouts steer and set the walking speed.
The method, in plain words
Every task follows the recipe used by the fly connectome demos that went viral in 2026, such as Fly Dino: the connectome is a fixed circuit, senses excite fixed input neurons, and a linear readout of chosen output neurons is trained with the cross entropy method. Each generation tries a population of readouts, keeps the best quarter, and samples the next generation around them.
We add what those demos usually leave out: held out test episodes that are never trained on, a hand written readout from known biology as a reference, and three controls (rewired, random and silenced circuits) you can train with one click and overlay on the same chart.