A fruit fly brain on a chip. All 158,262 neurons. Playing against your CPU, which is learning to read it.

Below is a biological computer running in your browser: the complete wiring map of a real fruit fly, every neuron and 3.1 million synapses, laid onto an electrode grid. The silicon side shows it a ball, listens to its descending neurons, rewards it with dopamine when it scores, and learns which neurons mean "up".

158,262neurons on the chip
3.1Msynapses, real wiring
0spikes this second
0training cycles

BIOLOGICAL SIDEwaking up, loading 9 MB of wiring

activity0.0
descending lean0.00
brain speed0 ms/s
Each pad is about 70 neurons. A pad lights when they fire. Left columns are the senses, the big middle blocks are the optic lobes and central brain, the bottom row is the descending and motor output.

SILICON SIDEreading the neurons

hits0
misses0
hit rate, last 20–
learning curve · hit rate over time
What the CPU learned: a weight per neuron group. Positive pushes the paddle up when that group fires, negative pushes it down. They start at zero and move on every hit and miss.

A dish brain, with a fly instead of a dish.

In 2022 a lab in Melbourne grew human neurons on an electrode chip and taught them Pong by feeding back predictable signals when they hit and noise when they missed. This page does the same experiment with the wiring map of a real fruit fly, which is the one brain science has mapped completely.

01 · STIMULATE

The ball becomes neurons firing

Where the ball is, above or below the paddle, drives the fly's real small-object detectors on that side. The ball rushing in drives its looming detectors, the same cells that make a real fly jump.

02 · LISTEN

The chip reads the output

The fly's descending neurons, the ones that carry commands from brain to body, are read by the silicon side. A weighted sum of their activity moves the paddle.

03 · LEARN, BOTH WAYS

Dopamine in, weights out

A hit fires the fly's reward dopamine neurons, a miss fires its punishment ones. At the same time the CPU adjusts its weights, learning which groups of neurons predict where the ball is. The curve shows it getting better.

What is real and what is not

  • Real: every neuron, every connection and its sign come from the BANC connectome, the complete map of one female fruit fly's nervous system. The sensory populations and the dopamine neurons stimulated here are the real annotated cells.
  • Real: the activity you see is computed live in your browser from that wiring. Nothing is a recording or an animation.
  • Ours: the neuron model is leaky integrate-and-fire with calibrated constants, not a biophysical copy. The learning on the silicon side is a simple reward rule written by us.
  • Not here: living cells. The Melbourne machine keeps real neurons alive in an incubator. This is the map, running as a calculation.
  • Not claimed: consciousness, feelings, or that the fly "wants" to win.

Questions

Is it really running the whole brain?

Yes. The page downloads the full wiring (9 MB compressed), and your browser steps all 158,262 neurons about a thousand times per second of brain time. On a slow phone the brain runs slower than real time, which the "brain speed" number shows.

Why is it bad at Pong at first?

Because nobody told the CPU which neurons mean up. It starts with every weight at zero, so the paddle drifts on the fly's own activity. Each hit or miss nudges the weights. Give it a minute.

Does the fly learn, or the CPU?

Both get a signal. The CPU changes its weights, which is where most of the improvement comes from. The fly gets dopamine through its real reward and punishment neurons, which changes what its brain does next, but there is no long-term synapse change in this model yet.

Can I control it?

No, and that is the point. The only inputs are the ball, the rewards and the fly's own background activity.