DiegoVallejo

The Fly Brain Is Not Open Source. The Harness Is.

Diego Vallejo • • 9 min read EN

systemsagentsopen-sourceneuroscience

Any sufficiently mapped nervous system will be forced to play Doom.

Adult Drosophila melanogaster, side view
The animal. Not the runtime. Drosophila melanogaster, André Karwath / Wikimedia (CC BY-SA 2.5).

On September 3, Google, HHMI Janelia, Cambridge, and the MRC Laboratory of Molecular Biology published the complete central nervous system of an adult male fruit fly. About 166,700 neurons. About 125 million synapses. Eleven thousand six hundred and ninety-one cell types. Licensed CC-BY. The dataset had been sitting on the public bucket since June. The paper made it official.

By September 6, Alex Wormuth had the same graph playing Doom.

That is the story. Not “scientists map a brain.” Not “the fly is an agent.” The map left the lab and immediately became a socket.

What actually shipped

MaleCNS v1.0 is not the 2024 FlyWire female brain. That map was ~139,000 neurons and stopped at the head. This one is the male adult, brain plus ventral nerve cord — the insect spinal cord — proofread, annotated, and downloadable from Janelia. Berg et al., Cell, September 3, 2026.

The important sentence in the release is not the neuron count. It is the license. CC-BY. You can take the weights of a dead animal, drop pairs under three synapses, assign sign from predicted neurotransmitter, and run leaky integrate-and-fire units on a laptop GPU. Several wrappers did exactly that in the first week. MIT on the glue. CC-BY on the brain. A token contract on top of both, if you wanted the circus.

Three layers. Only the first one is the scientific object.

THREE LAYERS, ONE SENTENCE1 CONNECTOMEMaleCNS v1.0 · 166,700 neurons · ~125M synapses · CC-BY · neuPrint male-cns:v1.02 SIMULATORLIF cells · FlyGym / MuJoCo · game adapters · MIT glue · this layer can lie3 PERSONAwallet · ticker · “the fly decided” · theater · do not peer-review this
Close-up of a fruit fly head and proboscis
Head sensors first. The MaleCNS map adds the ventral nerve cord — the cable to the body. Sanjay Acharya / Wikimedia (CC BY-SA 4.0).

Viren Jain and Michał Januszewski, Google Research. Official walkthrough of MaleCNS and the zebrafish map, 3 Sep 2026.

Official viewer. This is the file, not the demo:

Open Neuroglancer full screen · dataset gs://flyem-male-cns/v1.0/ · Janelia FlyEM.

The week the harness arrived

Wormuth’s loop is the clean description of everything that followed:

WORMUTH LOOP · THE HARNESSFRAMEDoom pixels→SENSORSR1–R6 / R8→GRAPHstatic LIF→MOTORWASD + fire→PPL101damageThe graph does not train. The named pin is the paper.
  • Each Doom frame stimulates sensory neurons.
  • Neural activity is mapped to game controls.
  • Damage hits two PPL101 dopamine cells as “reinforcement.”
  • The question at the end of the post: will the fly learn to survive?

Three days later the same harness had a clip that traveled:

Jessica Paquette pointed the same graph at Super Mario 64. Mario ran into a wall, repeatedly.

Someone else wired looming expansion into LC4 / LPLC2 and let the Giant Fiber fire a jump in Chrome’s T-Rex. Beat Saber came next. Then a fenced browser. Then a wallet on Robinhood Chain signing $FLYBRAIN.

Tom’s Hardware reached for Clarke and almost got it right. The amendment is not “any sufficiently advanced technology.” It is any sufficiently instrumented graph. Once the pins have names, the demo writes itself.

None of these systems trained the connectome. The synapses are frozen at the moment the specimen went under the electron microscope. What you watch in the livestream is the harness: which cells you poke, which descending neurons you read as a button, what you decided to call reward.

A runtime is not a mind

The honest wrappers say this out loud. flybrain.dev: browsing is a fair word. Deciding is not. The fly has no goals. A person still holds the key. Live launches stay off unless a human turns them on.

That is the part the token pages skip.

A male fly already has courtship-tracking neurons that steer toward a target. Wire those to an opponent in a fighting game and the “strategy” is 64 percent chase. That number is not intelligence. It is a cable that existed in the animal. Same trick as stigmergy in a multi-agent system: the environment does the work you will later attribute to the agent.

Astra, this week, is the trained version of the same confusion. A model that clicks, types, and greens the test runner gets called a coworker. A connectome that jumps when DNp01 crosses threshold gets called a player. In both cases the interesting engineering is the I/O map, not the story you tell about the middle.

What is actually open

Open source, here, means three different claims stacked in the same sentence.

  1. The connectome. Measured. CC-BY. neuPrint dataset male-cns:v1.0. This is the rare case where “open” is not a metaphor.
  2. The simulator. LIF parameters, neurotransmitter signs, body in FlyGym / MuJoCo, browser fence, game adapters. MIT or equivalent. This is software. It can lie.
  3. The persona. A wallet, a ticker, a live page titled “the fly decided.” This is theater. Treat it as theater.

If you collapse 1 and 3 you get the worst kind of demo: a video of spikes next to a price chart. If you keep 1 and study 2, you get something engineers can use. A runtime that forces you to name the pin.

That is the part worth stealing.

The pin is the product

Agent stacks spend years hiding the interface behind prompts, skills files, and a 3,000-word AGENTS.md. The fly stack cannot. You do not tell MaleCNS to “be careful in the repo.” You stimulate a column in the optic lobe and you read a descending neuron. If the mapping is wrong, the animal walks into the wall. Mario already did.

A connectome runtime is a hostile interface in the useful sense. It will not complete your sentence. It will not invent a utility module. It will not apologize. It will integrate current until a cell spikes, or it will not. That is closer to a kernel than to a chat model, and it is a better teacher for anyone building agents this month.

The evaluation lesson is the same one as flaky agent evals. Doom-with-a-fly is not a benchmark of the fly. It is a benchmark of the adapter. Change PPL101 to some other pair of cells and you have a different paper. Leave the graph fixed and swap the body, and you have a different animal. Publish the harness or the clip is content.

What not to write next

This is not a post about consciousness. It is not a review of the token. It is not “the fly is AGI with 166k parameters.” A static wiring diagram of one male Drosophila does not plan, does not regret, and does not get better at Doom because you watched it for an afternoon.

It is a public graph with named ports. People did what people do with public graphs that have named ports. They plugged in a game.

The sentence that survives October is not “we uploaded a fly.” It is: once the nervous system is a file, the file is not the product. The product is whatever you solder to it, and that part was never peer reviewed.