developers seeing this started to simulate the fly brain architecture on computer to train reinforcement-based models to achieve specific tasks; such as playing games, solving puzzles or doing actions. the system is simple: fly brain model is taken, fly’s neural pathways are linked with results of specific functions (e.g dopamine pathway is triggered when the fly captures the health orb in the game), and reinforcement learning steps are executed in order to “teach” the fly to “live” in the simulation setup that was built. people have been building very creative setups: from flies that parallel park to flies that writes computer programs, i’ve seen many interesting projects (i had built a doom simulation using older fafb v768 dataset by princeton before, although i used a frozen connectome and optimized doom in order to preserve real fly behavior inside doom instead of training a model on fly brain architecture to play doom).

in addition to the developers building those simulations, i’ve seen many non-developers arguing about the ethics of such experiments. the common point/argument was something close to “the connectome has a fly’s nervous system running with a ‘synthetic’ environment it interacts with; so how’s it different from using a real animal for those experiments?”. if we see the fly’s brain as “a carrier of a conscious mind”, the argument becomes acceptable on the grounds that a perfectly copied digital remake of the same structure would also carry such mind.

the question before this, however, would be “is the fly a conscious organism in the first place?”. in the scientific consensus, the answer would be yes; considering the 2012 Cambridge Declaration on Consciousness and more recent 2024 New York Declaration on Animal Consciousness. therefore, a real fly is a conscious organism, and if a simulated environment “perfectly” copies its brain, that copy could also be considered “conscious” & should be examined within the same standards.

even though it might make me “soulless”, reckless or rude; i’ll say that nobody gives a fuck about a fruit fly. sorry vegans. i dont.

the bigger, and much more concerning question here will open when (or if) somebody manages to model the human brain: what level of humanism we’ll approach with to a simulated human brain? and what separates a real human with its thoughts, feelings and rights from a simulated one? after all, the conscious human mind also consists of an extremely complex and fascinating biological network and it interacts with an external environment; not different from a potential synthetic brain model interacting with a synthetic environment perfectly copied. have fun.

The project is FlyDOOM. On 20 August, I took the FAFB fruit fly connectome—139,255 reconstructed neurons and roughly 2.7 million synapses—and fed DOOM into it through the fly’s visual pathway. I read the motor output back out as control signals.

A real biological wiring diagram was playing a 1993 shooter.

Nothing was trained. There was no reward function and no learning. The connections belonged to the fly, exactly as they had been measured, and they did not change. The interesting question was never whether a frozen fly brain could finish a level. It was what a real nervous system would do when dropped somewhere it had never evolved to be.

I published the repository and moved on. Four people on GitHub saw it.

being found

Then the connectome discussion took off online. A group I had never met went looking for prior work, found the repository, and decided the earlier version deserved credit. They built a funding mechanism around it and routed the proceeds to my GitHub account.

I only learned about it because I happened to check social media. I replied to posts, shared gameplay clips, and watched the displayed market cap of $FLYDOOM pass $65,000.

The money is not the interesting part. Recognition for a piece of research now travels through channels nobody designed for that purpose. It can cross research communities, open-source networks, social platforms, and on-chain markets, then reach you whether or not you were paying attention.

There is something strange and encouraging in that. Publishing unfinished work in public can create a trail. Sometimes that trail stays quiet. Sometimes strangers find it, preserve its provenance, and send a signal back.

finishing the work

I am treating this as a reason to finish FlyDOOM properly. I am rerunning the methodology on the HHMI Janelia dataset and writing up the paper.

The repository now includes the simulator, experiment suite, results, technical notes, and gameplay recordings. The results are more useful than a clean success story: motion computation appears in the visual system, but the original motor readout does not receive enough of it; touch reaches the steering pathway without preserving left-right direction; smell is the sensory path that behaves as the wiring predicts. Those failures expose exactly what a connectome provides—and what it does not.

If you have arXiv endorsement rights for cs.NE, I am looking for an endorsement. If you work anywhere near connectomics or neuromorphic computing, I would welcome the conversation.

explore the FlyDOOM repository · see the LinkedIn post · follow updates on X