Google Research reported about the completion of work that neuroscientists have been waiting for decades. For the first time, the brain and central nervous system of an adult male fruit fly have been completely mapped. The model is called MaleCNS v1.0.
By the first weekend after the announcement, the dataset was in the hands of the community, and already at the beginning of the next week, enthusiasts demonstrated how Drosophila connectomes are taught to play Doom and Super Mario 64. Arthur C. Clarke’s Third Law seems to require an amendment along the lines of “any sufficiently advanced technology will immediately be forced to play Doom.”
Fruit flies have served as laboratory test subjects for more than a century and remain a key model organism in genetics and neuroscience. To map the nervous system, Google Research worked with the Janelia Research Campus (HHMI) and the scientific community.
According to Google, AI was used to “combine millions of 2D images into 3D neural structures, reconstructing a record 166,000+ neurons.” For comparison, the human brain is estimated to have about 86 billion neurons, so humans are still a long way from a complete digital copy, and data centers will need much more memory.
Google called the result a major achievement in neuroscience that will accelerate our understanding of how the brain works. Practical application was found faster than expected.
Software engineer Alex Wormuth built a connection between the emulator Doom and simulation of the nervous system of a fly. Each frame of the game turns into stimulation of sensory neurons, and the activity of the neural network is reflected in the control of the character.
The training scheme is especially interesting. Receiving damage activates a pair of dopamine cells, PPL101, which in the real Drosophila brain are responsible for reinforcement and the formation of associative memory.
Every frame of Doom stimulates sensory neurons. Neural activity is mapped to game controls. Damage causes a stimulus for two PPL101 dopamine cells to act as a reinforcer. Will the fly learn to survive?
The experiment code is completely open, and the learning process of 166,700 neurons can be observed live on a special page.
The results are still far from passing the level. In the published clip, Mario runs up and crashes into the wall over and over again, which looks more like expected behavior to the fly’s intelligence.
Puckett admitted that her version was “literally 100% vibe-coding with GPT Astra… just for fun.” The sources of this work are also publicly available, and the developer promises to talk about the progress of training on social networks.
