Google's Complete Fruit Fly Brain Map Becomes Playground for Doom and Mario 64
Researchers have successfully mapped all 166,000+ neurons in an adult male fruit fly's brain and nervous system, and software engineers have already begun training the neural model to play classic video games.
The scientific community has relied on fruit flies as research subjects for more than a century, particularly in the study of genetics. These insects remain valuable to contemporary researchers working in genetics, neuroscience, and related fields throughout the 2020s.
https://twitter.com/cantworkitout/status/2095553014715093022
A watershed moment has arrived: researchers have now completed a full mapping of the brain and central nervous system belonging to an adult male fruit fly. Collaborating on this achievement were scientists from Google Research, HHMI Janelia Research Campus, and the broader scientific community. According to Google's announcement, artificial intelligence played a crucial role in the process, being deployed "to combine millions of 2D images into 3D neural shapes, reconstructing a record-breaking 166,000+ neurons." For perspective, the human brain contains an estimated 86 billion neurons, making this fruit fly connectome substantially smaller but no less significant.
Google characterized the breakthrough as something that would "accelerate our understanding of the brain, and is a major milestone in neuroscience," in a statement released last Thursday. Within days, however, the mapped neural structure had been repurposed for an entirely different purpose: playing video games.
https://twitter.com/cantworkitout/status/2096409780961059119
Software engineers have wasted no time putting the fruit fly brain model to work. Engineer Alex Wormuth and developer Jessica Paquette, who describes herself as a 'C++ ragebaiteur,' have each undertaken projects training the neural model on different games.
Wormuth's project focuses on Doom. Describing the approach, Wormuth explains that "Each Doom frame stimulates sensory neurons. Neural activity is mapped to game controls. Damage triggers a stimulus to two PPL101 dopamine cells as reinforcement." Wormuth posed the central question: "Will the fly learn to survive?" The implementation is available as open source code, and observers can monitor the training progress of the 166,700 neurons in real time.
https://twitter.com/cantworkitout/status/2097004115826200898
Paquette's effort involves Super Mario 64. In a demonstration video, the neural model repeatedly attempts to move forward only to collide with a wall. Paquette developed the project using an unconventional method, noting that the code was "literally 100% vibe coded with GPT Astra… just for fun." The source code for this experiment is also publicly available, and Paquette plans to share updates on the training's progress through social media channels.
Source: Tom's Hardware