Summary
Nick Saraev demonstrates how to train a simulated fruit fly brain — the publicly released connectome from Google DeepMind containing 166,000 neurons and over 125 million connections — to automatically classify and respond to emails. Using GPT-6 Astra in high mode as a supervising model, Saraev fed 864 labeled email examples across six categories (sponsorships, consulting, student support, and others) into the connectome simulation and applied reward-based training to map neural activity patterns to response labels.
The resulting system achieves roughly 80% accuracy on his email set, which Saraev acknowledges underperforms even a small conventional neural network — but that is not really the point. The project illustrates that publicly available biological connectome data can now be downloaded, simulated, and trained for practical tasks using general-purpose AI models as the supervising layer, with the motor and classification circuits kept deliberately separate to avoid interference.
Saraev also addresses the ethical dimension of running a simulated biological brain, describing how he provided the virtual fruit fly with a simulated environment, food, and companions as a precaution. The video blends practical automation demonstration with a broader provocation about where connectomics and machine learning are heading as mouse, rat, and eventually human brain maps come online.
📺 Source: Nick Saraev · Published September 13, 2026
🏷️ Format: Hands On Build
