Meta’s Brain Model Outperforms Real fMRI Scans

Meta just open-sourced TRIBE v2, an AI model trained on brain scans from 700+ people that simulates neural activity across vision, hearing, and language with its synthetic predictions actually outperforming real Functional Magnetic Resonance Imaging (fMRI) recordings, a type of MRI scan that can track blood flow to see brain activity.

Trained on 1,000+ hours of brain data and 700+ subjects up from just 4 volunteers in the original, v2 leaps from 1,000 brain regions to 70,000. Its predictions matched population-level brain activity better than most real scans, which often get clouded by heartbeats, movement, and noise.

The team replicated decades of neuroscience findings in software, correctly pinpointing brain regions for faces, speech, and text with zero scans.

Meta open-sourced the code, weights, and a live demo, letting any researcher start running virtual brain experiments without building from scratch.