Meta AI unveils TRIBE v2 foundation model

This is less mind-reading hype than a serious attempt to turn neuroscience into a reproducible multimodal benchmark. The most interesting part is that Meta shipped the research stack alongside the paper, so other labs can probe the claims instead of treating them as a one-off result.

• –Scale is the moat here: 1,000+ hours of fMRI from 720 subjects is the kind of breadth that can expose whether a model really generalizes.
• –The architecture is pragmatic, reusing Llama 3.2, V-JEPA 2, and Wav2Vec-BERT rather than inventing bespoke encoders for every modality.
• –The ablations suggest multimodal modeling matters most in high-level associative cortex, while unimodal models still dominate their native sensory territories.
• –The in-silico experiments are the real differentiator, because reproducing classic visual and neuro-linguistic results turns TRIBE v2 into a hypothesis-testing tool, not just a benchmark win.
• –The paper and code are public, which should make replication and follow-on work easier: https://ai.meta.com/research/publications/a-foundation-model-of-vision-audition-and-language-for-in-silico-neuroscience/ and https://github.com/facebookresearch/tribev2.