Brain2Qwerty

Meta FAIR's non-invasive brain-to-text line (v2 announced June 2026; v1 published in Nature Neuroscience): decodes typed sentences from MEG brain recordings by fine-tuning LLMs on neural data. v2 reaches 61% average word accuracy (78% best participant) versus ~8% for prior non-invasive approaches, trained on ~22K sentences from 9 participants. Full v1+v2 training code released. A milestone for non-invasive brain-computer interfaces and on-thesis neuroscience FM work.