For practitioners: Progress in non-invasive brain-to-text decoding changes the data and modeling trade-offs for BCI research, increasing the importance of large annotated neural-language corpora and robust transfer techniques. Meta published a research paper and engineering blog describing Brain2Qwerty v2, a non-invasive pipeline that decodes intended typed sentences from magnetoencephalography (MEG) recordings. According to Meta's paper and blog, the team trained on about 22,000 sentences collected from nine volunteers who each spent roughly 10 hours wearing an MEG device; the system uses end-to-end deep learning with fine-tuned large language models and agent-driven configuration search. Meta reports an average 61% word accuracy across participant-specific models (best single participant: 78%), versus about 8% for prior non-invasive approaches, and frames this as approaching accuracies previously seen only with invasive recordings (per Meta research and the company blog).
Brain2Qwerty v2 – Meta's non-invasive brain-computer interface