Meta introduced Brain2Qwerty v2, an AI system that translates brain activity into text using non-invasive magnetoencephalography (MEG) scans. Trained on 22,000 sentences from nine volunteers, the model achieved 61% average word accuracy, a significant leap from the ~8% of previous non-invasive methods. Meta released the training code for v1 and v2, while research partners provided the v1 dataset. The system uses end-to-end deep learning and fine-tunes large language models on neural data to improve accuracy. Meta aims to help people with communication loss due to brain lesions and is part of its Digital Brain Project, which includes a $5 million fund for open neuroscience datasets. The paper in Nature Neuroscience highlights that high-performing brain-computer interfaces typically require surgery, but Brain2Qwerty v2 approaches surgical-level accuracy non-invasively. This advances the growing field of non-invasive brain-computer interfaces, alongside competitors like Neuralink and AlterEgo.

Key facts

• Brain2Qwerty v2 achieves 61% average word accuracy, vs ~8% for prior non-invasive methods.
• Trained on 22,000 sentences from 9 volunteers using non-invasive MEG scans.
• Meta released training code for both v1 and v2; dataset for v1 also released.
• Part of Meta's Digital Brain Project with $5M fund for open neuroscience datasets.
• Approaches accuracy of invasive brain-computer interfaces without surgery.