Meta showed Brain2Qwerty v2 - it decodes sentences in real time straight from brain activity recorded non-invasively. 61% accuracy on average, 78% for the best participant. The code is open to everyone.
- Meta decodes writing from brain activity WITHOUT surgery: MEG recording + deep learning, in real time (June 29).
- 61% word accuracy on average and 78% for the best participant, trained on about 22,000 sentences from 9 volunteers.
- The code is open; the data comes out through partner BCBL. The limit: the MEG machine is lab equipment.
A second version of Brain2Qwerty has come out of Meta AI Research - a system that decodes what you're writing from your brain activity alone, recorded with MEG (magnetoencephalography). No surgery, no implants. Nine volunteers wear an MEG device for 10 hours while typing, and the model learns straight from the raw signals.
This is one of the quiet-but-tectonic stories. Surgical brain interfaces have been producing impressive results for years. Surgery, though, keeps them niche. A non-invasive method with this kind of accuracy opens the door to communication for people who lost it - no scalpel.
The brain-machine interface is slowly moving from science fiction to an engineering problem. It's still a lab - the MEG machine weighs tons. But once the code and the data are out, the next step is no longer made by one team. It's made by hundreds.