Meta publishes Brain2Qwerty v2 research and training code

Meta introduced Brain2Qwerty v2 on June 29, reporting real-time sentence decoding from non-invasive brain measurements while participants actively typed. The research announcement describes experiments involving nine volunteers and approximately 22,000 sentences.

Code release accompanies the research

The system uses magnetoencephalography, or MEG, and incorporates a language model adapted to neural data. Meta released training code for both versions of Brain2Qwerty. Its research partner BCBL also released the dataset from the first version; the announcement should not be read as a release of the second version’s dataset.

The typing condition is central to interpreting the result. This is a constrained experimental task with specialized recording equipment, rather than evidence of unrestricted thought decoding or a finished communication product.

Evaluate the whole communication loop

Our analysis: a useful reproduction should report performance per participant as well as an aggregate. With a small study group, a headline average gives little guidance about the experience of an additional user.

Sentence accuracy also leaves product questions unanswered. How would someone correct a wrong word, cancel an unintended output or signal that the system should stop? Those interaction costs belong in an evaluation alongside decoding speed and accuracy.

The contribution of the language model needs similarly careful treatment. A plausible sentence can be linguistically strong while differing from the intended content. Evaluation should therefore preserve the original target and examine systematic substitutions, not just whether outputs read naturally.

For engineers, the released code creates an opportunity to inspect the experimental pipeline. Moving beyond that setting would require evidence about new users, recording conditions and communication tasks. None of those steps should be silently inferred from a successful demonstration during active typing.