Meta's Brain2Qwerty v2: they've built spellcheck for your motor cortex

[smfadmin (OP)]

Meta's Brain2Qwerty v2: they've built spellcheck for your motor cortex

July 2026

Posted because it's the closest thing to reading minds we've got, and it's still nowhere near close

Meta dropped v2 of their non-invasive brain-to-text system this week (28-30 June). Worth a look if you've ever wondered what "reading thoughts" actually means in practice — spoiler: not much, yet.

What it does
Sticks you in a $2 million shielded MEG machine (magnetoencephalography — reads the magnetic field your neurons throw off, no drilling required) and decodes what you're typing, in real time, at the sentence level. Not character-by-character guessing like the old stuff — full semantic reconstruction, because they've bolted an LLM onto the back end that uses language context to fill in the noisy gaps. Think predictive text, but the "text" you're predicting is coming out of someone's skull.

Numbers

• 61% average word accuracy across the cohort
• 78% peak for the best subject, with most sentences at one error or fewer
• Scales log-linearly with training data — i.e. it's not hitting a wall, more data keeps helping

The catch
It's not reading thought. It's reading the neural signature of physically typing a memorized sentence. You have to sit dead still in a machine the size of a small room. This is motor-cortex pattern matching with a language model doing cleanup, not telepathy. Lab-only, miles from anything you'd wear.

How it stacks up against the invasive stuff
Non-invasive is still getting smoked on raw bandwidth, and for the obvious reason — the skull attenuates the signal, that's physics, not a software problem. Old EEG spellers managed 2-5 words a minute despite decent accuracy, because there's a hard bandwidth ceiling on anything reading through bone.

Implanted systems don't have that problem:

• Neuralink's first patient: ~40 WPM typing/cursor control. A "120 WPM" second-patient claim is doing the rounds but the source is dubious, wouldn't cite it as fact.
• Speech-BCI implants (electrodes straight into speech-motor cortex, decoding attempted speech): approaching 60 WPM.
• Paradromics' Connexus: 200+ bits/second, 11ms latency. Non-invasive EEG is stuck around 1 bit/second for comparison. That's not a gap, that's a different sport.

Brain2Qwerty's actual contribution isn't better signal — it's squeezing more usable output out of the same crap bandwidth by having an LLM do inference instead of raw decoding. Clever, but it's polishing a much smaller pie.

Why I bothered posting this
Meta's open-sourced the training code for both v1 and v2, and their research partner released the v1 dataset. Long-game target is assistive comms for stroke/ALS/locked-in patients who've lost speech entirely — which is a genuinely good use of the tech, even if the current form factor is a small building.

Also: if you've ever wanted a concrete example of "the bottleneck isn't the AI, it's the sensor," this is it.