How a model learns to read a brain

Strip away the jargon and a brain-to-text model is a translator. It reads a sequence of voltage measurements and writes the sentence the person was speaking or thinking. The hard part is that brain signal, unlike text, never repeats itself. Say the same word twice, one second apart, and the recordings look different: electrodes shift a millimeter, attention drifts, the room hums differently. A translator whose source language changes spelling mid-sentence needs to be very well read.

What the machine actually does

The work happens in four moves. First, slicing: the continuous signal from every channel is cut into short windows, a few milliseconds each, because words unfold in time and the model needs the timeline. Second, encoding: each window is compressed into a list of numbers, a feature vector, the way a camera turns light into pixels. Third, attention: a transformer network looks across all the windows at once and works out which moments belong together, the same machinery that lets a language model connect a pronoun to the name three sentences back. Fourth, output: the network emits word pieces, and a pretrained language model scores the candidates, cleaning up the result the way autocorrect wishes it could. This is not speculation; it is the published recipe behind Meta's Brain2Qwerty results, including the honest detail that they needed the language model in the loop to cross the finish line.

Training is just this, repeated millions of times: play the recording, read the guess, nudge every parameter toward less wrong.

The foundation-model move

The old approach built a small decoder per person, per session, per device, recalibrated in a lab ritual every time. It works, and it caps out. The foundation-model move is to train one large network across many brains, many sessions, many devices, so that somewhere in its middle layers it holds a representation of speech that no longer cares whose head the signal came from. Below that shared core, a few minutes of personal data adapts it to a new speaker. MANAS-1 demonstrated the pattern for EEG generally: sixty thousand hours of recording across setups, one four-hundred-million-parameter model that transfers. The same move on language-paired data is the open lane, and it is our Phase 2.

Why hours are the currency

Representation quality scales with diverse data, and diversity means accents, heads, days, moods, electrode placements. Every hour of paired recording buys tolerance for one more way a brain can be different from the last one. That is why we treat the dataset, not the architecture, as the company: the architecture is publishable, the hours are not. Whoever has the hours has the model that works on a stranger's head on a Tuesday morning, and nothing about that sentence is metaphor.