Decoders that generalize: the case for brain foundation models.

The neural decoding results that made headlines were, under the hood, language-modeling results. In August 2023, two landmark papers appeared in the same issue of Nature: a Stanford-led team decoded attempted speech at 62 words per minute from intracortical arrays, and a UCSF-led team drove text, synthesized voice, and a talking avatar at 78 words per minute from a cortical surface array. Both crossed the conversational-usefulness threshold the same way: recurrent networks trained on neural features, constrained by an n-gram language model — the decoder guesses what cortex is doing, and a language prior cleans it up.

The obvious next question is the one the rest of machine learning already answered: why train a bespoke model per patient, per array, per task? The field's answer is emerging under the name neural population foundation models. The clearest early statement is POYO, a transformer framework for neural population decoding that tokenizes individual spikes and trains across recordings from different sessions, subjects, and even electrode geometries — then transfers to a new session with minimal fine-tuning. One model, many brains: the pretrain-then-adapt recipe, applied to cortex.

Scaling pressure is arriving from the non-invasive side too. Meta's Brain2Qwerty showed that a deep network over MEG signals can decode full sentences as a participant types — well below implanted-array accuracy, but trained on a data modality you can collect at scale from healthy volunteers. The lesson generalizes: whoever assembles the largest, most diverse corpus of neural recordings gets the best prior over brain activity, and that prior lifts every downstream decoder, invasive or not.

This is the thesis we find most compelling in the field today. Session-to-session drift, the classic failure mode of clinical BCIs, becomes a domain-adaptation problem a pretrained model absorbs rather than a recalibration burden the patient carries. And it is why neural data governance is a capability question, not just a compliance one: the corpus is the model. Language models learned language from text; the decoders that matter next will learn the brain from the brain.

Update, November 9, 2026: newer evidence tempers the "one model, many brains" half of this thesis. In October 2026, Neuralink reported pretraining decoders on more than 50,000 hours of unlabeled intracortical data, which kept them working for weeks without recalibration and set an 11.32 bits-per-second cursor record. But those results came from models pretrained on a single participant's data, and the company wrote that it has "yet to improve performance online: decoders built from multi-participant models perform no better than their single-participant counterparts." It did report one successful transfer of a model from one participant to another. Pretraining is paying off within a brain; the gain from pooling across brains is not yet shown. The post is a company update, not a peer-reviewed paper.