The calibration tax
Every decoder you built in Volumes I and II — a velocity Kalman cursor, a Riemannian EEG classifier, a deep speech neuroprosthesis — was trained essentially from scratch for one person, and usually re-fit at the start of every session. That per-user, per-day cost is the field's quiet tax: it burns clinical time, exhausts patients, and caps how far neural decoding can scale. Foundation models are the frontier's bet that this tax can be amortized — paid once, across a large population, then reused cheaply.
What a foundation model promises
A neural foundation model is a large network pretrained with self-supervised objectives on a broad corpus of neural recordings, then adapted — by fine-tuning, a light readout head, or even zero-shot — to many downstream tasks. The value is not one clever architecture; it is the factorization of the learning problem into an expensive, once-off representation stage and a cheap, per-deployment task stage.
Stage 1: learn a representation f by minimizing a label-free self-supervised loss over a large corpus. Stage 2: freeze (or lightly adapt) f, and fit only a small task head g on scarce labels.
The economic argument is compelling: labels (intended movement, spoken words, mood ratings) are what cost clinical effort, while unlabelled neural activity is comparatively abundant. If the representation carries most of the structure, the label-hungry stage shrinks.
Why a brain is not a corpus of text
The analogy to language and vision foundation models is seductive but leaks in three places. First, there is no shared vocabulary or coordinate frame: electrode i in subject A samples different neurons than electrode i in subject B, and there is no canonical alphabet of neural symbols. Second, neural data is profoundly non-stationary — the distribution shifts within a day and across signal drift, violating the near-i.i.d. assumption text enjoys. Third, the total quantity of high-quality neural data is minute next to the trillions of tokens behind a large language model — the small-data regime.
A map of the frontier
The frontier is best read by modality. For spikes and intracortical activity, Neural Data Transformers and spike-tokenizing models (POYO) pretrain across sessions. For scalp EEG, masked models such as LaBraM and BENDR pretrain on thousands of hours. For intracranial EEG, BrainBERT learns self-supervised spectrogram representations; for fMRI, BrainLM is a masked-autoencoder foundation model. A separate lineage builds encoding foundation models — neural digital twins that predict how cortex responds to any stimulus. All share the same transformer-for-neural-data backbone idea.
Demonstrated versus aspirational
Be scrupulous about the ledger. What is demonstrated: multi-session and multi-subject pretraining measurably reduces the labelled data a new session needs, and encoding twins generalize to held-out stimuli. What is emerging: genuine cross-subject transfer and multimodal neural+language models. What remains aspiration: a single brain-scale model that zero-shots onto an arbitrary new person the way a language model zero-shots a new document. Do not let the last be quoted as the first.