Brain AI scales with data, not size – the opposite of ChatGPT

What this is about

For years, language models like ChatGPT followed a simple rule of thumb: more parameters, more compute, better model. These so-called scaling laws describe how performance rises smoothly with model size, data volume and compute. They're how the AI industry justifies its billion-dollar investments. Now research groups and companies are building the same kind of models for the brain. These neuro foundation models are large base models pretrained on recordings of brain activity. Afterwards they're supposed to predict, for example, how neurons respond to a video.

Several studies from this year show a different pattern: bigger models bring little or nothing here, while more recordings do. On November 11, the US site TechTimes put it bluntly: more compute doesn't help, only more data does. The same day, the trade magazine The Transmitter reported on a debate about the future of these models. Both reports announce a face-off: Bing Wen Brunton (University of Washington) and Patrick Mineault (Amaranth Foundation) are due to appear on November 15 at the annual meeting of the Society for Neuroscience in Washington, which runs from November 14 to 18. A point-counterpoint piece by the two is also coming out in the Journal of Neuroscience. We couldn't find the session in the public conference program ourselves; the information comes from the two reports.

We put the studies side by side. Our thesis: the bottleneck for brain AI is data. That explains why so much money is flowing into brain scans right now. And it explains why makers of sleep headbands consider their customers' recordings so valuable.

Finding 1: OmniMouse hits a wall at 80 million parameters

The cleanest measurement comes from Andreas Tolias's lab (Stanford, Enigma Project). The OmniMouse model was presented at the AI conference ICLR 2026. It learns from calcium imaging, in which active neurons light up under the microscope. The recordings come from the visual cortex of 73 mice and cover 323 sessions and more than 150 billion "neural tokens" – data points comparable to the text tokens of a language model. The number of recorded neurons varies by version: the arXiv paper says 2.3 million, the ICLR abstract 3.1 million. The model predicts responses to movies, reads behavior such as pupil size and running speed from brain activity, and forecasts future activity.

The team tested sizes from 1 to 300 million parameters. The metric is the per-trial correlation between prediction and actual measurement (1 would be perfect):

Beyond about 80 million parameters, the paper says, gains were minimal. The loss curves saturated or the model started memorizing data. With more sessions, on the other hand, every model got better – the large ones even more so. The authors conclude that brain models are "data-limited, not compute- or parameter-limited." They consider sudden new capabilities with even more data possible, but they haven't shown it.

Finding 2: With BrainLM, the bigger model does worse

A preprint from May 29, 2026 (Marraffini and colleagues) tests three brain foundation models for fMRI. This method uses blood oxygen levels to measure which brain regions are currently working. The models examined were BrainLM, Brain-JEPA and BrainMass. BrainLM was pretrained on around 6,700 hours of fMRI from the UK Biobank and the Human Connectome Project (HCP). The models had to predict individual people's cognitive performance from the scans. The comparison was a plain regression on functional connectivity – the information about which brain regions fluctuate in sync. It gets by with around 80,000 parameters and no GPU.

The result (correlation): on the HCP dataset, BrainLM-111M reached 0.050 using its internal features, while the 650M version managed just 0.003. On the Dutch AOMIC dataset it was 0.126 versus 0.065. According to the table, the simple connectivity regression scored 0.39 and 0.31 respectively. No foundation model beat it.

Important for our thesis: the authors don't blame a lack of data but the pretraining objective. The models learn the loudest parts of the signal and in the process destroy subtler statistical structures linked to cognitive performance. CortexMAE from the MedARC group (ICML 2026) paints a similar picture: data reconstruction improves along a clean power law. But when it comes to predicting personal traits, no model beats the simple connectivity baseline.

The counterevidence – and what it actually measures

Some studies do report scaling gains. Read closely, though, the picture shifts:

Our take: where growth with model size looks impressive, what's usually being measured is pretraining loss – how well the model fills in masked parts of a recording. On tasks with practical value, such as seizure detection or trait prediction, the curves flatten quickly. So the counterevidence doesn't refute the thesis; it refines it.

Predicting isn't understanding

Daniel Yamins (Stanford) cautions, according to The Transmitter, that better prediction doesn't make a model's inner layers more brain-like. Brunton argues that many different models can explain the same recordings equally well. A transformer that perfectly predicts falling apples still couldn't design a rocket. As an example, she cites an experiment in which a worm's nervous system was hooked up to a simulated fly body: the fly walked – but just as well with a random network. John Krakauer (Johns Hopkins) speaks of "a second brain you first have to understand." Even Tolias warns that models "can hallucinate too." Their value, he says, lies in testable discoveries. One example is his lab's prediction that dilated pupils make visual neurons more sensitive to UV light, which was later confirmed in animals.

Mineault thinks the criticism is premature. Language models, too, only showed new capabilities beyond a certain size. Grace Lindsay (NYU) indirectly supports the data thesis:

"None of the performance we're seeing is the peak performance a model could achieve. More data should help."

Why everyone is collecting brain data right now

A 2025 Meta study (Banville and colleagues) analyzed 8 datasets with 84 people and 498 hours. Decoding images from brain signals grew log-linearly with recording time, with no plateau. What mattered were the hours per person. Doubling the number of participants (from 24 to 48) brought no significant gain. Meta's TRIBE v2 model also keeps improving with more training data "without a plateau," according to the paper. It was trained on 451.6 hours from just 25 people. The frequently cited 720 people include the test data, around 1,118 hours in total.

That fits the new large-scale projects. Meta is funding the Digital Brain Project with $15 million; it aims to collect 15,000 hours of brain activity recorded while people play video games. It's coordinated by the Fondation Rothschild in Paris together with the Université de Montréal. According to The Transmitter, nine teams have been funded so far, including Harvard/MGH, Columbia and Georgia Tech. TechTimes writes that the money goes to Stanford. According to a corrected version of the The Transmitter article and the project website, that's wrong. Stanford runs its own Digital Brain Project: around 14,000 hours of 7-tesla fMRI and wearable magnetoencephalography, which measures the magnetic fields of brain currents, from about 100 people over five years. Yamins, who is involved, says:

"Even though our goal is a model, our main activity is collecting data."

Brunton counters: "What's the evidence that this works, other than that we hope it does?"

The link to sleep headbands

This makes clear why our report on EEG headbands is no side issue. With EEG, electrodes on the scalp measure the electrical voltages of brain activity. According to an investigation by Science, consumer neurotech companies have collected tens of millions of hours of such recordings. Nine companies say publicly that they're building brain foundation models. Interaxon (Muse) says it has used 50,000 hours of customer data for training and, according to a summary of the investigation, calls brain data "the most valuable part of the company." Muse says it has collected over a billion minutes, which works out to around 16 million hours. Piramidal has more than 4 million hours from open datasets and clinics.

For comparison: even Meta's ambitious project is aiming for 15,000 hours. Our take: if data is the bottleneck, companies with customer recordings have an enormous incentive to use them – especially overnight recordings, which deliver many hours per person. But the research also shows that hours alone aren't enough. Consumer devices have one to six electrodes, clinical ones 19 to 25. And nobody knows what the person was seeing or doing at the time. Experts interviewed by Science doubt that a useful model can be built from this.

The bigger picture

The evidence is remarkably consistent: with brain models, more parameters stop paying off early, while more recordings keep paying off. That's the opposite of the language model story, where text is cheap and available in bulk. TechTimes points out that an hour of brain recording costs orders of magnitude more than text. Progress here costs scanner time, not compute time.

Still, caution is warranted. Many of these papers are preprints without peer review. They measure different things with different methods, and an hour of fMRI isn't comparable to an hour of headband EEG. The BrainLM finding mainly argues against size, not automatically for more data, since the authors blame the training objective. It's also unclear how much data is enough. According to The Transmitter, not even Stanford can say. Whether Mineault turns out to be right about "jumps beyond a certain amount" will only become clear with projects like Meta's 15,000 hours. Until then, one thing stands: whoever owns brain data owns the scarcest raw material in this research. That makes it all the more important that headband users know what their nights are being used for.