Bhautik Patel/Unsplash.com

The same week that researchers gathering in Washington, D.C., for the Society for Neuroscience annual meeting will debate whether artificial intelligence can genuinely explain how the brain works, a quieter result from a mouse-brain study may have already given them the answer.

A team training one of the largest neural AI models ever built on rodent brain data — 3.1 million neurons, 73 mice, 150 billion recorded neural signals — discovered something that runs directly against the central premise of modern AI: adding more parameters to the model did not improve its performance. Data helped. Model size did not.

In AI, that is essentially heresy. The AI scaling laws behind language models are built on a simple, repeatedly confirmed relationship — more parameters, more compute, better results. That relationship has held across language, images, and video. It is the economic engine behind hundreds of billions of dollars of AI investment. And for modeling the human brain, it may not apply.

That finding, from the OmniMouse ICLR 2026 study presented at the International Conference on Learning Representations, sits at the center of a formal scientific dispute published today in The Journal of Neuroscience — a "Dual Perspectives" paper in which University of Washington computational neuroscientist Bing Wen Brunton and Amaranth Foundation NeuroAI researcher Patrick Mineault argue directly opposing positions about what AI brain models are actually capable of. The two will take the debate live before an audience of thousands on Sunday, November 15, at SfN's Washington D.C. convention center.

The Field's Biggest Bet

Neuroscience is in the middle of an expensive wager. The same foundation model approach that transformed natural language processing — train a large neural network on massive amounts of unlabeled data, then fine-tune it for specific tasks — is now being applied to the brain's own neural activity.

The logic is intuitive: if a model trained on enough neural recordings can accurately predict how a new brain responds to a new stimulus, perhaps that model has learned something fundamental about how brains compute.

The investments reflect that logic. Meta's AI research division has donated $15 million to a Stanford University initiative targeting 15,000 hours of high-resolution functional MRI data to train what researchers are calling a "digital twin" of the human brain. Meta released its TRIBE v2 brain model in March 2026 — a system trained on more than 450 hours of fMRI recordings across 25 participants and subsequently evaluated against more than 1,100 hours of data from 720 participants. TRIBE v2 achieves zero-shot prediction: given a new stimulus a participant has never seen, it predicts how that person's brain will respond before recording a single neuron.

Meanwhile, a Tolias Lab paper published in Nature in April 2025 and developed at Baylor College of Medicine, Stanford, and the Allen Institute demonstrated a foundation model that improved predictive accuracy in higher-order visual brain regions by 25 to 46 percent over prior leading models. Training on recordings from multiple mice, the model generalized to new mice after just 30 minutes of additional data — and made a prediction about UV light sensitivity in the visual cortex that was subsequently confirmed in a live animal experiment. Andreas Tolias, the study's senior author, described that prediction as "precisely what usefulness looks like."

Where the Scaling Story Breaks

Then came OmniMouse.

Trained on 3.1 million neurons from 73 mice across 323 recording sessions, with more than 150 billion individual neural data points — a dataset orders of magnitude larger than any prior brain model — OmniMouse was designed to test whether the standard AI playbook would hold. Build bigger. Train on more. Reap the gains.

The data helped. The model size did not.

The OmniMouse ICLR authors, presenting at the International Conference on Learning Representations in 2026, found that performance scaled with how much neural recording data they provided, but adding more parameters to the model architecture produced no meaningful improvement after a relatively modest parameter count. In standard AI scaling, those two curves move together. For brain data, they diverge — and the parameter scaling curve goes flat.

That divergence is not just a modeling inconvenience. It suggests that neural activity data is organized differently from language or image data at some fundamental level. The statistical regularities that language models exploit — the patterns that reward adding more parameters — may not be present in the same form in the electrical signals neurons exchange. If that is correct, the foundational transfer of AI methodology to brain science may be hitting a wall that more compute cannot solve.

A 2026 evaluation of BrainLM, a foundation model trained on an extraordinary 6,700 hours of fMRI data from across the neuroscience literature, compounded the concern. When researchers tested BrainLM's ability to predict cognitive performance from brain activity, its larger 650-million-parameter version performed worse than its smaller 111-million-parameter version. Both were outperformed by a simple linear model using functional connectivity — a matrix that describes which brain regions tend to activate together — that required no deep learning at all.

Two Machines That Don't Understand Each Other

The core argument animating the Brunton side of Sunday's debate is philosophical as much as empirical: prediction and explanation are not the same thing.

A model that accurately predicts what your visual cortex will do when it sees a red apple has not told you why your visual cortex does that. It has told you that the relationship between the input and the output follows a pattern — the same thing any good regression model tells you. What it has not told you is the mechanism: the circuit-level computations, the cellular properties, the evolutionary constraints, the feedback loops that produce the pattern.

John Krakauer, who directs the Brain, Learning, Animation and Movement Laboratory at Johns Hopkins University, has described the situation with characteristic bluntness: foundation models give us "a second brain we have to understand" — which is to say, two brains and no explanations. Brown University's Ellie Pavlick put the problem as "two black boxes": the original brain, and the AI model trained to predict it, neither of which you can fully see inside.

Brunton's critique goes further. Her work on embodied intelligence argues that any model of the brain that does not account for the body the brain evolved to control is structurally incomplete. The brain is not a pattern-completion machine that happens to run inside a body — it is a control system whose representations are shaped by the physics of locomotion, the mechanics of grasping, the demands of social coordination. A model trained on neural recordings captured while a mouse watches videos cannot learn what the brain is for in the way that matters.

Her illustration is pointed: a lab that transplants a worm's entire connectome — every neuron and every connection, fully mapped — into a robot body still does not get a worm. The parts are correct. The behavior is not.

The Case for Optimism — and Its Limits

Patrick Mineault's counterargument is empirical and direct: the pessimists are holding the standard too high too early.

Foundation models in natural language processing were initially criticized for predicting the next word without understanding language. Those criticisms were not wrong — the models were not understanding in any robust sense — but the critics were also wrong about what the models would eventually be able to do. Emergent capabilities that seemed impossible at smaller scales appeared reliably as models grew and training data expanded. The prediction-understanding gap, Mineault argues, is a gap between the current state of a technology and its trajectory, not evidence of a fundamental ceiling.

He points to what brain models have already produced: the Tolias Lab's UV sensitivity prediction, confirmed in vivo, represents a genuine scientific discovery that emerged from a purely data-driven model, without any mechanistic theory of why the effect existed. The model found it first. The experiments confirmed it after.

NYU's Grace Lindsay, one of the field's more measured voices, has noted that current predictive accuracy figures — even the improved numbers from the Tolias Lab — are not the ceiling. The benchmarks the field uses, including the Brain-Score visual benchmark, measure how well models predict activity in specific brain regions, primarily the visual cortex's inferotemporal area. The fraction of neural variance explained by even the best current models leaves substantial room. Lindsay has characterized the current state as a beginning, not a verdict.

Stanford's Daniel Yamins, who has worked on both task-trained and data-driven approaches to neural modeling, argues that the two camps are not as opposed as the debate framing suggests. Task-trained models — systems that learn to perform specific tasks like object recognition, then get compared to brain recordings — and data-driven foundation models address different questions. Combined, they may be more powerful than either alone: the task-trained models supply mechanistic hypotheses; the foundation models supply the scale to test them.

What "Useful" Would Actually Look Like

The clinical stakes in this debate are not abstract. Neurostimulation therapies for Parkinson's disease, treatment-resistant depression, and epilepsy currently require extensive individual calibration — iterative adjustments that take weeks or months. A model that could accurately predict how a patient's brain responds to a stimulus could dramatically accelerate that process. The same logic applies to psychiatric drug development, where a model of how a compound affects neural dynamics across a population could replace years of clinical trial guesswork.

But Krakauer's "second brain" problem does not disappear when the application is clinical. A foundation model that predicts brain responses to DBS parameters would need to predict them accurately across the full range of a patient's movement states, emotional states, and environmental contexts — not just in the controlled recording session that generated its training data. Brunton's embodied intelligence argument applies with particular force to clinical translation: the brain being treated is attached to a person moving through a world, and recordings made in a lab may not sample enough of that variation to generalize reliably.

The OmniMouse scaling results suggest a path through, but it is expensive. If more data helps and more parameters do not, then progress requires an enormous ongoing commitment to neural recording — not building bigger models, but building richer datasets. That inverts the economics that made AI foundation models viable in language and vision, where parameter scaling is cheap and data collection is relatively fixed. For brain data, it is the other way around: a recording-hour costs orders of magnitude more than a text token.

Sunday's Question

The Society for Neuroscience meeting that opens Thursday is the largest annual gathering in brain science. When Brunton and Mineault face each other Sunday, they will be arguing in front of colleagues who have staked careers, grants, and institutional reputations on the foundation model bet. The OmniMouse and BrainLM results will be in the room.

The question is not whether AI brain models produce useful predictions — the evidence is clear that they do. The question is whether a field that has described itself as building toward a mechanistic understanding of the brain is actually building that, or whether it is building ever-more-precise pattern matchers that predict the brain the way a tide table predicts the ocean: accurately, usefully, and without any theory of what moves the water.

The scaling laws that broke in OmniMouse suggest the analogy may be closer than the field's most ambitious practitioners would like to admit. What the brain's information structure resists — the neat statistical regularities that reward bigger models — may be precisely what makes it a brain rather than a very complicated lookup table. If so, the path to genuine mechanistic understanding may require not more of what has already worked, but something genuinely different.

That is Brunton's argument. Mineault's is that the field has not been trying long enough to know. Both of them will be there Sunday to make the case in person.

The Brunton/Mineault Dual Perspectives paper is published today in The Journal of Neuroscience, Vol. 46, Issue 45. The Society for Neuroscience annual meeting runs November 14–18, 2026, at the Walter E. Washington Convention Center in Washington, D.C.