Brain AI Models Break the Rule That Built ChatGPT: More Computing Power Doesn't Help, Only More Data Does
Reported by Tech Times1 source ↓
A mouse-brain modeling study found that adding model parameters did not improve prediction of neural activity, while supplying more neural recording data did.
Key points
- The OmniMouse study found performance improved with more neural recordings, not a larger model
- The study used recordings from 3.1 million neurons in 73 mice
- A separate BrainLM evaluation found its larger model performed worse than its smaller version
- Researchers disagree over whether predictive brain models can explain how the brain works
Details
The OmniMouse study, presented at ICLR 2026, tested whether scaling up AI models would improve predictions of brain activity. Tech Times reports that the researchers trained models on recordings from 3.1 million neurons across 73 mice and found that performance improved as they added data, but not as they added parameters beyond a modest model size.
Tech Times also reports that a 2026 evaluation of BrainLM found its 650-million-parameter version performed worse at predicting cognitive performance than its 111-million-parameter version. Both were outperformed by a linear model based on functional connectivity.
The findings appear amid a broader debate over what brain models can show. In a Journal of Neuroscience Dual Perspectives paper, University of Washington researcher Bing Wen Brunton and Amaranth Foundation researcher Patrick Mineault argued opposing positions on whether these models can explain brain function; the article says they were due to debate the issue at the Society for Neuroscience meeting in Washington, D.C.
Written from coverage as of