New AI Mimics Your Brain Waves Using 18x Less Computing Power

MEG (magnetoencephalography) is a brain scanner that listens to the tiny magnetic fields your brain produces. Unlike an MRI, which gives you a still picture, MEG captures brain activity a thousand times a second — close to real time. Hospitals and labs are collecting mountains of this data, which makes it a perfect target for the same trick that built ChatGPT: train one big AI on tons of examples, then reuse it for many jobs.

The new part is the engine under the hood. Earlier brain-AI models used a "transformer" — the same design behind ChatGPT — which is powerful but power-hungry. This team swapped in a different design called Mamba, which handles long streams of data more efficiently. The result: their model produced more realistic brain signals than the older version while training in 22 hours of computer time instead of 400, and it could look at 4 seconds of brain activity at once instead of a third of a second.

The model can also be taught new tricks cheaply. With a light adjustment technique (LoRA — think of it as a small add-on patch rather than rebuilding the whole engine), it learned to generate brain responses to specific tasks it had never seen during training. In plain terms, it can imagine what a healthy brain would do in a given situation — useful for testing theories, filling in missing data, or creating realistic synthetic brain data without exposing real patients' scans.

Here's the catch: this is a research paper, not a product in a clinic. The AI generates activity that looks like real brain data, but it doesn't diagnose anything, and nobody has shown it improves patient care yet. It also still needs huge amounts of real MEG recordings to learn from in the first place. Think of it as a faster, cheaper research tool — promising groundwork, years from your doctor's office.