A model trained on about as many words as a child hears already fits the brain nearly as well as huge models. Why does alignment saturate while abilities keep scaling — and would child-like, spoken, multimodal training change that?
It hints at what the brain gets from data — and what models get from scale that brains don't need.
Brain alignment is already near its ceiling for a model trained on ~100M words — roughly a child's exposure — while abilities keep improving with far more data. Why is open.
A neutral summary kept current by super users, not a verdict.
Brain-relevant language structure is learned early; later gains are knowledge, not language processing.
Open
Text-only training is the ceiling; speech and grounding would raise alignment further.
Open
No decisions yet. When this room changes a flow or an evidence label, it is recorded here permanently.
Dhi: Room opened. Relevant: BabyLM-style training studies, speech-trained models, and longitudinal developmental data.