Universality and transfer limits

The honest counterweight to the field's central metaphor. The promise of a neural foundation model rests on there being a shared 'language of neural activity' that generalizes across subjects, areas and tasks; the open question is how universal that language really is, and where transfer breaks. Several failure modes are documented: negative transfer, where pretraining on mismatched data actively hurts the target task; out-of-distribution collapse across species, brain areas, or task structures the model never saw; and brittleness when the recording modality itself changes.

The current, defensible reading is that transfer is real but bounded. It is strongest within a modality, species and task family, and weakest across them, with a large gap between within-distribution success and true cross-domain generality. A separate limit is identity: per-unit and per-subject embeddings that make transfer work also encode individuating structure, raising re-identification and privacy concerns. A truly universal neural foundation model — one that reads any brain doing anything with little calibration — remains aspirational, and honest reporting should say how far from that ideal a given result sits.