The episode argues that computational neuroscience has quietly inverted its old project: rather than decoding the brain on its own terms, researchers now treat frozen, pretrained AI networks as a fixed coordinate system and bend neural data to fit it—achieving striking results in fMRI mapping, spike-to-language decoding, and cross-subject alignment through simple contrastive matching rather than elaborate decoders. The non-obvious insight is that this success is also an admission of defeat: it works precisely because neural responses resist intuitive parameterization, so a ruler that measures meaning without explaining it now stands in tension with the slower mechanistic program of connectomes, conductances, and noise theories. The most revealing work sits where alignment and mechanism collide—where structure is derived from plasticity rather than borrowed from a black box.

Topics: foundation models, neural decoding, brain-AI alignment, connectomes, mechanistic interpretability