Pruning Mental Models through Sleep and Psychedelics

Within the Free energy principle framework, the brain is understood as a prediction machine that minimizes variational free energy — a quantity that decomposes into an accuracy term and a complexity term. Accuracy rewards models that fit incoming sensory data; complexity penalizes models with unnecessary parameters. During waking life, both terms are in play. Sleep, however, removes the accuracy constraint entirely: with sensory channels quieted, the brain can minimize complexity alone, pruning synaptic weights that encode redundant or overfitted structure. This is the computational rationale for sleep — not rest, but model compression. Dreaming represents the phenomenology of that compression: the rehearsal and consolidation of causal structure as the generative model is restructured.

Psychedelics operate on the complementary term. Compounds like psilocybin are understood to attenuate the precision weighting assigned to high-level prior beliefs — the compressed, iconic representations that normally dominate perception. When the precision of abstract priors is reduced, the relative precision of low-level sensory prediction errors increases, and perception shifts toward its constituent elements rather than its categorical summaries. The 'microphone' dissolves into surfaces, reflections, and textures. At the apex of the representational hierarchy sits the self-model — the most abstract and compressed prior of all. Dissolving its precision produces the spectrum of non-dual states, depersonalization, and Ego dissolution that characterize high-dose psychedelic experience, and that parallel the phenomenology reported in advanced contemplative practice.