Algorithmic Structure of Experience and the Unfolding Argument

Two complementary moves in one paper: a philosophical defense of why brain architecture type matters for consciousness, plus a geometric framework for how brains change over time — with psychedelics as the test case.

The unfolding argument is a clever attack on theories like Integrated Information Theory and Kolmogorov Theory (KT). It goes like this: any recurrent neural network (RNN) can in principle be "unfolded" into a feedforward network (FNN) that produces identical input-output behavior. If that's true, then the causal structure of a network can't be what determines consciousness — both architectures would generate the same reports. The first half of this paper pushes back hard. The key move is distinguishing between "can approximate one specific function" and "can access the same space of possible functions." Under realistic resource constraints, an RNN (which is Turing-complete — it can compute anything a universal computer can) has access to an astronomically richer repertoire of compressible models than an FNN (which is equivalent to a lookup table or primitive recursive function). The paper puts a number on this: for 1000-character reports, roughly 10^367 distinct outputs are possible — no physically realizable FNN could enumerate them. So the two architectures are not equivalent in any practical sense, and computational hierarchy (the Chomsky ladder from finite-state machines up to Turing machines) is an invariant of what is being computed, not of which wires happen to implement it.

KT's claim is that structured conscious experience tracks the ability to run highly compressive predictive models. The paper's corollary is clean: agents forced into FNN-class computation can only access simpler, less compressive models, so they have systematically lower potential for rich experience — not because recurrence is magic, but because the most parsimonious models of a complex world tend to require Turing-complete expressivity. The paper also proposes "first-person science" as an empirical path forward: directly perturb a subject's computational structure (via drugs, brain stimulation, meditation) and ask them what changed. The subject-as-researcher sidesteps the measurement problem that plagues third-person consciousness science.

The second half introduces neural geometrodynamics. The core idea is that brain connectivity is not a fixed backdrop but a dynamical geometric object — analogous to spacetime curvature in general relativity. Connectivity shapes the landscape of possible neural trajectories (fast timescale, seconds); neural activity reshapes connectivity via plasticity (slow timescale, minutes to hours); and the plasticity rules themselves adapt over longer periods via metaplasticity (ultraslow timescale, days to years). This three-layer hierarchy is written as three coupled differential equations, cleanly separating what the brain is doing now, how it is learning, and how it is learning to learn.

Psychedelics become a worked example. Psilocybin and LSD act as serotonin agonists on pyramidal neurons, producing a state-independent deformation of the connectivity landscape — flattening the deep valleys (pathological attractors like depressive rumination loops) so the brain can escape them. This is the acute phase. What follows is a post-acute window of enhanced Hebbian plasticity, during which new activity patterns can consolidate into new, healthier attractors. The therapeutic implication is direct: psychotherapy or behavioral intervention during this window should be unusually effective at reshaping the landscape permanently. The paper frames this geometrically as a wormhole — a temporary topological shortcut connecting otherwise inaccessible regions of phase space. The analogy earns its keep here because it captures both the transience and the structural nature of the effect.