Illuminating how brains construct inner worlds, and converting that into strategic blueprints for an artificial model of the same system.
The brain does not receive the world. It hypothesises the world, continuously, and uses sensation to correct the hypothesis.
The anatomy is blunt about this. Roughly 5% of the excitatory synapses in V1's input layer come from the thalamus; only about 7% of the synapses on LGN relay cells come from the retina — the visual "gateway" receives an order of magnitude more input from cortex than from the eye. Sensory input is a minority shareholder in the activity of sensory cortex. It does not build the percept; it votes on a percept the cortex is already proposing.
What follows from that is the central discovery: perception, imagination, memory, planning and dreaming are one machine in different couplings. A generative model that runs whether or not the world is there. Tethered tightly to sensory correction, we call it perception. Run free, we call it imagination or dreaming. The difference is a precision parameter, not a different mechanism — which is why turning that parameter causes a healthy brain to hallucinate on demand, and why hallucination, delusion, phantom limbs, and the sense of owning a rubber hand all become mechanically explicable rather than merely catalogued.
An inner world, then, has ten properties: it is generative, hierarchical in space and time, endogenously driven, precision-weighted, body-anchored, action-closed, decouplable, partly globally broadcast, rebuilt offline, and self-modelling.
Building an artificial version means implementing those ten properties — and not implementing the biology that merely happens to surround them.
| Document | What it gives you | |
|---|---|---|
| I | 01-mechanisms.md |
The biology, in eleven sections: the inference problem, precision-weighted hierarchies, laminar and dendritic circuits, thalamus as router, the hippocampal-entorhinal simulator, interoception and the self, action as causal probing, global availability, development and neuromodulation, and the clinical failure modes. Flags what is solid, what is framework, and what is contested. |
| II | 02-invariants.md |
The translation layer. Twelve load-bearing commitments to keep, nine biological contingencies to discard, and five genuinely open questions to decide deliberately rather than inherit. |
| III | 03-blueprints.md |
CIW-1: twelve modules, seven typed buses, the objective functions, and the per-tick control loop written out step by step. Then a three-tier build programme — Foundry, Keel, Cathedral — with pre-registrable falsifiable targets, and the conditions that would falsify the whole approach. |
| IV | 04-validation-battery.md |
Forty tests in seven suites, organised around dissociations rather than effects. Includes a six-test minimum viable battery. |
| V | 05-risks-and-ethics.md |
Risk register, the two integrity failures this architecture predicts of itself, and pre-specified capability tripwires for the moral-status question. |
What must be preserved for an artificial system to construct an inner world by the same means:
| Invariant | Omit it and | |
|---|---|---|
| A1 | Generative-first, not recognition-first | no imagination, no rollout planning, no graceful off-distribution failure |
| A2 | Segregated prediction and error channels | attention, sensory attenuation and self/world discrimination all become unimplementable |
| A3 | Explicit dynamic precision on every signal | all evidence weighted equally; cannot attend, arbitrate, or decide when to revise |
| A4 | Explicit multi-timescale hierarchy | every level chases the fastest signal; slow structure never isolated |
| A5 | Factorised structure ⊗ content binding | every environment learned from scratch |
| A6 | Two-store memory with offline distillation | catastrophic interference |
| A7 | Body/viability anchor generating endogenous valence | all objectives exogenous and gameable |
| A8 | Action-in-the-loop causal probing | correlational structure mistaken for causal structure |
| A9 | Scheduled decoupled simulation | the system can only react |
| A10 | Capacity-limited global broadcast | modules stay siloed, or connect all-to-all and never cohere |
| A11 | Global neuromodulatory control bus | every hyperparameter frozen at design time |
| A12 | Self-model: body, agency, attention schema, calibrated confidence | cannot distinguish self- from world-caused change; confabulates fluently |
Discard: spikes (unless energy efficiency is itself the goal) · cell-type taxonomy · absolute latencies · circadian scheduling · neuron and synapse counts as targets · biological pain.
Keep anyway, counter-intuitively: a hard binding resource budget. Scarcity is what produces abstraction — remove all resource pressure and the system has no reason to compress rather than memorise, or to attend rather than process everything.
1. The most implementable idea in the field is structure–content factorisation. Entorhinal cortex carries transferable relational structure; hippocampus binds it to sensory particulars. Learn the grammar of an environment once, reuse it wherever that grammar holds. It has a working formalism, clean behavioural predictions, and it is the difference between an agent needing a million episodes and one needing a handful.
2. The sharpest validation test is a dissociation, not an effect. Attenuate the top-down/apical channel: an architecture doing precision-weighted inference will show its integration measure collapse while feedforward classification survives. Generic breakage does not produce that pattern. If you cannot make your system hallucinate by raising top-down gain, it is not implementing this — whatever the loss curve says.
3. Confabulation is structural, and biology warns us about it explicitly. Split a brain's hemispheres and ask the speaking one to explain behaviour it did not cause: it produces a fluent, confident, false answer, with no sense of fabricating. Its confidence is unrelated to its access. This is the same failure as confident hallucination in current language models, and the lesson is that you do not fix it by making the report module better at reporting. You fix it by training the report head on calibration and never on fluency — or you accept that it lies.
Foundational documents complete. Nothing implemented.
The recommended first move is the instrument track from 03 §5 — build the
measurement suite before the thing being measured. Every brain-inspired programme that drifted into
folklore did so because it had narratives instead of measurements.