Feber · Architecture of Intelligence · 2026
How long does it take a human being to learn to tie a shoelace? The act looks trivial and is in fact highly complex. This instrument shows that complexity across eleven systems of the body, all working together — operating both sequentially and simultaneously. What runs is not a single model being fitted but a dynamic relationship between many systems and sub-systems: an electrochemical, bio-neurological process, and a form of computation.
Seven of these are dynamic systems that take turns leading; four are persistent (substrate) systems that run throughout. The substrate systems are not static — they vary too, but within tight bounds — and each system runs on its own time base.
The framework we are using here is active inference — a theory in computational neuroscience (Karl Friston and colleagues) in which the brain is a prediction machine: it continuously predicts what the world and the body will do next, and perceives, acts and learns so as to close the gap between the prediction and what actually arrives. It works in priors. A prior is an expectation the body has built from experience — a prediction about what will happen, weighted by how reliable it has proved. The intention to tie the bow is the top-level prior under which all eleven systems are organised: the frame of intention, or apex prior. (This is one of a set of such instruments; like each of them, it is written to be read on its own.)
This act has two timescales. The learning unfolds across five to seven years — the human being's period of deep learning. The doing takes twenty or thirty seconds. The learning that stands behind the act is on the order of five to eight million times longer than the act itself. But this is not a model mastered and then applied: learning and doing are not separate phases — every performance goes on adapting, tuning and re-remembering the systems it draws on.
Active inference is an error-correcting loop, and it never leaves the systems as it found them. Each enactment reinforces, adapts and improves the modelling. Shoelaces show this plainly: there are many knots and many kinds of lace — round, flat, waxed, slippery — and each is fine-tuned in the doing. (Ian Fieggen — “Professor Shoelace” documents twenty-five ways to tie the knot and over a hundred ways to lace a shoe.) The capability is general; we examine it here through an ordinary act we take for granted.
The eleven systems
Seven dynamic systems lead in turn; four persistent systems run as substrate. Each operates on a different time base within the apex prior — the frame of intention that is tying the bow.
1 · Diachronic development — five to seven years of deep learning
The seven dynamic systems named above are developed here through the coupling between the organism and its environment — and that environment is highly variable. Each line is one system's prior growing from shallow (leaning on live sensation) to deep (running on stored structure). The four substrate systems of the body run beneath as the dashed baselines. The shaded band around a line is the range the system can still vary within at that age — its room for variation: wide while the system is still being learned, narrowing as it settles. It stays wide only because of the error the system keeps meeting at its edge — and that error is what drives the learning. Ticks below mark rehearsal-experiments: the child's attempts, each returning information the system learns from. Untying precedes tying.
The capability this produces is transferable — portable and flexible, re-deploying to new shoes, new knots, and other tasks entirely. Transferable intelligence of this kind is a distinguishing mark of the human, setting it apart from every other form.
This settled state is not a fixed program but metastable: a basin the systems can enter when the frame of intention is active — stable enough to run the act, shallow enough to be left and re-entered. It is softly assembled afresh each time from the present state of the eleven systems and the actual shoe and lace, and reconsolidated — re-tuned — whenever it is enacted. The skill is never replayed; it is re-enacted. The loop is recursive: the model builds the act, the act re-tunes the model, and that revised model builds the next.
2 · Synchronic act — spending the priors (~20 seconds)
The knot-tying sequence — four phases, the same eleven systems re-weighted. Each curve is now a coupling coefficient — and what is actually changing beneath it is each system's threshold of activation: lowered when the system leads, raised when it runs in the background. Tying is bimanually asymmetric, so bilateral integration / midline crossing becomes a lead system, peaking at the Wrap where the two hands work differently across the midline. The endpoint is Tightening: a proprioceptive stop-condition, tension read to tightness. A canalised respiratory micro-hold appears at the Wrap — the tying analogue of swallow apnea.
3 · Learning and use on one timeline
Both timescales in a single static view. On the left, the five-to-seven-year development — the priors assembled through the organism's coupling with its environment — rising to the gate, where the assembly becomes action-ready and metastable. On the right, use across the lifespan: a long series of brief enactments (~20–30 s each). Each one recalls the assembly and reconsolidates it — re-tuning it a little. The loop is recursive: the model builds the act, the act re-tunes the model. The coupling between organism and environment runs throughout — and that coupling is what we are calling intelligence.
In the human this coupling has distinctive features. Our intelligence is not solitary: other human beings are themselves part of the field we learn, remember, experiment and think within — a collective intelligence. We extend it through symbolic systems and through intergenerational learning, so that what one generation assembles can be handed to the next. The result is an intelligence code that is flexible, transferable, adaptable, and continuously active — continuous in a literal sense: of the systems recruited for the shoelace-tying apex prior, most are either in action or quiescent and primed for action at all times, never switched off, always ready to be recalled.
Coda — reality without an object
A note for the psychoanalytic reader. The account above is a model-building account: learning is active inference, and reality is something the organism continuously infers. It is set deliberately against the object-relations account of how reality is acquired, as we can see it in the work of Donald Winnicott.
In The Use of an Object and Relating Through Identifications — given to the New York Psychoanalytic Society on 12 November 1968, first published as ‘The Use of an Object’ in the International Journal of Psycho-Analysis (1969, 50, 711–716) and reprinted in Playing and Reality (London: Tavistock, 1971) — the sequence is well known: the subject relates to the object, destroys it in unconscious fantasy, the object survives, and only a surviving object can be used; destruction is what places the object in external reality. This rests on Freud's reality principle: the psyche's move from the pleasure principle to an acceptance of an external reality that must be tested and accommodated. Winnicott takes that principle as given, and deepens it — for him destruction makes the object's externality felt — but he still arrives at an object that was “there all the time,” “a thing in itself.”
Our objection is to the reification. The habit runs well beyond that one paper: across his seminal work — the transitional object (1953), the true and false self (1960), the maturational processes and the facilitating environment (1965) — relational and developmental processes are repeatedly given the standing of entities with substance and agency. Object-relations theory, and the reality principle beneath it, treat “the object” and “external reality” as pre-given things the subject comes to recognise. But on the model-building account there is no object handed to the system, and no reality to cross over into. There is a single system minimising prediction error at its sensory boundary; “external reality” is only ever the best explanation the model maintains for what arrives there — continuously fitted, never inspected directly. What Winnicott calls destruction-and-survival is, without the drama, the world returning error the model did not predict; what he calls the object is a stable, high-precision inference, not a thing found. Even Freud's reality principle is reification: it posits a reality the psyche bumps into, where here reality is always already built. Reality is not encountered. It is constructed, and reconstructed, every time.
A word on destruction. Winnicott's mechanism for developing a stable psyche is inherently destructive — a quantum of aggression that, by attacking the object, creates its externality. He is explicit that it is “the destructive drive that creates the quality of externality,” and that he means “the actual impulse to destroy.”
This is troubling because it imports drive theory, and it dramatises a continuous process as a staged, violent event — attack. The model-building account doesn't use this invention. The work Winnicott assigns to ‘destruction’ is done via prediction modelling and error correction — not a force or an impulse but an informational computational process. Where he writes that the object survives destruction, we would say the prediction failed to absorb the error, the error persisted, the model updated.
What replaces the reified object is not nothing — it is a process. Intelligence, on this view, is not the recognition of a pre-given reality but the continuous, developmental, recursive coupling between organism and environment, through which the system actively and creatively builds and rebuilds its model — and with it a recursive self, made and remade in the modelling, never a thing found.
This does not deny the clinical phenomena Winnicott describes — the patient who can only feed on the self, who must come to find the analyst as real. However, it places the process in the narrow band where it belongs, as a subset of cognition and perceiving.
Illustrative model — Feber, Architecture of Intelligence (2026). Active inference: Friston · precision weighting: Feldman and Friston (2010) · developmental motor sequence after standard fine-motor milestone literature. Coupling values, maturation ages, and credible intervals are illustrative: no study has measured per-channel precision across eleven systems during either developmental assembly or a naturalistic tying act. The two rate-limiters (bilateral integration, finger isolation) set the gate; their wide early bands are inter-individual spread in experiment count, not the calendar.
Space play/pause · R reset · ← → scrub the active timeline · click either panel to scrub it.