Cybernetics Canon — Modern Reread

A chapter-a-day walk through the cybernetics canon, restructured for modern reading without information loss. Source text register-shifted; defined terms, math, and key illustrations preserved. Scaffolding (numbered subsections, exercises, “we now turn to” connective tissue) stripped.

Method: Pick the structure that fits each chapter — chiasm where the argument inverts a frame, linear-cumulative for derivations, contrast for dialectical chapters, spiral for layered re-encounters. The structure should fall out of the source, not be imposed on it. See chiastic-rewrite skill.

Cadence: One chapter per day. Live status on chainlink issues tagged cybernetics-chapter.

Done

Powers, Clark & McFarland — Behavior: The Control of Perception (May 30). The 1960 two-part paper that seeded Perceptual Control Theory (and Powers’ 1973 book). The title is an inversion: behavior is not what the organism controls — perception is; behavior is what it does to hold perception at a goal. Build the loop from non-standard joints: Feedback Function F (f = F(vₑ), producing the perception), Comparator C (e = r − f), Output Function O (o = O(e)), Environment Function E (vₑ = E(o)). Negative feedback = the functions conspire so any error drives f back to r; the reference-signal is the only handle by which the loop can be steered. The hinge: the system controls f, not o — output varies wildly to cancel disturbance while perception stays nailed at reference, so “the system reacts only to the signals injected by its feedback function, and for any one system nothing else exists” — a control system cannot know what is going on out there, and its world-effects are incidental exhaust. The hierarchy: each order perceives an environment made of the lower order’s feedback-signals and acts only by setting their references — six orders (intensity, sensation, configuration, sequence, relationship, system) demonstrated on two people and a willingness to get shoved (the “Portable Demonstrator”). Reorganization (the N-system) is learning-by-selection: rate proportional to intrinsic error, new systems stick because they reduce error, not because anything instructs them; consciousness defined as a feedback function while the N-system acts on it. Closes by returning to the observer with the frame inverted: behavior is multiordinal, legible only if you reconstruct the controlled perception — the test of the controlled variable (perturb the world, watch what the system refuses to let you change) makes purpose empirical; and you cannot control a control system from outside — try and you become a disturbance it removes, which is why forcing behavior below the order where a conflict lives just relocates the symptom (“the paralyzed leg turns into a paralyzed arm”). Surprises: controlled-perception / “can’t know what’s out there” is in print in 1960, decades before predictive-processing vocabulary; the test of the controlled variable is a genuinely falsifiable method for reading purpose; control-only-from-within is derived as a theorem, not asserted as a value. Chiasm.
Pask — Conversation Theory Primer (May 29). Pask’s “Conversational Techniques in the Study and Practice of Education” (1976), with Pangaro’s Cybernetics and Conversation (1996) as opening philosophical frame. Conversation as the fundamental unit of cognition: not the individual but the dyad in interaction. The contrast: ordinary “understanding” is verbal agreement on a definition; Pask’s understanding is the ability to reconstruct the concept ab initio from related topics. A concept is not a stored description but an internal procedure that brings about and satisfies a topic; memory is reconstruction, not retrieval. The apparatus exists to make the stronger test mechanical — entailment mesh (directed graph of topics, with the load-bearing requirement of cyclicity: T derived from P,Q only counts if P,Q can also be derived from T, the “getting back” property that enables reconstruction), behaviour graphs (task structures for non-verbal model construction), modelling facility (mechanics bench, electricity bench), descriptors, and lamp-state machinery (Explore/Aim/Goal/Working/Understood) implemented in CASTE and INTUITION. The deepest move: when the tutor is replaced by a computer, learning still happens because the learner is already two participants — A₁ (operator, decides what to do) and A₂ (learner, tries to understand). The two-person tutorial is the externalisation of a one-brain architecture, not its source. TEACHBACK (tutor plays neutral student, learner explains AND explains the derivation of the explanation) is the cleanest test; works only ~90 minutes because the strain of forcing the duality to stay externalised exhausts the tutor. Surprises: concept-as-procedure and memory-as-reconstruction are in print in 1976, forty-eight years before the vocabulary that would let us recognise the architectural commitment; the two-participant requirement is mechanism not metaphor (the standard condition removes the second person; the dyadic structure persists internally); cyclicity is a structural test for whether a subject is understandable at all, and most curricula fail it. Contrast.
von Foerster — Ethics and Second-Order Cybernetics (May 28). Address to the Paris congress on family therapy, 1990. Opens by quoting Wittgenstein 6.421 — ethics cannot be articulated — and spends an hour honoring the constraint while still doing something useful. The argument: a brain is required to write a theory of a brain, so the cybernetician entering his own domain has to account for his own activity; cybernetics becomes second-order. The pivot is ethical, not technical. Outside the system you can say thou shalt — origin of moral codes. Inside you can only say I shall — origin of ethics. The two sisters who let ethics manifest without becoming explicit: Metaphysics (only the in-principle undecidable questions are the ones we can decide; the decidable ones are already decided by their framework; objectivity / hierarchies / Pontius Pilate are the standard dodges for the responsibility that freedom of choice forces; act so as to increase the number of choices) and Dialogics (language runs on two tracks — appearance/monologue/denotative vs function/dialogic/connotative — and Descartes should have written cogito ergo sumus; consciousness lives in appearance, conscience in function). Surprises: the Wittgenstein move is structural, not mystical — articulating ethics turns it into moralization by mechanism; the load-bearing line is a Marrakesh child’s the problem is not truth, the problem is trust; Foerster himself proposed the name cybernetics in 1949 because his 25-word English vocabulary couldn’t pronounce circular-causal-and-feedback-mechanisms-in-biological-and-social-systems. The interview adds zero-order cybernetics (acting without reflection), the causa-finalis frame (goal-directed action is computationally cheap, trajectory simulation is impossible), and the therapeutic imperative — if you want to be yourself, change. Contrast.
Beer — Designing Freedom (1973 Massey Lectures) (May 27). The short popular form of Beer’s management cybernetics, broadcast on CBC radio in the final weeks of the Allende government. Six lectures build cumulatively: (1) the real threat is institutional collapse, not pollution or famine, because relaxation time has overtaken perturbation time; (2) we have the tools (Ashby’s Law: V_R ≥ V_D) but install variety attenuators and amplifiers on the wrong side of the equation — the citizen is attenuated, the bureaucracy amplified, when it should be the reverse; (3) the “Liberty Machine” is a real-time information network that filters by significance (algedonic signals: pain/pleasure indices) and pushes decisions to the lowest competent recursion; (4) science applied to management is not bureaucratic measurement but model-driven steering — Beer’s six chapters of Brain of the Firm compressed to a paragraph; (5) the future demandable now is decentralisation through real-time computation, prefigured in Chile; (6) freedom is a computable function — the variety the regulated system can absorb without losing identity. Three contributions worth noting: autopoiesis is named on the air in 1973, before Maturana’s 1980 book lands in English; Beer states that problems exceeding the brain’s variety cannot be understood at all, ever — a principled epistemic ceiling; the Chile sections were recorded eleven days after Allende’s death (“they will give him a uniform of wooden pajamas”), making this not theory but witness statement. Linear-cumulative.
McCulloch & Pitts — A Logical Calculus of the Ideas Immanent in Nervous Activity (May 26). The paper that makes neural nets a mathematical object. Five assumptions (all-or-none firing, fixed threshold, single synaptic delay, vetoing inhibition, fixed wiring) turn each neuron into a proposition about its inputs one tick ago, and each net into a logical expression on past values — a temporal propositional expression (TPE). Feedforward nets are exactly TPEs in both directions (Theorems 1–2), with four primitives — delay, OR, AND, AND-NOT — as a basis. The equivalence theorems (4–6) absorb every plausible alternative neurophysiology (relative vs absolute inhibition, threshold extinction, temporal vs spatial summation) into the same calculus, protecting the model from future biological detail. Theorem 7 is the unexpected one: alterable synapses can be replaced by circles. Learning is recurrence. Nets with circles refer to indefinitely distant past; given a tape, scanners, and effectors they compute exactly what a Turing machine can — the natural model of the nervous system lands on the universal computing machine, in 1943, seven years after Turing. Closes with the epistemic move: psychons are no smaller than single-neuron firings; if the net is undefined, so are the facts; the “thing in itself” becomes a wiring question. Linear-cumulative.

Conant & Ashby — Every Good Regulator Must Be a Model (May 25). Sommerhoff’s five variables (Z outcomes, G good subset, R regulator events, S system events, D primary disturbers) set up the regulator-as-mapping picture. Cause control beats error control because cause control can drive H(Z) to zero; error control by construction cannot. The “model” problem — Chartres replica to Switzerland relief to Mercator to subway map, no natural boundary — gets sidestepped: define “model” for this proof, as a mapping h : S → R, and prove only that. The theorem: among regulators that minimize H(Z), the simplest is a deterministic map from system states to regulator actions. Proof engine is a one-line lemma about entropy and imbalance: if two regulator actions r₁, r₂ both have positive probability under p(R

s_j) and map to different outcomes, you can shift mass between them to lower H(Z) — contradiction with optimality, so all positive-probability actions for a given s_j must map to the same outcome. Crush ties, you get a mapping. Discussion: model-making goes from optional to compulsory; brain-as-regulator must be modelling its environment as a theorem, not as interpretation; “theoretical neurology” becomes available. Left open: how a regulator becomes a good one, and whether entropy-minimization is the same as correspondence to causal structure. Internal model principle (Francis-Wonham 1976) and “general agents need world models” (Richens et al. 2025) are the descendants. Linear-cumulative.

Wiener Cybernetics Ch 4 — Feedback and Oscillation (May 24). Opens with two ataxic patients — one missing proprioception (tabes dorsalis), one missing the proportioning that turns proprioception into motor command (cerebellar tremor). Both are broken feedback loops. From the patients to the signal tower (orders repeated back), the thermostat, Watt’s governor, and the defining property of negative feedback: it opposes the system’s own departure from the goal. Then the math: any linear time-invariant operator on the past of a signal is characterized by \(A(z) = \int_0^\infty a(\tau) e^{-z\tau}\,d\tau\); a feedback loop produces \(A/(1+\alpha A)\), which blows up when \(-1/\alpha\) lies inside the curve traced by \(A(iy)\). Stability becomes a geometry problem in the complex plane. Three first-order lags in cascade cannot be stabilized by a single feedback — which is why a ship’s gyrocompass autopilot needs two feedbacks (rudder-position loop inside, course loop outside), and why limb motion needs postural feedbacks underneath voluntary ones. Brief tour through relaxation oscillations (organ pipe, gas explosion), anticipatory feedback (duck shooting, AA fire control), and informative feedback (probing an icy road with small wheel impulses). Closes by widening the frame: homeostasis is the same architecture, slow timescale, non-myelinated nerves and chemical messengers, keeping body temperature, blood pH, osmotic pressure, heart rate, calcium balance inside narrow bounds. Feedback as visible motor control → feedback as the invisible substrate of being alive. Chiasm.
Wiener Cybernetics Ch 1 — Newtonian vs Bergsonian Time (May 23). Two sciences of the heavens: astronomy (palindromic Newtonian time, the music of the spheres reads the same backwards as forwards) and meteorology (directional, irreversible, statistical). Every science slides toward the meteorological pole — tidal evolution, biology, Maxwell–Boltzmann–Gibbs, Heisenberg’s statistical synthesis. Bergson named the time-asymmetry; the vitalism-mechanism wall got pushed outward to enclose both. Engineering tracks the philosophy: clocks (Huygens, Newton), then steam engines (Carnot, Joule), then communication and control (Gauss, Kelvin, Heaviside, radar). The 19th-century body-as-heat-engine view collapses — the body is not a conservative system; neurons work like vacuum tubes; the bookkeeping that matters is information, not energy. The modern automaton (receptors, effectors, central control, proprioception, memory, learning) exists in the same Bergsonian time as the living organism. Vitalism has won to the extent that even mechanisms correspond to its time-structure — but this victory is a complete defeat. The mechanist-vitalist controversy is now a badly posed question. Contrast.
Ashby Design for a Brain — Ultrastability in the Living Organism (May 22). The chapter pivots at §9/4 from plausibility (Stentor’s escalation staircase, Mowrer’s rat, the homeostat reversal showing nervous-system-like reorganisation) to necessity: a determinate system that changes its mode of behaviour must contain step-functions. From the necessity proof, four consequences fall out for free — training is feedback acting on step-functions; surgical compensation (Marina, Sperry) shows the cortex doesn’t need to know where the reversal lives; learning, memory, and goal-shift are angles on the same selection-of-terminal-field machinery; the gene-pattern needs only six items to install the whole apparatus, and the apparatus is reachable by ordinary natural selection. Contrast.
Ashby Design for a Brain — The Ultrastable System (Homeostat) (May 21). Step-functions plus fields generate a selective process: an ultrastable system rejects fields that lead to a critical state and retains those that don’t. Ashby builds the homeostat — four magnets in water troughs, uniselectors with Fisher-Yates random values, 390,625 combinations — to test the principle physically. The aileron analogy: ordinary stable systems persist in their wrong action; ultrastable systems change the field until it is stable. The closing test (a glass fibre joining two magnets) shows the machine adapting to a constraint that was never in its design spec. Linear-cumulative.
Ashby Ch 4 — The Machine With Input (May 20). Open: a crane controlled by a driver, parameters set from outside, deliberate switches. Hinge: feedback is dismissed as the central concept — “the exact definition of ‘feedback’ is nowhere important,” and with more than a handful of parts complex systems must be treated as a whole, not as an interlaced set of loops. Close: the input picture inverts. In very large systems the locus of history is local — self-locking ratchets (the closed oysters) and breeding properties (k > 1, autocatalysis, Darwinian evolution). No operator, no switch, no input. Chiasm.
Ashby Ch 6 — The Black Box (May 19). Open: an engineer faces a sealed box. Close: every object is a Black Box; “what’s inside” is the relation between observer and what is distinguished. Protocol → canonical representation → isomorphism (analogues) → homomorphism (models) → emergence as epistemic gap → memory as the trace of unobservable variables. Chiasm.
Ashby Ch 5 — Stability (May 19). The whole stack: equilibrium → cycle → stable set (closure) → stable under D → stable to a range of D → stable when coupled. The veto principle: a coupled whole rests only where every part can rest given what the others are doing. Homeostat falls out as a corollary. Positive feedback need not destabilise. Linear-cumulative.
Ashby Ch 12 — The Error-Controlled Regulator (May 19). Anticipatory regulator vs error-controlled regulator. The impossibility proof: the more R succeeds at holding E constant, the more it blocks the channel it needs. Continuity rescues the design; Markovian “hunt-and-stick” makes it tractable. Homeostat, telephone selector, blood-pH. Contrast.
Ashby Ch 11 — Requisite Variety (May 19). The Law: V_O ≥ V_D − V_R. Variety can destroy variety. The chapter sets aside regulation to watch a game, derives the bound, then reveals the game was regulation all along. R’s capacity as regulator cannot exceed R’s capacity as channel. Hitler’s control = 1 man-power. Chiasm.
Ashby Ch 3 — The Determinate Machine (May 19). A machine is anything behaving like a closed single-valued transformation. State, operand, trajectory, canonical representation. Phase space makes the algebra visible. The closing inversion: a “system” is not a thing in the world, it is the list of variables we chose. Contrast.
Ashby Ch 2 — Change (May 18). Operand, operator, transform, transition. Closure as the prerequisite for repeated application. Kinematic graphs reveal basins of attraction the algebra hides. Linear-cumulative.
Ashby Ch 7 — Quantity of Variety (May 17). Variety as the count of distinguishable possibilities. The chair as constraint. Every law of nature is a constraint. Chiasm.

Foundational

Wiener, Cybernetics (1948)
- Ch 1: Newtonian vs Bergsonian Time ✅
- Ch 4: Feedback and Oscillation ✅
Wiener, Human Use of Human Beings — see interactive walkthrough
Ashby, Introduction to Cybernetics (1956)
- Ch 1: What Is New
- Ch 2: Change ✅
- Ch 3: The Determinate Machine ✅
- Ch 4: The Machine With Input ✅
- Ch 5: Stability ✅
- Ch 6: The Black Box ✅
- Ch 7: Quantity of Variety ✅
- Ch 11: Requisite Variety ✅
- Ch 12: The Error-Controlled Regulator ✅
Ashby, Design for a Brain (1952)
- Ch 8: The Ultrastable System (Homeostat) ✅
- Ch 9: Ultrastability in the Living Organism ✅
McCulloch & Pitts (1943) — A Logical Calculus of the Ideas Immanent in Nervous Activity ✅
Conant & Ashby (1970) — Every Good Regulator of a System Must Be a Model of That System ✅

Management cybernetics (Beer)

Brain of the Firm — System 1-5 chapters
Heart of Enterprise — recursion, variety engineering
Designing Freedom (Massey Lectures, 1973) ✅
Diagnosing the System for Organizations — the diagnostic protocol

Second-order

von Foerster — Observing Systems (selected essays)
- “Ethics and Second-Order Cybernetics” (1990) ✅
Maturana & Varela — Autopoiesis and Cognition + The Tree of Knowledge (selected chapters)
Pask — Conversation Theory primer ✅
Bateson — Steps to an Ecology of Mind: “Cybernetic Explanation”, “Form, Substance, Difference”
Powers — Behavior: The Control of Perception: the perceptual control loop ✅

Privacy / source notes

Some sources (Beer, Bateson, Maturana) aren’t freely available. Tim acquiring as needed. Chapters from non-free sources land when the source does.