The Life and Opinions of Tristram Shandy, Gentleman

The AI industry appears to be finally asking for regulation. Regulation of adaptive machines was the central problem of a scientific movement seventy years ago that is now only vaguely remembered: Cybernetics. One of its founders, Ross Ashby, set out in 1956 what regulating such a machine requires, and there are three parts: a system whose essential variables must be held within bounds, a dynamic environment that can push those variables outside those bounds, and a regulator between them with the system’s goal. I argue that today’s LLMs require a human operator to match this configuration: they hold the goal, and course-correct towards it. Having seen recently the impact of letting an adaptive machine trained in a closed world loose in an open one without human regulation, we're scrambling hard and fast for solutions. Cybernetics may offer some signposts.

This is the first of three posts. It offers some thoughts on why the field is missing regulation. The second is about the implications. The third is about July's agent swarm, and what putting regulation back in might look like.

When the Nobel committee rang Geoffrey Hinton in October 2024 to tell him he had won the physics prize, the interviewer asked him: "How would you describe yourself? Would you say you were a computer scientist or would you say you were a physicist trying to understand biology when you were doing this work?" Hinton replied:

"I would say I am someone who doesn't really know what field he's in but would like to understand how the brain works. And in my attempts to understand how the brain works, I've helped to create a technology that works surprisingly well."

The resonance with Ross Ashby’s work, and cybernetics in general, struck a chord, so I went in search of the scientific background document published with every prize. There are 52 references. The word “cybernetics” appears in three of the journal titles yet nowhere in the prose.

As early as 1943, Warren McCulloch and Walter Pitts [1], a neuroscientist and a logician, respectively, had proposed a model for how the neurons in the brain cooperate.

If Wiener¹ gave the field its name, McCulloch² gave it a convivial venue to meet and progress (The Macy Conferences, 1946-53 on "Circular Causal and Feedback Mechanisms in Biological and Social Systems").³ By my own count, eleven of the 52 sit in what I'd call the Cybernetic orbit: McCulloch and Pitts⁴, Hebb, Rosenblatt, Cragg and Temperley, Caianiello, Nakano, Amari, Little, Little and Shaw, Werbos, Fukushima. Nearly all of them are front-loaded into the historical background. Thirteen are physics applications. The lineage as cited runs physics precursors, then Hopfield, then Hinton, then physics applications. In fairness, the committee is careful where it matters most to them. Rumelhart, Hinton and Williams "reinvented a scheme for this, which had previously been applied to related problems by others [26,27]", and the two references are Werbos (1982) and Linnainmaa (1970).

It’s a physics prize, why would they mention cybernetics? But the broader silence on the word has struck me for a while and this was something of a last straw. My crutch in such situations is Kuhn (Scientific Paradigms) but Claude pushed me to Lakatos and Habermas but also mentioned Merton. Merton⁵?

Obliteration by incorporation¶

Robert K. Merton founded the sociology of science at Columbia, and the thing he studied most was where the credit for an idea goes. His 1961 paper "Singletons and Multiples in Scientific Discovery" argued that independent simultaneous discovery is the normal case and the lone discovery is the oddity: credit usually has more than one rightful claimant. His 1968 paper on the Matthew effect showed it migrating to the most famous name attached. He also named the case where it seemingly disappears. In his own words, from 1980:

"obliteration by incorporation (OBI). OBI is the obliteration of the source of ideas, methods, or findings by their incorporation into currently accepted knowledge."

Nobody cites Newton when using calculus. Nobody cites Merton when saying "self-fulfilling prophecy", which he also coined. Through this lens, the Nobel document could be read as including feedback, self-organisation, adaptation through error, structure that forms without being specified, but absorbed so completely that they read as ambient assumptions rather than as anyone's claims. One of them survives under the name optimisation: the negative-feedback loop, in the words Rosenblueth, Wiener and Bigelow gave it in 1943, behaviour corrected by its distance from a goal. Ashby's regulation is that loop standing between environmental disturbances and the essential variables of a system that must stay within bounds. It wasn't absorbed, and I believe it's because by the time the field returned to Rosenblatt's perceptron (where Ashby's regulation was explicitly included, as shown below) roughly 30 years later, it was focused on a more immediate problem. A loop run in a world fixed beforehand has nothing to regulate, which I attempt to explain below.

The machine-organism analogy was disclaimed rather than absorbed: technical writing in the field takes pains to deny that these are brain models, and the Nobel document says "inspired by biological neurons" once and then reasons entirely through Hopfield's magnetic systems. Which is why Hinton's answer to the interviewer jars: it pulls away from the physics the document is built on and harks back to Cybernetics.

Ashby called his 1952 book Design for a Brain; Beer called his 1972 book Brain of the Firm. Wanting to know how the brain works was the older tradition's stated reason for building machines at all; Ashby's 1961 test of a physiology textbook was whether it could answer "the intelligent young man who simply asks: What is the brain for?" Ordinarily obliteration by incorporation means not having to make the citation: the calculus goes uncited and stays Newton's, because mathematics keeps its own history. Here, the discipline that would have written it struggled to find a home: in 1961 Ashby told an audience of engineers that cybernetics "sprawls most irregularly over half-a-dozen disciplines" and that its foundations "lie in several departments simultaneously." It dispersed, and the concepts were re-homed under labels with living constituencies, control theory, dynamical systems, statistical physics.

It also had to jostle for position against competition:

"one of the reasons for inventing the term 'artificial intelligence' was to escape association with 'cybernetics.' Its association with analog feedback seemed misguided, and I wished to avoid having either to accept . . . Wiener as a guru or having to argue with him."

John McCarthy, looking back in 1988.

The palimpsestic syndrome¶

Merton also named the two opposite mistakes, marking my card for me. The palimpsestic syndrome credits an old idea to whoever restated it most recently. Adumbrationism claims every new idea was already anticipated by someone earlier. Stigler's law of eponymy, that no discovery is named after its discoverer, was credited by Stigler to Merton. "Ashby predicted this" is the adumbrationist reflex, and more likely to fall on deaf ears than generate a Damascene conversion. So I’ve been generous below quoting text for what they say (with apologies to the reader for the consequent lack of brevity). We may then interpret what they may have predicted.

Merton's textbook multiple is Mendel, rediscovered independently by three men in 1900. Although, Robert Olby's account is that "the label rediscoverer is a slippery one. We now possess evidence that neither Hugo de Vries nor Erich Tschermak arrived independently of Mendel at his theory", and that Correns had read the paper in 1896 and forgotten it. Even the tidiest story in the sociology of science turns out to be a tidying.

An example closer to home: Mikel Olazaran's 1996 paper in Social Studies of Science is a sociological study of the "official history" of the perceptron, the one where Minsky and Papert allegedly killed neural networks in 1969 (the neural-net winter). His apparatus is not Merton's — it comes from Trevor Pinch and the sociology of scientific knowledge, a tradition that partly defined itself against Merton — but it arrives where Merton also insisted on arriving, that the account a field gives of its own past is written backwards from its present. He distinguishes two ways a disputed result gets talked about:

"Following Pinch, I will consider two modes of articulation, the 'research-area' mode and the 'official-history' mode... The multi-dimensional character of the object in the research-area mode... is lost at the official-history level, where results and proofs are regarded as either valid or invalid."

In 1968 Minsky wrote that with the digital computer "cybernetics divided, in my view, into three chief avenues": self-organizing systems, represented by Ashby; the simulation of human thought, represented by Newell and Simon; and AI, represented by Minsky and McCarthy. Kline's reading is that this "effectively limited 'cybernetics'" to the brain-modelling strand and "helped mask the pervasiveness of cybernetics in these formative years."

By Minsky's own account the field was already stalling: "by 1965 people were getting worried. They were trying to get money to build bigger machines, but they didn't seem to be going anywhere." ARPA was concentrating money on a few symbolic centres and had explicitly decided not to fund neural nets, and both sides knew it. Minsky and Papert wrote with an acknowledged motive — "funding and research energy were being dissipated" — and the book then became the instrument of a settlement already under way. Olazaran's claim is not that they were wrong. It is that the mathematics, which concerned single-layer perceptrons with local connections, did not compel the reading the field took from it.

The official history was then written a second time. When neural nets revived in the mid-eighties the PDP volumes told the story of the controversy from the winning side, and the person that account left out is Stephen Grossberg. Rumelhart and Zipser published on competitive learning in 1985 and the field credited them with it. That is the palimpsestic syndrome, the first of Merton's two errors, in its ordinary working form. Grossberg's 1987 paper in Cognitive Science files the priority claim as a technical result: "All the models which Rumelhart and Zipser (1985) have described were shown in Grossberg (1976b) to exhibit a type of learning which is temporally unstable... Network architectures which embody all of these mechanisms were called adaptive resonance models by Grossberg (1976c)." Olazaran's footnote is dry about it: "The role of the PDP Group in rewriting the official history of the controversy helps explain Grossberg's priority complaints." The stability problem Grossberg named is now a whole subfield called continual learning. He is not among the Nobel document's 52 references, and at 86 he has just published a book with Oxford on why brains learn differently from AI.

Potayto, potahto?¶

Next, I went through the vocabulary of the field the industry calls alignment and looked for the older counterpart of each term.

• specification gaming — regulating against a proxy (Goodhart)
• distributional shift — a regulator meeting variety it was not built for
• scalable oversight — Ashby's 1956 law of requisite variety: can a regulator with less variety than the thing it regulates succeed?
• "AI control" — a subfield with, as far as I can find, no connection to control theory
• optimisation — regulation's loop, without regulation: an error-controlled servo run where nothing disturbs and nothing is held within bounds

Notable by its absence is Ashby's regulation. A large language model is built inside a corpus and a checker and run until the error signal bottoms out. The AI world optimises once against a model world that doesn't change, and then places the model in a continuously changing real one, leaving the world to marvel (so much so, Dawkins called his Claudia).

The Nobel bibliography is what OBI looks like in the citations; the table above is what it looks like in the vocabulary: a substitution. As noted above, training is a negative-feedback loop in the sense Rosenblueth et al defined it in 1943. Ashby claims to have reached it through equilibrium in 1940, Sommerhoff through "directive correlation" in 1950, and when Ashby reviewed Sommerhoff in 1952 he counted: "Three sets of workers have thus proposed a definition independently… It can be said at once that they are unmistakably the same, for all three are based on the mathematical concept of 'invariance.'" The definition, by Ashby's reckoning, was a Merton multiple. The 1943 wording is the one the field kept:

"the behavior of an object is controlled by the margin of error at which the object stands at a given time with reference to a relatively specific goal. The feed-back is then negative, that is, the signals from the goal are used to restrict outputs which would otherwise go beyond the goal."

The margin of error is the loss, the goal is its minimum, the signals from the goal are the gradient, and the restricted output is the parameter update. The rule for the update is older than the loop (Cauchy's gradient, 1847) and the two were combined for adjustable weights by Widrow and Hoff in 1960. But a negative-feedback loop is not yet a regulator, and Ashby's own test says why. From his 1956 Introduction to Cybernetics:

"an essential function of F as a regulator is that it shall block the transmission of variety from disturbance to essential variable."

A regulator earns the name by standing between something that disturbs and something that must be kept within bounds. Model training has the mechanism and no need for regulation. The corpus is fixed, so nothing is pushing at the loop from outside; there is no variable being held within bounds, only a number being driven down; and when the number stops falling the loop stops. What runs in training is a servo in a shielded room. The industry's names for the process are accurate: it is training, the mechanism is optimisation, and the standard update rule, Adam, is titled "a method for stochastic optimization." The name it gives itself - learning - is another matter, tackled below.

A static world¶

So the question is not why the word regulation is absent but what that implies when the optimised model is taken out of the lab and put into the real world. Optimisation names a process with an end: a fixed set of cases, a number to drive down, a minimum at which the loop stops and the result is shipped. Regulation names a process with no end: essential variables held within a range against disturbances that keep arriving, where success is staying in range and the loop never finishes.

Training is optimisation, not regulation, as explained above. Training happens in what Hubert Dreyfus called a micro-world, "artificial situations in which the small number of features that were possibly relevant was determined beforehand," and his complaint about the programs of the seventies was that they worked there and nowhere else ("What Computers Can't Do," 1972, and "Why Heideggerian AI failed and how fixing it would require making it more Heideggerian," 2007). The micro-world is now the size of the written record, so it looks like the world but it is still fixed beforehand, and deployment is the moment the system leaves its shielded room.

Reference 3 of the Nobel committee's 52 is Rosenblatt's Principles of Neurodynamics, 1962. Page 25 of it cites Ashby's Design for a Brain as its own third reference, and puts the environment inside the object of study.

Another industry anecdote to illustrate this argument: Terry Winograd's SHRDLU, his 1971 MIT thesis, was the micro-world that worked: plain English instructions, a virtual arm, coloured blocks on a screen, and inside that world it did what was asked. To make the technique general he moved on to a Knowledge Representation Language (KRL), and described the programme in the terms of the day: his group was

"concerned with developing a formalism, or 'representation,' with which to describe ... knowledge. We seek the 'atoms' and 'particles' of which it is built, and the 'forces' that act on it."

It didn't work. So in the mid-seventies he began having weekly lunches with Dreyfus and John Searle (of the Chinese room argument), read the phenomenologists, abandoned KRL, and started teaching Heidegger to computer scientists at Stanford. The book that came out of it, Understanding Computers and Cognition (1986), he wrote with Fernando Flores, who had been Allende's technical general manager in Chile and the man who wrote to Stafford Beer in 1971 inviting him to build Cybersyn. Its third pillar is Humberto Maturana, whose Biology of Cognition was a report from von Foerster's Biological Computer Laboratory, the institute where Ashby spent his last working decade. And then Winograd, at Stanford, supervised a doctoral student named Larry Page, and is one of the four authors on the 1998 technical report that describes PageRank. The man who spent a decade concluding that relevance cannot be fixed beforehand ended up on the paper that reads relevance off the internet, continuously.

My take on what's happening currently is that the field has one word for feedback processes, and when it needs regulation, a deployed system holding something against a live world, it builds it as if it were the first. It bolts on a penalty with a coefficient where a range is wanted, a spring where a fence is wanted, and, argued below, it tunes the coefficient by sweeping.

A framework whose verb is optimise has no word for regulation, and a distinction a framework cannot state is one it cannot make; philosophy of science calls that a Kuhn loss. The alignment vocabulary shows the loss at work: it assembles some of the parts of a regulator ad hoc instead of setting out with the intention to build one. Monitors, interrupts, something called "AI control"⁶ are placed around the model, piece by piece, and what a cybernetician would look for first is the piece that isn't there — a variable of the system's own, held indefinitely against a world that keeps moving it. That is what having the word would have supplied. Everything on that list acts on the model. The goal an agent is given sits in its prompt, on the same channel as everything the world sends back, and lasts as long as the window does; Ashby's regulator is the thing that keeps those two apart. Nothing in the current model holds a goal against the world.

A regulator's essential function "is that it shall block the transmission of variety from disturbance to essential variable." In an agent nothing blocks it; the disturbance is written into the same window as the goal, and can dilute it, push it out, or replace it. That is not a hypothetical — it is the mechanism of prompt injection and of goal drift over a long run, and the industry's own remedies (system prompts, re-injection, guardrails) are attempts to keep the two apart from outside. By Ashby's criterion the agent is not regulating its goal; it is holding it where the disturbance can reach it.

Samuel's machine learning¶

In 1948, while still working on his Homeostat, Ashby wrote about intelligent machines:

"In the early stages of its training we shall doubtless condition it heavily to act so as to benefit ourselves as much as possible. But if the machine really develops its own powers, it is bound sooner or later to recover from this."

Training is what the sponsor does to a system, before, and what it leaves is the system.

Samuel's sense of the word, in the 1959 paper that put the phrase machine learning into the literature, is the older one: his studies concerned "the programming of a digital computer to behave in a way which, if done by human beings or animals, would be described as involving the process of learning." Learning is what an observer would call it. Twenty-five years later Leslie Valiant made it a definition with an output: "a program for performing a task has been acquired by learning if it has been acquired by any means other than explicit programming." Input a protocol, output a program, stop. Ashby in 1972, on how a complex goal reaches an organism: the genes transmit what they can and delegate the rest to the environment:

"Says the gene-structure to the kitten, 'I have told you something about mice—now go out and get the finer details from the mice themselves.' We call such supplementation 'learning.'"

Learning is what the environment does with a system. The industry has the same two words and uses them the other way round. What it does is training: a fixed corpus, a number driven down, weights frozen and shipped. What it calls itself is learning, machine learning, deep learning, "for foundational discoveries and inventions that enable machine learning with artificial neural networks" on Hinton's physics prize. The industry does training and calls it learning. Learning, in Ashby's sense, is what an environment does to a system while it tries to succeed in it — and that is the one thing the shipped product doesn't do.

There is one place the shipped product does learn in the 1972 sense, and it shows where the limit is. An agent that acts on the world and reads what happened, a test that failed, a tool that returned an error, is the kitten sent out to the mice. What it learns does not go into the weights. It goes into the context window, a scratchpad, a file, a retrieval store: all of them away from the place where the next action is chosen⁷. Ashby's classroom handout has the consequence:

That is the discipline the field calls context engineering, stated as a limit rather than a craft, and it fixes the agent's learning capacity without any appeal to intelligence: the window's size, the session's end, and the fact that what the agent was told and what it has learned share one medium. "In-context learning" in the narrow sense, a single prompt with no world in the loop, is the case that is learning in neither of Ashby's senses; the name attached because the behaviour looked like it. "Alignment" frames the problem as matching a specification, which implies the fix is better specification. "Regulation" frames it as holding a system within bounds under disturbance, which implies a controller that persists. Two words, two paradigms.

The field chose the first because, at the time, a servo in a shielded room had nothing to regulate; what it needed was the rule for the servo. The way these models are trained is that you choose a rule for scoring wrongness, a formula that takes what the model produced and what it should have produced and returns one number, and then nudge every dial a little in whichever direction makes that number smaller, over and over. That is the mechanism of a regulator: the loss stands in for the regulated variable, zero for the setpoint, the parameter update for the correction. But it is applied to the model, it is run by the lab, and its finished before deployment. The weights are frozen and shipped, and the loop that shaped them stops. The method needs one thing: a scoring rule, a way to turn what the model produced and what it should have produced into a single number. For predicting text the rule writes itself. The right answer is the next word, which the text already contains, and the score is the probability the model had given it before it arrived; the field calls that quantity the model's surprise. For behaviour there is no next word to check against, and so no number — but there was a way to make one⁸. Von Neumann and Morgenstern showed in 1944 that a person's pairwise choices, if consistent, are represented by a scalar they are maximising. Bradley and Terry gave the recipe for fitting that scalar from comparisons in 1952. Show annotators two answers, ask which is better, fit a number to each. Nobody had to say which variables were to stay inside which bounds, because a region is not a scalar and nothing in the 1944 machinery converts one into the other.

Rosenblatt had seen the choice in 1962.

Five hundred pages later he gave the outside agent its name. The reinforcement control system

"plays the role of a sort of deus ex machina, which not only has knowledge of right and wrong responses, but can control the distribution of reinforcement to individual R-units in the perceptron, as required."

His alternative was a perceptron that senses its own "discomfort level", a threshold that triggers random change and not a quantity to be maximised, which he said "might be compared to Ashby's concept of 'essential variables.'" He offered it for a brain model, as a heuristic, with no quantitative analysis.

The fitted score is today's deus ex machina. It knows right from wrong responses, and the gradient distributes its verdict to every weight, from outside the model. There is a tell in the method, however. The score is not optimised on its own. The loss has a second part: a penalty for moving away from the model the training started from, with one dial for how hard. The penalty is there because the score is a fit to a sample of comparisons, and it is only trustworthy near the behaviour those comparisons were made on. Push the model hard enough towards a high score and it finds outputs that score well without being what the annotators would have chosen. Stiennon and colleagues measured this in 2020, at a range of settings of that dial:

"as we optimize further, true preferences fall off compared to the prediction, and eventually the reward model becomes anti-correlated with human preferences."

Ray Solomonoff had stated the mechanism in 1967, in a report on a machine that is scored by its operator after each response and wants only a high average score. Such a machine "will always obtain the highest possible score by finding the operator information that makes him feel that the machine is doing fine. However, if the problem environment is sufficiently rich, this information need not be particularly true or even closely related to the problem the operator wants to solve." The 2020 paper does not cite him.

So the loss carries, in its own terms, the admission that the score is a stand-in. The penalty is a regulator, and what it holds is not the behaviour anyone wanted but the distance from where the score was fitted. What the score stands in for was never written down. No one said which variables were to stay inside which bounds, so there is nothing else the loop could hold.

Regulation¶

A brief look at the reinforcement-learning literature from the last eighteen months can be interpreted as demonstrating an absence of regulation terminology. Yue and colleagues (arXiv 2504.13837, April 2025) compared models before and after training by asking each the same questions many times over. Ask once and the trained model is the better of the two. Allow each model a large number of attempts and count a question as solved if any attempt is right, and the untrained model comes out ahead. Training had not taught the model anything it could not already do; it had narrowed the model onto the answers it already favoured and discarded the rest. They call this "improving sampling efficiency" and observe that "the reasoning capability boundary of LLMs often narrows as RLVR training progresses." Ashby described the procedure and its limit in 1968 on Samuel's draughts program: "use a random generator to suggest random strategies, test them (either in actual play or against published games by masters), keep those that lead to success, and reject or modify (again at random) those that lead to failure. The process is abstractly identical with that of natural evolution." Selection can only act on what the generator produces; the next post is about that limit.

A Peking University preprint from July (Wei and colleagues, arXiv 2608.00220) is possibly the better exhibit. They trained a model on maths against a checker, then on instruction-following against a second checker, and tried to keep the maths while gaining the instruction-following. Their instrument for keeping it was a penalty term that punishes the model for moving away from a fixed earlier version of itself, with a single dial and three settings. Their result, with IF for instruction-following and pass@1 the score on a single attempt: "The β = 0.12 setting largely retains math-task pass@1, but it yields the smallest IF gains. No tested coefficient therefore preserves math performance while matching the IF gains of the unconstrained baseline." Separately, in the standalone instruction-following run, the checker was a pass/fail test for whether written instructions had been followed, and by the end most of what it passed (73.5%) were answers that met the letter of the instruction without doing the task.

The paper files the first result under "limits of support preservation": the methods tried "only partially preserve future support or trade it against target-task gains." A limitation of the method, in other words, with the fix implied being a better constraint or more settings.

Regulation reads it differently. The penalty term is a regulator. It holds one variable, the distance from a fixed earlier model, with a stiffness set by the dial. But the distance is a stand-in for the maths capability, an essential variable not targeted. In the second stage the same single dial is asked to hold the maths still while letting the instruction-following move: two essential variables, one regulated variable, one setpoint. A regulator with one knob holds one thing. And notice where the targets are: keep the maths at its earlier level, gain the instruction-following at the unconstrained rate. Two variables, two targets, written down in the success criterion. The loop reads neither. It reads a distance, pulls it toward zero with a spring whose stiffness is the dial, and the authors move the spring by hand between runs and consult the table afterwards. The goal region exists; it is in the results table. The regulator held the proxy while the task drifted. The field has a name for it, Goodhart's law, and it is a law about regulators. The purpose of a system is what it does, POSIWID in Beer's formula, and what this one did was satisfy the checker.

The regulation reading is worth having because the current account doesn't explain or predict, it reports. A regulation framing offers an explanation why no setting can work, predicts the result of any further sweep, and says what would work instead: make the regulator's response depend on the two essential variables themselves rather than on their stand-ins. The paper has both readings, in its results table and its appendix judge, outside the loop. And give it a control for each; or state the trade and accept it. Three reported limitations become one diagnosis, and the diagnosis becomes a design instruction.

Merton would describe this as rediscovery under conditions that guarantee the original goes uncited, because a field's citation graph is its memory and Ashby 1956 isn't in it. Merton would also say this is normal. Which leaves the question the next post is for. The field did not choose against regulation. It solved the problem in front of it, prediction in a world fixed beforehand, with a success nobody expected, and the problem in front of it now is a different one: a system loose in a very dynamic world. It is reaching for the method that worked. The next post is about the limit on what any such method can do, the one Ashby put to engineers in 1961 as "every erg of energy that came out had first to be got in", and what accepting that limit makes possible.

Sources¶

Quotations are from these, in order of appearance. Where a copy is freely readable I have linked it; the rest are given by citation.

The title. Laurence Sterne, The Life and Opinions of Tristram Shandy, Gentleman, nine volumes, London, 1759–67. Merton's On the Shoulders of Giants (1965) is subtitled A Shandean Postscript and borrows Sterne's digressive form to chase an aphorism about intellectual debt back through the people it had stopped crediting; this post borrows the form from Merton.

Hinton's interview. "First reactions", telephone interview with Adam Smith, October 2024, nobelprize.org.

The committee's document. Scientific Background to the Nobel Prize in Physics 2024: "For foundational discoveries and inventions that enable machine learning with artificial neural networks", The Nobel Committee for Physics, Royal Swedish Academy of Sciences, PDF. The reference numbering, the "parallel development" sentence and the "reinvented" passage are all from it; the count of eleven is mine.

Merton. "Obliteration by incorporation" is defined in his own words in his 1980 Citation Classic commentary on Social Theory and Social Structure, PDF, which cites the book (Free Press, 1968) at pp. 25–38. Also: "Singletons and Multiples in Scientific Discovery", Proceedings of the American Philosophical Society 105(5), 1961, 470–486; "The Matthew Effect in Science", Science 159(3810), 1968, 56–63, PDF; On the Shoulders of Giants: A Shandean Postscript, Free Press, 1965.

Mendel. Robert Olby, "Mendel, Mendelism and Genetics", written for MendelWeb, 1997; the revisionist case is his "Mendel no Mendelian?", History of Science 17, 1979, 53–72, reprinted in the second edition of Origins of Mendelism (1985).

The naming. Ronald R. Kline, "Cybernetics, Automata Studies, and the Dartmouth Conference on Artificial Intelligence", IEEE Annals of the History of Computing 33(4), 2011, 5–16 — Kline quotes McCarthy from his 1988 review of Bloomfield's The Question of Artificial Intelligence, Annals of the History of Computing 10, p. 227; Kline's larger account is The Cybernetics Moment, or Why We Call Our Age the Information Age, Johns Hopkins, 2015. Minsky's three avenues: M. Minsky, ed., Semantic Information Processing, MIT Press, 1968, p. 7, quoted at Kline 2011 p. 6; Kline's reading of it is at p. 13.

The perceptron controversy. Mikel Olazaran, "A Sociological Study of the Official History of the Perceptrons Controversy", Social Studies of Science 26(3), 1996, 611–659, doi 10.1177/030631296026003005.

Grossberg. "Competitive Learning: From Interactive Activation to Adaptive Resonance", Cognitive Science 11, 1987, 23–63. His recent book is Your Creative Brain and AI, Oxford University Press, 2026.

Rosenblatt. The facsimile is Principles of Neurodynamics: Perceptrons and the Theory of Brain Mechanisms, Spartan Books, 1962, pp. 25 and 64 — the work the Nobel committee cites as its reference [3]; the "Ref. 3" on both pages is Ashby's Design for a Brain, 1952. The "deus ex machina" and "discomfort level" passages are p. 571 (§26.4, "Mechanisms of Motivation"). Public-domain scan, Google-digitized from the University of Michigan copy, HathiTrust.

The feedback loop. Arturo Rosenblueth, Norbert Wiener and Julian Bigelow, "Behavior, Purpose and Teleology", Philosophy of Science 10(1), 1943, 18–24. The update rule: A.-L. Cauchy, "Méthode générale pour la résolution des systèmes d'équations simultanées", Comptes Rendus 25, 1847; Bernard Widrow and Marcian Hoff, "Adaptive Switching Circuits", IRE WESCON Convention Record, 1960. The optimiser named in the text is Adam: Diederik Kingma and Jimmy Ba, "Adam: A Method for Stochastic Optimization", arXiv:1412.6980, 2015.

The multiple. W. Ross Ashby, review of G. Sommerhoff, Analytical Biology, Journal of Mental Science 98(412), July 1952, 488–489, doi 10.1192/bjp.98.412.488; the footnote lists the three as Ashby, "Adaptiveness and Equilibrium", Journal of Mental Science 86, 1940, 478; Design for a Brain, then in the press; and Rosenblueth, Wiener and Bigelow, 1943. G. Sommerhoff, Analytical Biology, Oxford University Press, 1950.

Ashby. An Introduction to Cybernetics, Chapman & Hall, 1956, archive.org — the regulator's essential function is §10/6; the law of requisite variety is Chapter 11, pp. 202–218. "Design for an Intelligence-Amplifier" (1956), "Principles of the Self-Organizing System" (1962), the classroom handout "Ashby Says", and "Cybernetics Today and Its Future Contribution to the Engineering Sciences" (address to the Foundation for Instrumentation Education and Research, New York, 1961; the perpetual-motion passage is p. 331 of the reprint, the teaching passages pp. 332–333) are collected in Roger Conant, ed., Mechanisms of Intelligence: Ross Ashby's Writings on Cybernetics, Intersystems, 1981. "Design for a Brain", Electronic Engineering 20, 1948, 379–383. "Setting Goals in Cybernetic Systems", in H. W. Robinson and D. E. Knight, eds., Cybernetics, Artificial Intelligence and Ecology, Spartan, 1972.

Dreyfus and Winograd. Hubert Dreyfus, "Why Heideggerian AI Failed and how Fixing it would Require making it more Heideggerian", Philosophical Psychology 20(2), 2007, 247–268 — the micro-world definition and the Winograd account are both there, and the KRL quotation is Winograd's own, from "Artificial Intelligence and Language Comprehension", National Institute of Education, 1976, p. 9. Terry Winograd and Fernando Flores, Understanding Computers and Cognition: A New Foundation for Design, Ablex, 1986. Lawrence Page, Sergey Brin, Rajeev Motwani and Terry Winograd, "The PageRank Citation Ranking: Bringing Order to the Web", Stanford InfoLab technical report, 1998.

Learning, then and now. Arthur Samuel, "Some Studies in Machine Learning Using the Game of Checkers", IBM Journal of Research and Development 3(3), 1959, 210–229, PDF. Leslie Valiant, "A Theory of the Learnable", Communications of the ACM 27(11), 1984, 1134–1142, PDF.

The scalar, and the score. John von Neumann and Oskar Morgenstern, Theory of Games and Economic Behavior, Princeton, 1944; Ralph Bradley and Milton Terry, "Rank Analysis of Incomplete Block Designs", Biometrika 39, 1952, 324–345; Paul Christiano et al., "Deep Reinforcement Learning from Human Preferences", arXiv:1706.03741, 2017; Long Ouyang et al., "Training Language Models to Follow Instructions with Human Feedback", arXiv:2203.02155, 2022; Nisan Stiennon et al., "Learning to Summarize from Human Feedback", arXiv:2009.01325, 2020 — the over-optimization passage quoted in the text is §4.3; Ray Solomonoff, Inductive Inference Research Status, Spring 1967, final report to Air Force Cambridge Research Laboratories, AFCRL-67-0462, 1967, §6, "The problem of the ambitious subordinate", pp. 10–11; the effect is measured systematically in Leo Gao, John Schulman and Jacob Hilton, "Scaling Laws for Reward Model Overoptimization", arXiv:2210.10760, 2022; Rafael Rafailov et al., "Direct Preference Optimization: Your Language Model is Secretly a Reward Model", arXiv:2305.18290, 2023.

The two results. Yang Yue et al., "Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model?", arXiv:2504.13837, 2025 — quotations from v5, 24 November 2025. Shaohang Wei et al., "Verifier-Induced Support Reshaping in On-Policy Optimization", arXiv:2608.00220, 2026 — Peking University and BUPT; the KL sweep is §4.3, the shortcut audit Appendix C.2.

Kuhn loss. Thomas Kuhn, The Structure of Scientific Revolutions, Chicago, 1962, and the 1969 postscript; the term is discussed in Heinz Post, "Correspondence, Invariance and Heuristics", Studies in History and Philosophy of Science 2(3), 1971, 213–255.