History of Cognitive Science: From Macy to Machines

[!NOTE] TL;DR: Deep learning is not a rupture in the history of cognitive science but the return of one of its two founding pathways. This article traces the genealogy from the 1943 triple birth and the Macy conferences (1946-1953) to the LLM debates of 2026, with live-verified citation numbers.

The seventy-four-year error

An ML engineer recently told me attention was "the first brain-inspired architecture". He was off by seventy-four years. In 1943, Warren McCulloch and Walter Pitts had already flattened thought into the firing of idealized neurons, and the entire modern vocabulary of the mechanized mind was in place. I made the same mistake for a decade, with the serene confidence of an engineer who cites nothing older than AlexNet.

This is not a personal failing, it is a disciplinary one. Ask a room of ML practitioners when the history of cognitive science begins and you will get Dartmouth 1956, or worse, 2012. Margaret Boden, whose keynote The History of Cognitive Science: Seven Key Dates remains the sharpest short chronology available, starts instead in 1943. The gap between what the field remembers and what its archives contain is the real subject here.

The question

My main question was whether the LLM era is a rupture or a resurgence. I read the sources as a resurgence. Two pathways were born inside cybernetics between 1943 and 1953, one symbolic and one connectionist; the community that carried them then split, the symbolic heir took the estate, and the disinherited heir came back sixty years later and took everything. The deep learning victory is not a new paradigm. It is the empirical vindication of the cybernetic lineage that Boden traced back to 1943 and that Jean-Pierre Dupuy, in On the Origins of Cognitive Science, accuses later cognitive science of repressing.

flowchart TD
  A["1943: Logical Calculus, purpose as feedback, Craik's cerebral models"] --> B["1946-1953: Macy conferences, one cybernetic community"]
  B --> C["1956: Dartmouth workshop and the MIT symposium"]
  C --> D["Symbolic heir: Logic Theorist, GOFAI, expert systems"]
  C --> E["Connectionist heir: Hebb, Rosenblatt, reverberating circuits"]
  D --> F["1969-1973-1987: the AI winters"]
  E --> G["1969: Perceptrons, the exile"]
  G --> H["1986: PDP volumes, backprop returns"]
  H --> I["1989-2012: convnets to AlexNet"]
  I --> J["2017-2026: transformers, LLMs, agents"]
  F -. "survivors rejoin the convergence" .-> J

1943, the triple birth

That single year saw the three texts Boden marks as the founding gesture of mind-as-machine. The first, McCulloch and Pitts' "A Logical Calculus of the Ideas Immanent in Nervous Activity", fused three ideas by declaring them formally equivalent: neuron theory; the Turing machine; and the Russell-Whitehead propositional calculus. Logic, computation and the brain were now all of a piece, and von Neumann read the paper closely enough to ground the EDVAC design in binary logic instead of decimal arithmetic. In the beginning was the propositional calculus, and von Neumann saw that it was good.

The second text, Rosenblueth, Wiener and Bigelow on purposeful behavior, dissolved teleology into negative feedback: a system reaches a goal because its outputs are corrected against the gap separating it from that goal. The third, Craik's "The Nature of Explanation", proposed that thought runs on small-scale models of the world. His brain is a workshop that builds a miniature of a terrain before acting on the terrain itself, the way an engineer tests a wing in a wind tunnel before committing the wing to the sky. Dupuy's phrase for the whole gesture is exact: cybernetics was not the anthropomorphization of the machine, it was the mechanization of the human.

One community, ten conferences

Between March 1946 and April 1953 the Josiah Macy Jr. Foundation funded ten meetings in New York, with a core group of roughly twenty-five people, Wiener and von Neumann beside Mead and Bateson, Heinz von Foerster as secretary from 1949, Shannon demonstrating a maze-solving machine in 1951. The first five meetings kept no published record, and the last five appeared under the name the group finally accepted: cybernetics. Claus Pias edited the complete transactions in 2016, and they read like the minutes of a family before the divorce.

Because the schism did not exist yet. Boden insists that at this stage "there was still one intellectual community", some members leaning toward logic and meaning, others toward adaptation and self-organization, McCulloch leaning toward both. The most underappreciated member was Rafael Lorente de Nó, Cajal's youngest student. His reverberating circuits, closed chains of interneurons capable of self-sustained activity, were the first neurobiological demonstration of a feedback loop, and a 2023 reconstruction by Espinosa-Sánchez and colleagues in The Neuroscientist restores him to the place he held at the first conference, where he presented the state of neurophysiology right after von Neumann presented the stored-program computer. The temporal depth of your favorite recurrent network runs through a Spanish neuroanatomist who died in 1990.

The tone was not always serene. When Ralph Gerard ventured that synapses do not act digitally, the transcripts preserve what may be the only refusal of the floor in 750 pages of discussion: "May I speak now?" "No, not now." Very human, and very cybernetic at once.

1956 and the two heirs

1956 was the annus mirabilis. The Dartmouth summer project gave the discipline its name and its first manifesto, and the MIT symposium in September assembled, in one room, Miller's magical number seven, Newell and Simon's Logic Theorist, and Chomsky's formal grammars. Miller walked out convinced that experimental psychology, theoretical linguistics and computer simulation were, in his words, "pieces of a larger whole". Chomsky's 1959 demolition of Skinner's "Verbal Behavior" ended behaviorism as an imperial program, and Neisser's "Cognitive Psychology" (1967) named the science that would institutionalize the inheritance.

The estate, however, had two heirs. One read 1943 as: thought is symbol manipulation, so program it top-down. The other read it as: thought is network dynamics, so grow it bottom-up. The table below is the family settlement.

The winters of the symbolic heir

The symbolic heir ran the field for twenty years, then lost its funding twice. Minsky and Papert's "Perceptrons" (1969) proved the single-layer limitation on XOR and froze neural network research for a decade; Boden's sixth key date is precisely this trenchant attack. The Lighthill Report (1973) did comparable damage in Britain, and the expert-systems boom of the 1980s collapsed with the LISP machine market in 1987. I cannot recount that period better than the primary testimony in "La plus belle histoire de l'intelligence" (Robert Laffont, 2018), where Stanislas Dehaene interviews Yann Le Cun. Le Cun describes convolutional networks surviving the winter underground, reading bank checks and postal zip codes while the discipline's official literature declared the approach dead. A paradigm, it turns out, can be clinically dead in the journals and quietly processing your mail.

The connectionist return

The lineage never actually broke. Hebb's learning rule (1949), Rosenblatt's perceptron (1958), Werbos' backpropagation (1974, in a Harvard thesis that went unread for a decade), the PDP volumes (1986), LeCun's zip code network (1989), AlexNet (2012), the transformer (2017): each generation re-engineered the same cybernetic wager, that structure can be grown from data by feedback rather than written by hand. The counts I pulled on 3 November 2026 tell the story of the vindication.

{
  "type": "bar",
  "data": {
    "labels": ["McCulloch and Pitts 1943", "LeCun et al. 1989", "Rumelhart et al. 1986", "Vaswani et al. 2017"],
    "datasets": [{ "label": "Citations", "data": [18560, 11882, 31903, 201060], "backgroundColor": ["#3b82f6", "#3b82f6", "#3b82f6", "#3b82f6"] }]
  },
  "options": {
    "responsive": true,
    "plugins": { "title": { "display": true, "text": "Citations of connectionist milestones, retrieved 3 November 2026" } },
    "scales": { "y": { "title": { "display": true, "text": "Citation count (log-scale venue mix: OpenAlex and Semantic Scholar)" } } }
  }
}

Two readings of this chart. The obvious one: a 2017 architecture paper now carries 201,060 citations (Semantic Scholar), an order of magnitude above the founding logical calculus at 18,560 (OpenAlex). The one I care about: the 1943 paper still accrues citations eighty-three years after publication, which is the signature of a founder rather than a fashion.

The embodied detour

The internal critique arrived in 1991, from inside the family. Varela, Thompson and Rosch's "The Embodied Mind" attacked both GOFAI and classical cognitivism for treating cognition as disembodied symbol crunching, and the enactive program, later the broader 4E and predictive-processing movement, spent thirty years being right in an unfashionable way. I find it striking that the grounding debate over LLMs in the 2020s, whether a model trained only on text can mean anything, replays the 1991 argument almost term for term, with the two camps wearing new jerseys. Although the embodied tradition produced nothing at the engineering scale of a convnet, it anticipated the exact terms of today's loudest AI disagreement.

The convergence, live in 2026

Everything in this section I verified live, because it moves fast. Michael Frank's 2025 review on cognitive modeling with AI argues that today's models are, for the first time, "stimulus-computable": they can receive roughly the inputs a human participant receives, which makes them candidate cognitive models, provided someone builds the model-mechanism mapping he asks for, which nobody has yet. Tuckute, Kanwisher and Fedorenko's Annual Review of Neuroscience piece documents the other direction: language model representations now predict human brain responses well enough to support neural encoding and decoding, and the authors treat these models as "in silico model organisms" for testing hypotheses about language. Dehaene, in a 2023 interview, put the fit honestly: current AI is a good model of about a third of a second of cerebral computation, while the brain runs that third of a second on some 20 watts. His lab now uses machine learning to decode MEG signals. The tool that grew out of the 1943 mechanization has become an instrument pointed back at the organ that inspired it. Cognitive science begat AI, and AI is returning the favor. In my own practice the exchange is concrete: the agent memory architectures I build borrow hippocampal indexing directly, a concept from the same neuroscience tradition cybernetics seeded, which I developed in an earlier piece on the hippocampal index as a scaffold and in a follow-up on rebuilding AI memory systems.

Limits of this genealogy

Four limits I want on the record. First, citation counts mix OpenAlex and Semantic Scholar, retrieved on 3 November 2026, and citations measure visibility, not influence; a 1943 paper cited mostly by historians counts differently than one cited by builders. Second, every retrospective genealogy risks the very teleology Wiener spent 1943 dismantling: reading history backward from the winners flatters the winners, and the losers, Ashby's homeostat or von Foerster's second-order cybernetics, may hold the unexploited ideas. Third, I am a practitioner reading archives, not a historian of science; Dupuy, Boden and Pias are the rigorous guides here, and this piece is a field note built on their scholarship. Fourth, whether LLMs are cognitive models is genuinely open: Frank says the mechanism mapping does not exist yet, and the "GPT-ology" wave of probing studies has known reproducibility problems.

The final zoom

Dupuy closes his excavation with a warning I keep quoting: ignorance of the history of one's discipline, far from being a source of dynamism, acts as a brake on it. He adds that a history of cognitive science which omits its cybernetic origins "gives the impression that connectionism is a new paradigm". It is not. The circle opened in 1943, neurons as logic and purpose as feedback, is closing in 2026, with models grown from data turned back onto the brain as measuring instruments. More generally, architectures die and founding gestures recur: the feedback loop has now outlived every program that declared it dead. If you build agent systems and want this depth of memory design in your stack, see my services page or write to me.