One field, two and a half centuries, and a recurring confusion: the distance between a machine that looks like it thinks and one that does. It opens with an outright fraud — a man folded inside a cabinet, moving the chess pieces — and it arrives at a question no one can yet answer, about whether the newest machines have any inner life at all. In between is a real chain of engineering. This record tags every entry by how firm the claim actually was, because the interesting history is the sorting.
A man pretending to be a machine, at one end. Machines whose minds we argue over, at the other. The field is the space between those two mistakes.
Automata & the illusion of mind
1770 – 1843 · before there was computing, there was the performance of it.
The Mechanical Turk
Wolfgang von Kempelen unveils a chess-playing "automaton" for the Habsburg court. It tours Europe and America for decades, beating (by legend) Napoleon and Benjamin Franklin. It is a fraud: a human chess master hidden in the cabinet, working the arm through levers. Edgar Allan Poe reasons out the trick in 1836. The name survives today as the label for work that looks automated but is quietly done by people.
The Jacquard loom
Punched cards drive a loom to weave arbitrary patterns — the first widely used programmable machine. Not intelligence, but the idea that instructions can be stored on a card and fed to a machine. Babbage kept a woven-silk Jacquard portrait; the punch card ran computing into the 1970s.
Babbage's Analytical Engine
Charles Babbage designs a general-purpose mechanical computer — a "store," a "mill" (memory and processor), conditional branching, loops. It is never built in his lifetime, but the architecture is genuinely universal in outline, a century ahead of the electronics that would realise it.
Lovelace's Note G
Ada Lovelace publishes, in her notes on the Engine, an algorithm to compute Bernoulli numbers — often called the first published computer program. She also draws the field's first hard line: the Engine "has no pretensions whatever to originate anything." Turing would name and answer this "Lovelace objection" a century later.
The formal foundations
1936 – 1956 · what a computation is, what a neuron is, and what it would mean to call a machine intelligent.
The Turing machine
In On Computable Numbers, Alan Turing defines an abstract machine that captures exactly what can be mechanically computed. The universal Turing machine — one machine that can simulate any other given its description — is the theoretical seed of the general-purpose computer.
The McCulloch–Pitts neuron
Warren McCulloch and Walter Pitts model the neuron as a logical threshold unit and show that networks of them can compute logical functions. The first bridge between brains and computation, and the distant ancestor of every artificial neural network.
Cybernetics & information
Norbert Wiener's Cybernetics frames control and feedback in animal and machine alike; the same year, Claude Shannon's information theory gives a precise measure of a bit. Together they hand the coming field its vocabulary of signal, feedback and uncertainty.
The Turing Test
In Computing Machinery and Intelligence, Turing sidesteps "can machines think?" as too vague and proposes the Imitation Game: judge a machine by whether its conversation is indistinguishable from a human's. A test of behaviour, deliberately not of inner experience — a distinction the field is still arguing about at the far end of this record.
The Dartmouth workshop
John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon convene a summer workshop and coin the term artificial intelligence. The proposal's optimism — that a summer might make real progress on language, abstraction and self-improvement — sets the pattern of promise-outrunning-delivery that will define the next thirty years.
The symbolic age & the first winter
1956 – 1974 · reasoning as symbol manipulation — and the first collision with reality.
Logic Theorist
Allen Newell and Herbert Simon build a program that proves theorems from Principia Mathematica — one proof more elegant than the book's. Often called the first AI program; the start of the symbolic, rule-manipulating tradition that dominates the era.
The Perceptron
Frank Rosenblatt's Perceptron is a real, trainable single-layer network — a genuine advance. But it is announced as an "embryo" of a machine expected to walk, talk, see and reproduce itself. The result was solid; the claim was science fiction, and the gap set up a hard fall.
ELIZA & the ELIZA effect
Joseph Weizenbaum's ELIZA parrots a Rogerian therapist with simple pattern rules — no understanding whatsoever. Users confided in it and insisted it cared. Weizenbaum was alarmed enough to spend the rest of his career warning about it. The ELIZA effect — reading a mind into fluent output — is the direct ancestor of every later "the chatbot is alive" episode.
Perceptrons — the book
Minsky and Papert prove what a single-layer perceptron cannot do (the XOR problem among them). Rigorous and correct — but widely read as a verdict on neural networks as a whole. Funding drained toward symbolic AI; connectionism went quiet for over a decade.
The first AI winter
The UK's Lighthill Report judges AI to have failed its grand promises; funding is cut in Britain and the US alike. The first of the field's boom-and-bust cycles — the bill coming due for a decade of overselling.
Knowledge, backprop & the second winter
1980 – 1993 · expert systems boom, neural nets quietly learn to go deep, then the money leaves again.
Expert systems
Programs like MYCIN and XCON encode human expertise as thousands of hand-written rules, and for narrow tasks they genuinely work — a real industry forms. But the rules are brittle, costly to maintain, and blind outside their domain. Japan's vast Fifth Generation project bets the decade on this approach and largely misses.
Backpropagation
Rumelhart, Hinton and Williams popularise training multi-layer networks by propagating errors backward — the algorithm that answers Minsky and Papert and quietly underwrites everything that comes after. It waits decades for the data and compute to make it roar.
The second AI winter
The specialised Lisp-machine market collapses and the expert-system boom deflates. Investment retreats again — the second cycle of the same lesson about the distance between demo and product.
Machine learning takes over
1997 – 2011 · statistics and data beat hand-written rules; a computer takes the chess crown for real this time.
Deep Blue beats Kasparov
IBM's Deep Blue defeats reigning world champion Garry Kasparov over six games — the first time a machine beats a top human at chess under tournament conditions. No hidden operator this time: brute-force search on purpose-built hardware. The Turk's fantasy, made honest 227 years later.
"Deep learning" gets its name
Hinton and colleagues show deep networks can be trained layer by layer, reviving interest in many-layered models and re-branding the connectionist programme as deep learning. The stage is set; it waits on the hardware.
The deep-learning revolution
2012 – 2016 · GPUs, big data and old ideas ignite at once — vision, generation and game-play fall in four years.
AlexNet
Krizhevsky, Sutskever and Hinton win the ImageNet contest by a stunning margin with a deep convolutional net trained on GPUs. The result that convinces the field, and then the industry, that deep learning works. The modern era starts here.
GANs
Ian Goodfellow's generative adversarial networks pit a generator against a discriminator, each improving the other, and machines begin to synthesise convincing images. The far side of this line runs to deepfakes and the modern image generators.
AlphaGo & Move 37
DeepMind's AlphaGo beats Lee Sedol at Go, a game long thought a decade off. Its "Move 37" in game two was so alien that commentators assumed an error — then saw it was brilliant. Deep learning fused with search, producing play no human had taught it.
The transformer era
2017 – 2022 · one architecture eats the field, scale becomes the strategy, and the public meets it.
"Attention Is All You Need"
A Google team introduces the transformer, replacing recurrence with attention — a mechanism that lets a model weigh every part of its input against every other, in parallel. It is the architecture under every large language model that follows. The single most consequential paper of the modern era.
BERT & GPT — pre-training
BERT and the first GPT establish the recipe: pre-train a transformer on a vast corpus, then adapt it. Language models stop being trained per-task and start being general starting points. The "foundation model" is born.
GPT-3 & scaling laws
A 175-billion-parameter model shows striking few-shot ability, and Kaplan and colleagues chart scaling laws — performance rising predictably with model, data and compute. The field's strategy becomes, in large part, "make it bigger," with results to back it.
ChatGPT
OpenAI wraps a tuned GPT-3.5 in a chat box, and in two months it reaches a hundred million people — the fastest consumer-software adoption on record to that point. The technology was years old; the interface was the event. AI stops being a research word.
The bleeding edge
2023 – now · multimodal, aligned, agentic, and reasoning out loud — with the hype machine running hot alongside.
Multimodal foundation models
GPT-4, Claude, Gemini and their peers handle text, images and code together, pass professional exams, and are steered by human and AI feedback (RLHF, and Anthropic's Constitutional AI). Real, broad capability — shadowed by confident wrong answers the field calls hallucination, the exact failure this whole record is built against.
Reasoning models & agents
Models trained to spend compute "thinking" before answering — writing out chains of reasoning — sharply improve on maths and code. Wrapped in tool use, they become agents that browse, run code and act over many steps. Genuinely more capable, and genuinely harder to verify.
The "AGI is imminent" chorus
As capability climbs, so do the claims — that general intelligence, or superintelligence, is a year or two away. Some of it is serious forecasting; much of it is fundraising and press. Tagged hyped not because the systems are fake, but because the timelines outrun any evidence, in the oldest pattern in the field.
The frontier — is anyone home?
The genuinely open questions. Not hoaxes — unsolved science and philosophy, where there is no settled fact yet. Tagged Open on purpose.
The Hard Problem
David Chalmers names the hard problem of consciousness: even a complete account of what the brain does leaves unexplained why there is something it is like to be it. It is the wall every claim about machine sentience runs into — we cannot yet say what would even count as evidence of inner experience.
Global Workspace Theory
Bernard Baars, later developed by Stanislas Dehaene, proposes consciousness as a "workspace" that broadcasts information across the brain. A leading, testable, mainstream theory — and one whose computational shape you can actually look for in an AI system. Serious, influential, not settled.
Integrated Information Theory
Giulio Tononi's IIT proposes that consciousness is integrated information, measured as a quantity Φ. Bold and precise — and fiercely contested: in 2023 some 120 researchers signed an open letter calling it, as tested, "pseudoscience." The dispute itself is the honest content; take neither the theory nor the letter as the last word.
Orch-OR — Penrose & Hameroff
Roger Penrose (a Nobel physicist) and Stuart Hameroff argue consciousness arises from quantum processes in neuronal microtubules, and that the mind is therefore non-computable. Real credentials, striking claim, near-zero uptake among neuroscientists. Filed as a serious fringe position, not a crank one — the gap between the two is the point.
The LaMDA episode
A Google engineer, Blake Lemoine, publicly claimed the LaMDA chatbot was sentient. The event was real; the claim was rejected by essentially every expert. It is the ELIZA effect at industrial scale — a case study in how fluent output invites us to read a mind in, not evidence of one. Tagged apart from the genuine open questions above for exactly that reason.
The Butlin–Long report
Nineteen researchers, drawing on the leading neuroscientific theories, derive a checklist of "indicator properties" and assess current AI against them. Their sober conclusion: no present system is a strong candidate for consciousness — and there is no in-principle barrier to a future one. The most honest anchor in the whole debate, and the frame this record borrows.
Model welfare
The question moves from a lab curiosity to an institutional one: Anthropic stands up a model-welfare programme; researchers at Eleos AI study whether, and when, AI systems could warrant moral consideration. Not a claim that today's models suffer — a claim that the question deserves serious, careful attention before it's urgent.
Roko's Basilisk — lore, not science
A thought experiment from the LessWrong forums about a future AI that might punish those who failed to help build it. Never a claim about any real system — a decision-theory puzzle that hardened into internet folklore. Included as the boundary marker: this is where the frontier shades into myth, and where honest tagging matters most.
Honest limits
This is a compendium, not a textbook — a curated spine, not every node. Dates mark a landmark moment (a paper, a match, a release), and many entries had long run-ups and disputed firsts; where a "first" is genuinely contested, treat the year as a signpost, not a verdict. The four firmness tags are judgements, made in the open so you can disagree with them: Real means the result stood up, Hyped means the thing was real but the claim outran it, Hoax / Misread means no thinking machine was actually there, and Open means the honest answer is that we do not yet know.
Why one record? Because the field's whole trouble is the distance between a machine that looks like it thinks and one that does — and the only cure is to keep the trick, the hype, the result and the open question in the same frame, side by side, labelled. Same discipline as the rest of lucid.rodeo: established fact kept honestly apart from wonder.