More than seventy-five years of artificial intelligence, from Turing's question to today's frontier models.
- The core ideas are old. Turing's test question, the perceptron and backpropagation all came long before today's models.
- Scale is what changed the results. AlexNet showed that more data, more compute and bigger networks work, and the transformer was built in a way that scales.
- ChatGPT changed access, not the science. The architecture already existed; what was new was a text box anyone could type into.
- The field has misjudged its own timing in both directions. It was too optimistic early on and too dismissive later, so confident forecasts deserve caution.
- The goalposts move. Chess, image recognition and conversation each counted as a test of intelligence until a machine did them.
More than seventy-five years, nine turning points. Most of what feels sudden about artificial intelligence was assembled slowly, by people who mostly did not expect this — and twice the whole field was written off as a dead end. Scroll to descend through the decades.
Alan Turing publishes Computing Machinery and Intelligence and sidesteps the unanswerable question of whether machines can think.Not read at source: A. M. Turing, “Computing Machinery and Intelligence”, Mind LIX(236), October 1950. Not read at source: the journal returns 403 to an automated request, so no link is given for a page this site did not open. Checked 16 Sep 2026. Instead he proposes a test: if a machine's replies are indistinguishable from a person's, on what grounds do you deny it?
More than seventy-five years later, that swap — from what a thing is to what it can be observed to do — still shapes every argument about AI, including the arguments on this site.
A summer workshop at Dartmouth College gathers a handful of researchers around a proposal that machines could be made to simulate learning and intelligence.Not read at source: the Dartmouth Summer Research Project on Artificial Intelligence, proposed August 1955 by McCarthy, Minsky, Rochester and Shannon, held summer 1956. Checked 16 Sep 2026. The phrase "artificial intelligence" is coined largely to distinguish the work from existing fields.
The proposal estimated significant progress in a couple of months. That optimism became a permanent feature of the field.
Frank Rosenblatt builds the perceptron:Not read at source: F. Rosenblatt, “The Perceptron: A Probabilistic Model for Information Storage and Organization in the Brain”, Psychological Review 65(6), 1958. Checked 16 Sep 2026. rather than being programmed with rules, it adjusts internal weights based on examples until it classifies them correctly. The press coverage promised far more than the device delivered.
The idea was right and a decade early. Every neural network since is a descendant.
After a long stretch of disillusionment now called an AI winter, a paper popularises backpropagationRumelhart, Hinton & Williams, “Learning representations by back-propagating errors”, Nature 323, 9 October 1986. Link checked 16 Sep 2026. — a method for assigning credit and blame through many layers of a network, so deep systems can actually learn.
This is the mathematical engine underneath everything that follows. The idea existed earlier; 1986 is when the field took it seriously.
IBM's Deep Blue defeats world champion Garry Kasparov.Not read at source: the rematch ran 3–11 May 1997 in New York; Deep Blue won 3½–2½. Checked 16 Sep 2026. It is not learning in the modern sense — it is enormous, specialised search — but it lands as a cultural rupture.
It also sets a pattern that repeats to this day: a task is proof of intelligence until a machine does it, and afterwards it was "just calculation."
A neural network called AlexNet wins the ImageNet image-recognition competition by a wide margin,Krizhevsky, Sutskever & Hinton, “ImageNet Classification with Deep Convolutional Neural Networks”, NeurIPS 2012. Link checked 16 Sep 2026. using graphics processors to train at a scale that had not been practical before.
The lesson the field drew was blunt and consequential: scale works. More data, more compute, bigger networks. Everything after 2012 is partly a response to that finding.
A paper titled Attention Is All You Need introduces the transformer architecture,Vaswani et al., arXiv:1706.03762, June 2017. Link checked 16 Sep 2026. which processes a whole sequence at once and learns which parts to attend to, rather than reading strictly left to right.
It parallelises well, which means it scales — and scale had already been shown to work. Nearly every model you have heard of since is a transformer. See how one is trained.
ChatGPT launches on 30 November 2022 and reaches a mass audience within weeks.Not read at source: the launch date is OpenAI’s own, widely recorded. An earlier version of this line said it reached an audience “faster than almost any product in history” — a superlative this site could not source, so it is narrowed to what is not in dispute. Checked 16 Sep 2026. Nothing fundamental was invented that day: the architecture was five years old and the model already existed. What changed was the interface — a text box anyone could type into.
This is where the internet's composition begins to shift, and where the record on this site starts counting.
Frontier models ship at a pace measured in months rather than years, with an open-weights wave following close behind. Automated traffic has passed half the web, roughly half of new articles are machine-written, and the tools for telling the difference remain too weak to accuse anyone with.Imperva / Thales, 2026 Bad Bot Report, read at source 17 Sep 2026: automated traffic “accounting for more than 53% of all web traffic in 2025, up from 51% the year before” · Graphite, May 2026, read at source 17 Sep 2026: “since Q1 2025 the percentage of primarily AI-generated articles has plateaued at roughly 50%”.
Read this stratum as unsettled. It is the one layer on this page still being poured.
What the pattern actually shows
Three things recur across more than seventy-five years, and they are more useful than any single date.
Progress is lumpy. Two long winters, then sudden jumps. People inside the field were repeatedly wrong about timing in both directions — too optimistic in the 1960s, too dismissive in the 1990s. Anyone confidently telling you what 2030 looks like is doing the same thing.
The goalposts move by design. Chess, then image recognition, then conversation: each was treated as the frontier of intelligence until it was crossed, then reclassified as mere mechanism. This is not dishonesty — it is what happens when you define intelligence by whatever machines cannot yet do.
The breakthroughs were mostly old ideas meeting enough compute. Neural networks date to the 1950s, backpropagation to the 1980s. What changed was hardware and data volume. That is worth remembering when a new capability appears to come from nowhere.
Read any confident AI forecast against this record. The field has been wrong about its own pace in both directions, and a capability that seems to come from nowhere is often an old idea that finally met enough computing power.
This history is not a countdown to something. It is a record of a field that has been wrong about its own timeline in both directions for more than seventy-five years, and is still moving. The useful posture is neither the 1956 promise of imminent thinking machines nor the 1990s certainty that none of it would work — it is paying attention to what is measurable, dating it, and revising when the numbers move.