4.2 How AI got here — and the two winters

What this is and why it exists

This field has collapsed twice, and both collapses followed the same pattern: a demonstration that worked on a small problem, promises made about the general case, then hardware, data or maintenance costs that could not deliver on them. Knowing that pattern is the cheapest defence available against believing the current round of promises uncritically — including your own, when you are the one demonstrating something that works on a small problem.

The vocabulary

• AI winter — a period of collapsed funding and interest following unmet expectations.
• Symbolic AI — approaches representing knowledge as symbols and rules over them.
• Expert system — a program encoding a specialist's rules to give advice in a narrow area.
• Knowledge acquisition bottleneck — the difficulty of getting expert knowledge into a system.
• Combinatorial explosion — the possibilities growing faster than any hardware can enumerate.
• Toy problem — a small, clean problem whose solution may not scale.

The mental model

The founding decade set the pattern of ambition. Turing's paper appeared in Mind in 1950. The field acquired its name six years later: the Stanford Encyclopedia records that "the field of artificial intelligence (AI) officially started in 1956, launched by a small but now-famous summer conference at Dartmouth College, in Hanover, New Hampshire". The proposal for that meeting suggested a significant advance could be made by a small group over a single summer, which is worth remembering whenever you read a confident timeline.

What followed was genuinely impressive on small problems. Programs proved theorems, solved algebra, and manipulated blocks in a simplified world. A conversational program worked by pattern-matching and reflecting the user's own words back as questions, and people confided in it and believed it understood them — a result about human interpretation rather than machine capability, and one of the most instructive findings of the era. A mobile robot combined perception, planning and action into one system for the first time, and revealed how much of the difficulty was in the perception nobody had budgeted for.

The first winter, roughly 1974 to 1980, came from promising the general case. Two documents mark it. Machine translation had been funded on the expectation of automatic, high-quality translation; a 1966 review "concluded, in a famous 1966 report, that machine translation was more expensive, less accurate and slower than human translation", and funding was withdrawn. In Britain, "in 1973, professor Sir James Lighthill was asked by the UK Parliament to evaluate the state of AI research in the United Kingdom", and the resulting report was sharply negative about whether the demonstrations would scale.

Two causes, and both are still live. Combinatorial explosion: the methods worked on small state spaces and the possibilities grow exponentially, so a technique that solved a puzzle could not touch a real problem, no matter what hardware arrived. And the limits of the early neural models: a formal analysis showed what a single-layer network could not represent, and although the fix — layers, with a way of training them — was later found, the perception at the time was that the approach had been closed off. A negative result about the simplest version of an idea was read as a negative result about the idea.

The second winter began in 1987 and came from maintenance rather than capability. Expert systems had been the commercial success of the early 1980s: encode a specialist's rules and give useful advice in a narrow area, and some genuinely paid for themselves. An industry grew around specialised hardware built to run them. Then, in the Wikipedia account, "in 1987, three years after Minsky and Schank's prediction, the market for specialized LISP-based AI hardware collapsed" — general-purpose workstations had become fast enough that the specialised machines had no reason to exist.

The deeper failure was in the systems themselves, and it is the more useful lesson. They worked in demonstration and failed in maintenance. Rules interacted in ways nobody could predict as their number grew; updating them required the original expert and the original engineer; the systems were brittle at the edges of their competence and gave confident wrong answers rather than declining; and getting the knowledge in at all — the acquisition bottleneck — was slower and more expensive than anyone had costed. Accounts differ on when this period ended, with the recovery generally placed through the 1990s.

What ended it was measurable progress on a public benchmark. The milestones people cite came first: a chess machine beat the world champion in a 1997 rematch in New York, having lost to him in 1996; a question-answering system won a televised quiz against the strongest human champions in 2011; a Go-playing system defeated a leading professional over five games in March 2016.

But the turn is better dated to 2012, when a convolutional network won a large image-recognition challenge by a margin far outside the year-on-year improvements that had preceded it — and, as the record notes, "suddenly people started to pay attention" — across the technology industry rather than only within the field. The reason that mattered more than the game-playing milestones is that it was general. A chess result is about chess. A method that recognises images better than everything before it applies to medicine, agriculture, manufacturing and everything else with a camera, and the same approach then worked on speech and on text. Three things had arrived together — much more data, hardware suited to the arithmetic, and training techniques that made depth workable — and none of them alone would have been enough.

So the pattern, stated for use. Both winters followed the same sequence: a demonstration on a clean small problem, a promise about the general case, a scaling wall that was structural rather than incidental, and then withdrawal of funding by people who had been promised something specific. The defence is not pessimism. It is to be precise about what has been demonstrated and on what, to state what would have to be true for it to scale, and to name the thing you do not yet know how to do. A statement that names its own limits survives contact with reality; one that does not is how a winter starts.

What you should now be able to explain or do

Place the founding landmarks and say what each revealed rather than only when it happened. Give the two documents that mark the first winter and the two causes behind it. Explain why a negative result about single-layer networks was over-read. Describe the second winter's hardware collapse and the deeper maintenance failure, including the acquisition bottleneck. Say why the image-recognition result mattered more than the game-playing milestones, and name the three things that arrived together. State the recurring pattern and the defence against it.

Check yourself