{"course":{"slug":"ai-actually-explained","title":"AI, Actually Explained","url":"https://addaly.com/learn/ai-actually-explained","level":"beginner","track":"start"},"lesson":{"slug":"the-two-winters","title":"Two winters, and what actually froze","url":"https://addaly.com/learn/ai-actually-explained/the-two-winters","position":43,"of":92,"minutes":9,"answer":"A winter is what happens when a system that works on the demonstration meets the long tail of real cases, and its cost of maintenance outruns its value.","keyIdea":"A winter is what happens when a system that works on the demonstration meets the long tail of real cases, and its cost of maintenance outruns its value."},"body":{"format":"markdown","text":"## A winter is a funding event\n\nThe phrase \"AI winter\" makes it sound like the ideas stopped working. They did not. What stopped was money, and it stopped for reasons that are still the reasons a project fails today. Two winters, two mechanisms, both still active.\n\n## The first winter, 1974 to 1980: the long tail\n\nThe early programs solved problems in tiny worlds: a few blocks on a table, a proof from a textbook, a conversation that only reflected your own words back. Each success was real. Each was also a demonstration, and the assumption was that the real world was the demonstration, only larger.\n\nIt was not. The number of situations grows explosively as a world gets bigger, and a rule written for a toy world has to be joined by a thousand more for the real one. Machine translation showed this first. A 1966 report to the US government concluded that after a decade of funding, translation by computer was slower, worse and more expensive than a human, and recommended the money stop. It did. In 1973 a report commissioned by the British government reached a similar conclusion about the field as a whole, and most British AI research lost its funding within a year.\n\nThe mechanism to keep: **a system that works on the cases you showed it, and breaks on the cases you did not, has not solved the problem. It has solved the demonstration.** You met the same mechanism in this course when a model answered fluently inside a familiar structure and invented the specifics.\n\n## The second winter, 1987 to 1993: the cost of hand-written rules\n\nThe 1980s boom was expert systems: software that captured what a specialist knew as a set of if-then rules. One at Digital Equipment Corporation configured computer orders and was reported to save the company tens of millions of dollars a year. It started with a few hundred rules and grew past ten thousand. Every one was written by a person, after interviewing an expert, and every one had to be revised when the products changed.\n\nThat was the problem. The rules did not learn. When a case fell outside them, the system did not degrade gracefully; it gave a confident wrong answer or nothing. And the people who could write the rules were scarce and expensive. The field called this the knowledge acquisition bottleneck, and no amount of cleverness removed it, because the bottleneck was the decision to write rather than learn.\n\nIn 1987 the market for the specialised computers these systems ran on collapsed in a year. Japan's national Fifth Generation project, launched in 1982 with the aim of machines that reasoned in human language, wound down in 1992 without reaching it.\n\n## What thawed it was not a new idea\n\nHere is the part that matters for reading the present. The method for training multi-layer networks (the thing underneath every model in this course) was popularised in 1986, in the middle of the second winter. The idea existed. It did not work well enough to matter, because the networks that could be trained on the computers of the day were tiny and the labelled examples to train them on did not exist.\n\nWhat changed over the following twenty-five years was not the idea. It was the hardware, which roughly doubled in capability every couple of years, and the data, which the internet made abundant. The idea sat waiting for the world to catch up with it. That is uncomfortable for a field that likes to credit insight, and it is why the next lessons are about a photograph competition and a graphics chip rather than about a theory.\n\n## Is another one coming?\n\nNobody knows, and the honest way to hold the question is to check both mechanisms against the present.\n\nThe long-tail mechanism is alive: a model that passes a demonstration and fails in deployment is the commonest way an AI project dies today, and you have already seen why. The maintenance mechanism has flipped: a learned system is cheaper to build than a rule-based one and does not need an expert per rule, which is the whole reason the current boom exists.\n\nWhat is new is scale on both sides of the ledger. There is real revenue (billions a year across the main labs) and there is capital spending announced in the hundreds of billions, on data centres that must earn a return. Whether the first grows fast enough to justify the second is the open question, and it is a business question rather than a scientific one. A winter, if it comes, will come from that ledger."},"check":{"question":"The expert systems of the 1980s often worked, and some saved real money. Why did the approach still collapse?","options":[{"text":"The computers of the time were too slow to run more than a few hundred rules.","correct":false,"tempting":"Blames hardware; the systems ran, and the largest carried over ten thousand rules in production."},{"text":"Neural networks arrived in 1987 and replaced them.","correct":false,"tempting":"Misplaces the thaw; learned systems did not displace rules commercially until after 2012."},{"text":"Users and their professional bodies refused to trust software that gave advice on medicine, engineering and finance.","correct":false,"tempting":"Picks the plausible social story; the documented failure was economic, in the cost of writing and revising the rules."},{"text":"Every rule was written by hand, so maintenance outran the savings and unfamiliar cases got confident wrong answers.","correct":true}]},"prev":{"slug":"seventy-years-in-one-page","title":"Seventy years in one page","url":"https://addaly.com/learn/ai-actually-explained/seventy-years-in-one-page"},"next":{"slug":"rules-give-way-to-data","title":"The quiet turn: from rules to statistics","url":"https://addaly.com/learn/ai-actually-explained/rules-give-way-to-data"},"authorship":"Founder and series editor: Dr Mustafa Mun. Written by AI models to a written standard. Exam keys checked blind by a second model: 3,665 of 3,666 agreed. Academic review invited.","review":{"status":"Draft, open for academic review"},"cite":"Addaly (2026) 'Two winters, and what actually froze', in AI, Actually Explained. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/ai-actually-explained/the-two-winters (Accessed: 13 October 2026).","licence":{"name":"CC BY-NC-SA 4.0","url":"https://creativecommons.org/licenses/by-nc-sa/4.0/"},"reader":{"schema":1,"title":"Two winters, and what actually froze","courseTitle":"AI, Actually Explained","module":{"index":5,"slug":"how-it-got-here","title":"How it got here, and what changed in 2022"},"answer":null,"keyIdea":"A winter is what happens when a system that works on the demonstration meets the long tail of real cases, and its cost of maintenance outruns its value.","hero":{"kind":"plate","image":null,"alt":null,"caption":null},"blocks":[{"t":"h2","id":"a-winter-is-a-funding-event","runs":[{"text":"A winter is a funding event"}]},{"t":"p","runs":[{"text":"The phrase “AI winter” makes it sound like the ideas stopped working. They did not. What stopped was money, and it stopped for reasons that are still the reasons a project fails today. Two winters, two mechanisms, both still active."}]},{"t":"h2","id":"the-first-winter-1974-to-1980-the-long-tail","runs":[{"text":"The first winter, 1974 to 1980: the long tail"}]},{"t":"p","runs":[{"text":"The early programs solved problems in tiny worlds: a few blocks on a table, a proof from a textbook, a conversation that only reflected your own words back. Each success was real. Each was also a demonstration, and the assumption was that the real world was the demonstration, only larger."}]},{"t":"p","runs":[{"text":"It was not. The number of situations grows explosively as a world gets bigger, and a rule written for a toy world has to be joined by a thousand more for the real one. Machine translation showed this first. A 1966 report to the US government concluded that after a decade of funding, translation by computer was slower, worse and more expensive than a human, and recommended the money stop. It did. In 1973 a report commissioned by the British government reached a similar conclusion about the field as a whole, and most British AI research lost its funding within a year."}]},{"t":"p","runs":[{"text":"The mechanism to keep: "},{"text":"a system that works on the cases you showed it, and breaks on the cases you did not, has not solved the problem. It has solved the demonstration.","b":true},{"text":" You met the same mechanism in this course when a model answered fluently inside a familiar structure and invented the specifics."}]},{"t":"h2","id":"the-second-winter-1987-to-1993-the-cost-of-hand-written-rule","runs":[{"text":"The second winter, 1987 to 1993: the cost of hand-written rules"}]},{"t":"p","runs":[{"text":"The 1980s boom was expert systems: software that captured what a specialist knew as a set of if-then rules. One at Digital Equipment Corporation configured computer orders and was reported to save the company tens of millions of dollars a year. It started with a few hundred rules and grew past ten thousand. Every one was written by a person, after interviewing an expert, and every one had to be revised when the products changed."}]},{"t":"p","runs":[{"text":"That was the problem. The rules did not learn. When a case fell outside them, the system did not degrade gracefully; it gave a confident wrong answer or nothing. And the people who could write the rules were scarce and expensive. The field called this the knowledge acquisition bottleneck, and no amount of cleverness removed it, because the bottleneck was the decision to write rather than learn."}]},{"t":"p","runs":[{"text":"In 1987 the market for the specialised computers these systems ran on collapsed in a year. Japan’s national Fifth Generation project, launched in 1982 with the aim of machines that reasoned in human language, wound down in 1992 without reaching it."}]},{"t":"h2","id":"what-thawed-it-was-not-a-new-idea","runs":[{"text":"What thawed it was not a new idea"}]},{"t":"p","runs":[{"text":"Here is the part that matters for reading the present. The method for training multi-layer networks (the thing underneath every model in this course) was popularised in 1986, in the middle of the second winter. The idea existed. It did not work well enough to matter, because the networks that could be trained on the computers of the day were tiny and the labelled examples to train them on did not exist."}]},{"t":"p","runs":[{"text":"What changed over the following twenty-five years was not the idea. It was the hardware, which roughly doubled in capability every couple of years, and the data, which the internet made abundant. The idea sat waiting for the world to catch up with it. That is uncomfortable for a field that likes to credit insight, and it is why the next lessons are about a photograph competition and a graphics chip rather than about a theory."}]},{"t":"h2","id":"is-another-one-coming","runs":[{"text":"Is another one coming?"}]},{"t":"p","runs":[{"text":"Nobody knows, and the honest way to hold the question is to check both mechanisms against the present."}]},{"t":"p","runs":[{"text":"The long-tail mechanism is alive: a model that passes a demonstration and fails in deployment is the commonest way an AI project dies today, and you have already seen why. The maintenance mechanism has flipped: a learned system is cheaper to build than a rule-based one and does not need an expert per rule, which is the whole reason the current boom exists."}]},{"t":"p","runs":[{"text":"What is new is scale on both sides of the ledger. There is real revenue (billions a year across the main labs) and there is capital spending announced in the hundreds of billions, on data centres that must earn a return. Whether the first grows fast enough to justify the second is the open question, and it is a business question rather than a scientific one. A winter, if it comes, will come from that ledger."}]}],"check":{"question":[{"text":"The expert systems of the 1980s often worked, and some saved real money. Why did the approach still collapse?"}],"options":[{"text":[{"text":"The computers of the time were too slow to run more than a few hundred rules."}],"correct":false,"tempting":[{"text":"Blames hardware; the systems ran, and the largest carried over ten thousand rules in production."}]},{"text":[{"text":"Neural networks arrived in 1987 and replaced them."}],"correct":false,"tempting":[{"text":"Misplaces the thaw; learned systems did not displace rules commercially until after 2012."}]},{"text":[{"text":"Users and their professional bodies refused to trust software that gave advice on medicine, engineering and finance."}],"correct":false,"tempting":[{"text":"Picks the plausible social story; the documented failure was economic, in the cost of writing and revising the rules."}]},{"text":[{"text":"Every rule was written by hand, so maintenance outran the savings and unfamiliar cases got confident wrong answers."}],"correct":true,"tempting":null}],"order":[2,0,1,3],"why":null},"prev":{"slug":"seventy-years-in-one-page","title":"Seventy years in one page","minutes":8},"next":{"slug":"rules-give-way-to-data","title":"The quiet turn: from rules to statistics","minutes":8}}}