The beginnings: 1950 to the first AI winters

What happenedCopy Anchor Link

In 1950 Alan Turing asked whether machines can think and proposed the imitation game, now called the Turing test. In 1956 a summer workshop at Dartmouth coined the term "artificial intelligence". Early programs proved theorems and played checkers, and researchers predicted human-level AI within a generation.

Two approaches formed. Symbolic AI wrote rules by hand: if the patient has a fever and a rash, then consider measles. Connectionism copied the brain's idea instead. In 1958 Frank Rosenblatt built the perceptron, a single artificial neuron that learns its weights from examples. In 1969 Minsky and Papert showed a single perceptron cannot learn simple functions like XOR. Many researchers and funders moved away from neural networks afterward.

The wintersCopy Anchor Link

Promises ran far ahead of results. Machine translation and general problem solving turned out to be much harder than expected. Government reports cut funding in the 1970s, the first AI winter. Expert systems, hand-written rule bases for narrow jobs, brought a boom in the 1980s. They were expensive to maintain and brittle outside their narrow domain, and the market collapsed around 1987, the second winter.

IBM's Deep Blue beat world chess champion Garry Kasparov in 1997, but it won by searching millions of positions per second with hand-tuned evaluation. It did not learn.

Why it mattersCopy Anchor Link

The pattern repeats: a method works on a narrow task, expectations balloon, then reality corrects. The shift that finally broke the cycle was not a smarter rule. It was letting models learn from large amounts of data with a lot of compute, which is the next chapter of the story.

In plain wordsCopy Anchor Link

Early AI was a bet on two different ideas. One bet said: intelligence is rules, so write enough of them. The other said: intelligence is learning, so build something that can learn from examples, like a brain does. For about forty years the rules side won the money, because rules were easier to run on slow computers. Learning needed two things nobody had yet: huge amounts of data and fast chips.

A worked example: why rules breakCopy Anchor Link

Suppose you write rules to spot spam: "if the email contains FREE and MONEY, it is spam." It works for a week. Then spammers write "fr33 m0ney". You add a rule. They change again. A rule-writer is always one step behind, and a real inbox needs thousands of rules that interact in strange ways. A learning system instead sees ten thousand examples of spam and not-spam and finds the patterns itself, including ones no person would think to write down.

Tracing a perceptronCopy Anchor Link

A perceptron multiplies each input by a weight, adds the results, and outputs 1 if the total passes a threshold. Learning means nudging weights after each mistake. Invented numbers for illustration: teach it OR (output 1 if either input is 1) with a threshold of 0.5 and both weights starting at 0.

1. Input (0,0), correct answer 0. The total is 0, so the output is 0. Correct, no change.
2. Input (0,1), correct answer 1. The total is 0, so the output is 0. Wrong, so add the input to the weights. They become (0,1).
3. Input (1,0), correct answer 1. The total is 0, so the output is 0. Wrong, so the weights become (1,1).
4. Check all four inputs again. The totals are 0, 1, 1 and 2, so every answer is right.

Now try XOR, which means "exactly one input is 1". It needs 1 for (0,1) and (1,0) but 0 for (1,1). Each weight alone must beat the threshold, while the two together must not. With a threshold of zero or more, no weights do both. That is the limit Minsky and Papert proved.

A second example: when rules are the right toolCopy Anchor Link

Not every problem needs learning. Take leap years: a year is a leap year if it divides by 4, except century years (ending in 00), unless it also divides by 400. So 1900 was not a leap year and 2000 was. The rule is short, complete and exact. A model trained on examples would only approximate it and might get 1900 wrong.

Rules win when the logic is fully known, small and must be explained to an auditor. Learning wins when nobody can state the logic, as with recognizing a face. Early AI went wrong by using rules on the second kind of problem.

What goes wrong: the demo trapCopy Anchor Link

Early systems were shown on a few hand-picked cases. The symptom is an impressive demo followed by failures on the first real inputs. A rule base for one disease has no rule for a patient who has two.

To confirm the cause, collect inputs the builders did not choose and count how many the system handles. A large gap means the demo covered one small corner. The same check applies today: a polished demo shows an idea can work, not that it works on your data.

Deep Blue shows the honest label. The mechanism was fast search plus hand-tuned scoring. Calling that "thinking" stretches the word, since the mechanism is search and scoring, not human-style reasoning.

Common mistakeCopy Anchor Link

Thinking the winters happened because the ideas were wrong. Neural networks were a good idea in 1958. They were starved of data and computing power. Good ideas can look like failures until the ingredients arrive.