In the beginning: from the dream of automata to the artificial neuron

The first truly scientific milestone is laid by two British figures of the nineteenth century. Charles Babbage designed the "analytical engine," a programmable mechanical calculator never completed in his lifetime, yet one that already contained the conceptual ingredients of the computer. Ada Lovelace, a mathematician, wrote in 1843 what is often considered the first algorithm intended to be executed by a machine. Lovelace also formulated a famous objection: a machine, she argued in essence, can do only what we order it to do; it can create nothing original. This "Lovelace objection" would be debated for a century and a half, and the apparent creativity of the generative AIs of 2026 gives it fresh relief.

Turing understood that this question is badly framed, for we have no way to define "to think." He therefore replaced it with a concrete test, the imitation game, which has remained famous under the name of the Turing test: if, conversing in writing with a machine without seeing it, a human cannot tell it apart from another human, then honesty requires that we credit it with a form of intelligence. Turing thus shifted the debate from "what is thinking?" (insoluble) to "what observable behavior would we accept as proof?" (testable).

The first successes seemed to prove them right. As early as 1956, the Logic Theorist program by Allen Newell and Herbert Simon proved theorems of logic. In 1958, the psychologist Frank Rosenblatt built the Perceptron, a machine inspired by the biological neuron and capable of learning to recognize simple patterns. The press grew feverish: machines were announced that would walk, talk and be conscious. This excess of optimism, recurrent throughout the history of AI, sets the stage for the fall.

A few striking successes: DENDRAL (1960s) identifies chemical structures; MYCIN (1970s) diagnoses blood infections and recommends antibiotics, sometimes better than junior doctors. In the 1980s, companies invested heavily in this technology.

The apex of a certain idea of brute-force reasoning came in 1997, when IBM's Deep Blue computer beat the world chess champion Garry Kasparov. But beware: Deep Blue "learns" nothing. It explores millions of positions per second thanks to rules written by humans. It is a feat of engineering, not a general intelligence: it would be incapable of playing checkers without being entirely reprogrammed.

But the idea arrived too soon: in the 1980s and 1990s, two indispensable fuels were sorely lacking: data (the Internet did not yet exist on a large scale) and computing power (computers were too slow). The pioneers—such as Geoffrey Hinton in Canada, Yann LeCun (a Frenchman who applied neural networks to reading bank checks) or Yoshua Bengio—plowed a field that would not bear fruit until decades later. They would be nicknamed, much later, the "godfathers of deep learning."

The convergence of these three elements, in the early 2010s, would trigger an explosion that the next chapter recounts in detail.

---

Key takeaways (Chapter 1)

• AI is the heir to a millennia-old dream, which became a science in 1956 at Dartmouth.
• Two philosophies are pitted against each other: symbolic AI (explicit rules) and connectionist AI (learning through neural networks).
• The symbolic approach dominated until the 1980s, with successes (expert systems, Deep Blue in 1997) but fatal limitations: brittleness and the cost of encoding knowledge by hand.
• The "AI winters" are a reminder that the gap between promises and reality always has to be paid for.
• Connectionism was reborn with backpropagation (1986), but would have to wait for the data and the compute of the 2010s to triumph.

In the next chapter, we open up the hood of this connectionist approach: how, concretely, does a machine "learn" from examples?

---