Machines That Think: AI's Four Summers

In 1955, the 28-year-old mathematician John McCarthy drafted a proposal to the Rockefeller Foundation. Needing a neutral name for the new field—one that sidestepped Norbert Wiener's "cybernetics"—he chose Artificial Intelligence. The proposal contained the famous claim: that every aspect of intelligence could in principle be described precisely enough for a machine to simulate it. Co-organizers included Marvin Minsky and Claude Shannon, father of information theory.

In the summer of 1956, ten researchers gathered at Dartmouth College, planning to make "significant progress" in two months. The real stars were not the theorists: Allen Newell and Herbert Simon brought a working program, the Logic Theorist, which proved theorems from Principia Mathematica—one proof even more elegant than the original. The meeting solved nothing large, yet it set the mainstream path for three decades: symbolism—intelligence as logical operations on symbols. The neural-network camp was pushed to the margins.

Today's disputes over when AGI arrives replay Dartmouth's optimism almost exactly: extrapolating today's local breakthroughs into promises of general intelligence.

Broken promises carried a price. In 1969, Minsky and Papert published Perceptrons, proving mathematically that a single-layer neural network cannot even learn "exclusive or"—funding for neural-network research dried up almost overnight. Britain's 1973 Lighthill Report then suspended much of Europe's AI funding. The first winter had arrived.

In the 1980s, AI warmed up again through expert systems: encoding an expert's judgment into thousands of "if…then" rules. DEC's XCON configured VAX orders and reportedly saved tens of millions a year; the medical system MYCIN briefly dazzled. Industry poured money into dedicated LISP machines. But the bubble was fragile: rules multiplied and clashed, anything outside the rules caused collapse, and experts had to hand-feed knowledge—the "knowledge-acquisition bottleneck." In 1987 the LISP-machine market crashed, and the second winter set in.

The "knowledge-acquisition bottleneck" is precisely the old problem today's large models try to bypass: rather than have humans write rules one by one, let the machine grow rules from vast data. Hand-written rules vs. data-driven emergence are two generations' answers to one question.

Go was long seen as AI's Everest. Its board allows 10¹⁷⁰ positions—more than the atoms in the universe—so brute force is hopeless; the 1997 machine that beat Kasparov, "Deep Blue," and its exhaustive search would fail here. Masters rely on an inexpressible "feel." DeepMind's Demis Hassabis chose to attack it anyway: combining deep neural networks (to evaluate positions) with Monte Carlo tree search, then letting the program play millions of games against itself, learning backward from wins and losses.

From March 9–15, 2016, AlphaGo faced world-elite player Lee Sedol in a best-of-five. In game two, at Move 37, AlphaGo played where almost no human would consider—commentators were stunned, professionals said "no human would play that," yet it proved brilliant. AlphaGo won 4–1. In game four Lee answered with his own divine "Move 78" to claw back a win—among humanity's last victories over a top AI.

"Machines proposing options humans never imagined, yet verified as better" is spreading: AlphaFold predicting protein structures, AI designing chip layouts. The real value of human-machine collaboration is not replacement but expanding the human search space—showing you the move beyond your own field of vision.

In 2017 a Google team published Attention Is All You Need, introducing the Transformer—an architecture that lets a model process long text in parallel, the foundation of every large model since. OpenAI scaled this path, building GPT-3 in 2020. But raw GPT-3 was hard to steer, often off-topic. The key step was RLHF (reinforcement learning from human feedback): using human preferences to align the model, teaching it to follow instructions and speak well.

On November 30, 2022, OpenAI wrapped this capability in a chat box and released it as a "low-key research preview," internally expecting a muted response. Instead: one million users in five days, one hundred million in two months—the fastest-growing consumer app in history. For the first time, ordinary people could summon AI in natural language. The technology had existed for years; what ignited it was the layer that "anyone can use."

This is where the "AI super-individual" begins: when the cost of summoning a top model falls to a chat box, leverage sinks from the organization down to the individual. Every past democratization of a general-purpose tool—printing, the personal computer, the internet—redistributed who gets to create.