On May 11, 1997, a machine made a world chess champion get up from the board and stalk off in disgust. Garry Kasparov, the most dominant player the game had ever produced, had just resigned a match to a refrigerator-sized IBM computer named Deep Blue. It was one of the most public humiliations in the history of human-versus-machine competition, and it remains one of the most misunderstood.
Two Matches, Two Outcomes
There were, in fact, two matches, and conflating them is the first mistake people make. In February 1996, in Philadelphia, Kasparov faced an earlier version of Deep Blue and won comfortably, 4–2. The machine startled everyone by taking the very first game—the first time a computer had beaten a reigning world champion under tournament time controls—but Kasparov recovered, adapted, and ground it down over the remaining five. The lesson he took away was that the machine could be steered into positions it did not understand.
The rematch came in May 1997 in New York, against a substantially upgraded machine the team had nicknamed “Deeper Blue.” This time Kasparov lost, 3.5–2.5. The margin was a single point, decided in the sixth and final game when a rattled Kasparov blundered into a known opening trap and resigned in under twenty moves. For the first time, a computer had beaten the world champion in a full match. The symbolism was enormous; the chess, in the end, was decided by human nerves as much as machine strength.
How Deep Blue Actually Worked
Here is the part that gets romanticized into something it wasn’t. Deep Blue did not learn. It did not reason about chess the way Kasparov did, nor anything close. It was a triumph of raw search and specialized engineering:
- Custom silicon. The 1997 machine was an IBM RS/6000 SP system with 30 processors augmented by 480 special-purpose chess chips, each one designed to do nothing but evaluate chess positions in hardware.
- Brute-force search. It evaluated roughly 200 million chess positions per second, exploring the game tree far deeper than any human could, using alpha–beta search to prune branches that couldn’t improve on what it had already found.
- Handcrafted evaluation. At the leaves of that search sat an evaluation function with thousands of tunable parameters—material, king safety, pawn structure, control of the center—hand-engineered and refined with input from grandmasters, notably Joel Benjamin, who helped the team teach the machine what “good” positions looked like.
That is the whole trick. No neural network, no self-play, no machine learning in any modern sense. Deep Blue was a calculator of extraordinary speed pointed at a problem that happens to reward calculation. It “knew” nothing it had not been explicitly told, and it could do nothing but play chess.
The Move That Broke Kasparov’s Nerve
The psychological turning point of the 1997 match came in Game 2. Deep Blue, instead of grabbing a pawn that virtually every chess engine of the era would have snatched, played a quiet, positional, almost human move. Kasparov was unnerved. A machine, he believed, should have been greedy; this one had shown something that looked like judgment. He later resigned that game in a position that post-match analysis suggested he could have salvaged with a draw—he had been beaten partly by his own conviction that the machine was seeing more than it was.
From there, the suspicion curdled. Kasparov suggested that the machine’s play had been too subtle, too human, and implied—without ever proving—that IBM might have had a grandmaster intervening behind the curtain. He demanded Deep Blue’s internal logs. IBM declined to hand over much, which did nothing to cool the conspiracy theories. The likeliest explanation has always been the dull one: the anomalous move was the product of a complex evaluation function and, by some accounts, a fallback triggered by a bug, not a hidden human hand.
The Controversy and the Quiet Exit
What happened next told its own story. Kasparov asked for a rematch. IBM refused, and shortly afterward dismantled Deep Blue entirely. The company had gotten exactly what it wanted—a landmark, a surge in its stock, and a permanent place in the popular imagination—and saw no upside in giving the champion a chance to even the score. For a scientific milestone, the ending was conspicuously corporate. Deep Blue retired undefeated in its final match, which is a tidy way to never lose again.
What It Was, and What It Wasn’t
It is tempting, even now, to file Deep Blue under “the day machines became smarter than us.” That reading is wrong, and the engineers who built it would be the first to say so. Deep Blue was narrow intelligence in its purest form: superhuman inside a 64-square box, useless one inch beyond it. It did not understand chess; it out-searched a human who did. The achievement was real, but it belonged to the people who designed the hardware and tuned the evaluation function, not to any spark of cognition in the machine.
The contrast with what came later is instructive. When DeepMind’s AlphaGo defeated Lee Sedol at Go in 2016, it did so in a game whose branching factor makes brute-force search hopeless. AlphaGo could not simply out-calculate the board; it had to learn a sense of position from millions of games and, in its successor AlphaZero, from playing against itself with no human games at all. Deep Blue was told what good looked like. AlphaZero figured it out. That is the line between engineering a search and growing a capability—between a system that executes human judgment very fast and one that develops judgment of its own.
What the Moment Really Proved
Deep Blue didn’t prove that machines could think. It proved something subtler and, in hindsight, more durable: that a hard human problem could be dissolved by enough specialized engineering and speed, and that the result would feel like intelligence to everyone watching. The lesson for the decades since is that the interesting question is never whether a machine can beat us at a single game, but how it got there—by being told, or by learning. Deep Blue was the last great victory of being told. Almost everything that mattered afterward came from teaching machines to learn for themselves.