How did IBM's Deep Blue beat world chess champion Garry Kasparov in 1997?

Chess had been a target for AI from the very beginning. Claude Shannon wrote about programming a computer to play it in 1950, and for decades a win over the world champion was treated as a milestone. On 11 May 1997, it happened. Garry Kasparov, the reigning world champion, resigned the sixth game of his rematch with IBM's Deep Blue after fewer than twenty moves.

The result, 3.5 points to 2.5, made Deep Blue the first computer to beat a reigning world champion in a match played under standard tournament time controls.

Who built Deep Blue

The machine grew out of a student project. In 1985, Feng-hsiung Hsu, a graduate student at Carnegie Mellon University, helped build a chess machine called ChipTest, which took the 1987 North American title for computer chess programs. With Murray Campbell, a computer scientist and strong chess player, and other colleagues, Hsu went on to build Deep Thought, the first program to beat a grandmaster.

In 1989, IBM Research hired Hsu and Campbell to build a chess computer that could beat the best human in the world. Other IBM scientists joined the effort, among them C. J. Tan, Joe Hoane and Jerry Brody. The new name, Deep Blue, played on Big Blue, the company's nickname.

Two matches, one year apart

The first match took place in February 1996 at the Pennsylvania Convention Center in Philadelphia. Deep Blue won the very first game, the first time a computer had beaten a reigning world champion in a game with classical time controls. But Kasparov adapted, winning three games and drawing two to take the match 4 to 2.

The IBM team spent the next year improving the machine. They strengthened its databases for endgames, gave it a more powerful way of scoring positions, hired more grandmasters as advisers, and found ways to make its strategy harder to read.

The rematch was played in May 1997 at New York's Equitable Center, with the world's media watching:

Game 2 shook Kasparov. Deep Blue declined to grab material and instead played quiet, positional moves that he felt were too subtle for a computer, and he suggested afterward that a human might have intervened. Analysts later found that he had resigned a position he could have drawn. In game 6, with the score level, Kasparov chose an opening line he hoped would push Deep Blue out of familiar territory. The machine answered with a known knight sacrifice that broke his defense.

Prize money came from IBM: $700,000 for the winner and $400,000 for the loser, and the machine's winnings simply returned to IBM. The team also received a $100,000 prize created at Carnegie Mellon by Edward Fredkin in 1980 for the first program to beat a reigning world champion.

How Deep Blue played

At heart, Deep Blue worked like most chess programs: look ahead and score the results. Its edge was raw speed. It used 32 processors working in parallel and could evaluate about 200 million chess positions per second. IBM puts its speed at 11.38 billion floating-point operations (calculations with decimal numbers) each second.

Here is the basic loop, in simplified form:

1. List every legal move in the current position.
2. For each move, imagine every reply, then every reply to those replies, as deep as time allows.
3. At the end of each line, score the position with an evaluation function, a formula that weighs material, piece activity, king safety and similar factors.
4. Assume both sides choose their best options, and work back up the tree to find the move that leads to the best guaranteed outcome.
5. Play that move.

Speed alone was not enough. For the opening, Deep Blue relied on prepared books; in the 1996 match it also had an "extended book" that scored moves using a database of 700,000 earlier grandmaster games. For the endgame, it had databases of solved positions. And the evaluation function itself was tuned with the help of grandmasters. As IBM's C. J. Tan put it, Kasparov had prepared to face a computer, but "we programmed it to play like a grandmaster."

Why it mattered

Kasparov said he had seen "something similar to an artificial intellect" for the first time. Newspapers treated the match as a contest between humanity and machines, and Deep Blue became a celebrity. The machine itself was later retired to the Smithsonian in Washington, D.C.

IBM applied lessons from Deep Blue's parallel design to problems such as financial modeling, data mining and drug discovery, and went on to build Blue Gene and Watson. In the game world, attention shifted to Go, which has far more possible moves than chess and resisted brute-force search. It took until 2015 and 2016 for DeepMind's AlphaGo to beat Fan Hui and then Lee Sedol, and it did so with learning rather than mostly hand-built knowledge.

Was Deep Blue intelligent, or just fast?

Both answers have defenders. Deep Blue could not explain its moves, learn from a single game the way a person does, or do anything except play chess. Its strength came from searching far more positions than any human could, guided by rules that people wrote. In that sense it was a superb calculator.

But the line between "just fast" and "intelligent" is blurry. Kasparov's own reaction shows that, from the other side of the board, deep calculation can look like insight, as in the quiet moves of game 2. Later commentary also tended to give some credit for the result to Kasparov's uncharacteristic mistakes.

A useful way to see it: Deep Blue proved that a narrow task long thought to need human genius could be mastered by search plus expert knowledge. It did not show how to build a machine that learns, which is the question most AI research is pursuing now.