Deep Blue vs Kasparov
♟️ Deep Blue vs Kasparov
🖥️ How Deep Blue Actually Worked
IBM's Deep Blue was a massively parallel supercomputer built around custom "chess chips," capable of evaluating roughly 200 million chess positions per second. Its core technique was brute-force minimax search with alpha-beta pruning — systematically looking ahead through the tree of possible future moves, several moves deeper than any human could consciously track, and pruning away branches that couldn't possibly affect the outcome. Grandmaster Joel Benjamin helped hand-tune its position-evaluation function (the formula scoring how good a given board looks), and the system was loaded with a large database of opening moves and endgame tablebases covering positions with few pieces remaining.
Notice what's not in that description: no learning from a dataset of past games, no statistical inference, none of Chapter 1's machine learning techniques. Deep Blue's chess knowledge was searched and hand-tuned, not learned — in spirit, it's a direct descendant of Course 2 Chapter 3's Logic Theorist and General Problem Solver, doing tree search over a well-defined problem space, just with four decades of computing power behind it instead of 1950s hardware.
📅 May 11, 1997
Deep Blue defeated Garry Kasparov, the reigning world chess champion, in a six-game rematch (Kasparov had won their first match in 1996), with a final score of 3.5–2.5. It was the first time a computer had beaten a reigning world champion in a full match under standard tournament time controls — a genuinely historic, widely reported result that made front-page news around the world.
Kasparov later alleged that a mysteriously "too human" move IBM's system played in Game 2 suggested outside human intervention during the match. He formally requested a rematch; IBM declined and dismantled Deep Blue shortly afterward, rather than defending the result or allowing a follow-up match. The dispute was never conclusively resolved, and it left a lingering asterisk over an otherwise landmark achievement.
🧠 Moravec's Paradox
Deep Blue's victory is best understood through an insight roboticist Hans Moravec articulated in 1988, now known as Moravec's Paradox: tasks that feel intellectually demanding to humans — chess, formal logic, calculus — tend to be computationally cheap for machines, while tasks humans perform effortlessly and unconsciously — recognizing a face, walking across a room, understanding a spoken sentence in a noisy room — tend to be extraordinarily computationally expensive. Chess felt like the ultimate test of machine intelligence for decades precisely because it's hard for humans. For a sufficiently fast search algorithm, it turned out to be one of the more tractable problems available.
Chess has a small, fully-defined rule set, a fully observable board, no hidden information, and a state space that — while enormous — is small enough for search plus raw computing power to conquer. That's exactly the profile of a closed, well-defined domain that Course 2 Chapter 4's compare-table already flagged as symbolic search's genuine strength. Many contemporary AI researchers pointed out that Deep Blue's win, however impressive as engineering, demonstrated brute-force search scaling with hardware — not a new kind of machine understanding. It's the Mechanical Turk's question again, from Course 1 Chapter 3: a sufficiently convincing performance gets read as "real" intelligence by the public, regardless of what's actually happening underneath.
📜 Three Ways to Win at a Game
| Approach | Knowledge Source | Where It Appears |
|---|---|---|
| Symbolic tree search | Hand-tuned evaluation function + brute-force lookahead | Logic Theorist / GPS (Course 2, Ch. 3); Deep Blue (this chapter) |
| Statistical learning | Patterns inferred from a fixed dataset | Bayesian networks, SVMs, statistical MT (Chapter 1) |
| Learning through self-play | A system generates its own training data by playing itself | Not yet covered — this is what makes Chapter 5's AlphaGo different from both |
Deep Blue's trick — throw enough hardware at exhaustive search — worked for chess because chess's state space, though vast, was small enough for 1997-era computing power to search deeply enough to win. It would not work nearly as well for the ancient game of Go, whose state space dwarfs chess by many orders of magnitude, making pure brute-force search hopeless no matter how much hardware you throw at it. Chapter 5 covers what it actually took to beat a top human at Go in 2016 — and the answer looks far more like Chapter 1's learning-from-data philosophy than this chapter's brute-force search.
🤔 Questions to Sit With
Moravec's Paradox suggests our intuitions about what's "hard" for a machine are often backwards — chess fell decades before basic vision or common sense. Can you think of another task widely assumed to require "real intelligence" that might turn out to be more tractable than it looks, once you know what's actually happening underneath?
IBM declined Kasparov's rematch request and dismantled Deep Blue immediately after winning. Does that decision change how you read the result — does refusing a chance to confirm the win under scrutiny make the achievement more or less convincing?
Many AI researchers considered Deep Blue impressive engineering but not a meaningful step toward general intelligence, while the public largely read it as a landmark "machines beat humans" moment. When expert and public interpretations of an AI milestone diverge this sharply, whose read do you think history tends to remember — and should it?
🎯 What's Next
Next chapter: The Neural Network Revival — the backpropagation algorithm, and why deep learning had to wait years after the underlying math was already known for enough data and compute to finally catch up to it.