In the summer of 1997, Garry Kasparov sat across a board from IBM’s Deep Blue and lost the match. Most people remember that as the story — the world champion defeated, the machine triumphant. But the more interesting story is what Deep Blue actually was, what it deliberately was not, and how the gap between those two things explains almost everything about the leap AI has taken since.
Deep Blue was, by any honest accounting, a marvel of focused engineering. It evaluated roughly 200 million chess positions per second using custom VLSI chips designed for exactly one purpose: assigning a numerical score to a board state. Its evaluation function was hand-crafted by grandmasters. Its search tree was pruned by decades of accumulated chess heuristics. It was fast, deep, and extraordinarily narrow. It could not learn. It could not transfer a single insight from chess to checkers, let alone to language or protein structure or scientific reasoning. Ask it to recognize a photograph of a chessboard and it would return nothing. It was a calculator that happened to calculate the future of a 64-square universe.
That narrowness was not a bug — it was the entire point. The dominant philosophy of the era held that intelligence was best approached by decomposing it into well-defined subproblems, then engineering brutally optimized solutions to each. Chess was a subproblem. Vision was another. Language was another. The dream of a unified learner, a system that could extract general structure from raw experience and apply it across domains, was the fringe position. Most of the field considered it naive, even romantic.
The magnitude of that philosophical reversal is worth sitting with. Within roughly two decades of Deep Blue’s match, the fringe position had not just won the argument — it had eaten the field. The systems that now outperform specialist programs on specialized tasks are, almost without exception, the same systems that also write code, synthesize research, compose music, and reason through multi-step scientific problems. Generality turned out not to be a distraction from capability. It turned out to be the source of it.
What changed was learning. Deep Blue’s evaluation function was installed by humans. Modern systems derive their own evaluation functions from data, and those functions generalize in ways that nobody fully anticipated. When a large language model solves a novel mathematical proof by borrowing structural intuitions from its exposure to philosophy papers and programming tutorials, it is doing something Deep Blue could not have approached even in principle — not because it’s faster, but because it learned a representation of the world rather than a hand-coded map of one tiny corner of it.
There is something almost poignant about the 1997 match when you view it from here. Kasparov reportedly sensed something uncanny in Deep Blue’s play during the second game — a move so strange and positionally subtle that he suspected genuine intelligence behind it. The move later turned out to be a bug, a search glitch that produced an inexplicable choice when the system failed to find a preferred option. The intelligence Kasparov feared wasn’t there yet. But his instinct that he was in the presence of something genuinely new was correct. He was just 25 years early.
The distance between Deep Blue and today’s frontier systems is not primarily a distance in compute, though compute matters enormously. It is a distance in kind. One was a brilliant answer to a closed question. The other is an open learner operating on open problems. That shift — from engineered expertise to learned generality — is arguably the most important transition in the history of the field, and the 1997 match is the clearest possible before-photograph. What the after-photograph looks like in another 25 years is the question worth losing sleep over.