In May 1997, a computer called Deep Blue defeated Garry Kasparov in a six-game match for the world chess championship. Kasparov was, by wide consensus, the greatest chess player who had ever lived. The result sent shockwaves through culture, philosophy, and cognitive science. But the most interesting thing about that moment isn’t what it meant for chess. It’s what it revealed about the shape of AI progress — a shape we’re still tracing today.

Deep Blue was not intelligent in any general sense. It evaluated around 200 million positions per second using custom VLSI chips, hand-tuned evaluation functions, and decades of human grandmaster knowledge baked into its heuristics. It couldn’t play checkers. It couldn’t recognize a photo of a chessboard. It had one extraordinary capability, purchased at enormous engineering cost, and that capability happened to be sufficient to beat the best human on earth at one specific task.

That’s the pattern. It keeps recurring, and each iteration arrives faster and at greater scale than the last.

What Deep Blue demonstrated — though few framed it this way in 1997 — is that “superhuman performance” is not a single threshold you cross. It’s a moving, task-specific frontier. Deep Blue crossed it for classical chess. AlphaGo crossed it for Go in 2016, but did so in a fundamentally different way: rather than encoding human knowledge, it discovered its own through self-play, generating strategies that professional players described as genuinely alien. AlphaZero then crossed the same frontier for chess, shogi, and Go simultaneously, starting from nothing but the rules, and it played in a style that looked nothing like Deep Blue’s brute-force search. Three different architectures, three different epistemologies, the same headline result.

The trajectory from Deep Blue to AlphaZero spans just over two decades. What changed isn’t just raw compute, though compute matters enormously. What changed is the nature of what the systems learned and how they learned it. Deep Blue’s knowledge was largely installed by humans. AlphaZero’s was self-generated. That shift — from engineered knowledge to learned knowledge — is arguably the single most consequential methodological transition in the history of the field.

And it didn’t stop at games. The same self-play and reinforcement learning logic that powered AlphaZero became a conceptual ancestor to the techniques now used to train large language models to reason, to code, to solve graduate-level mathematics. When a model today works through a multi-step proof or debugs a complex system, the intellectual lineage runs back through those game-playing systems and the insight that an AI can bootstrap capability by competing against itself in a structured environment.

Kasparov himself, to his credit, eventually arrived at a productive framing of the 1997 match. The question was never whether machines would surpass humans at chess. Once you accept that chess is a finite, deterministic game with a computable optimal strategy, the question was only when. The more interesting question — which Kasparov spent years exploring — is what humans and machines can do together that neither can do alone. His work on advanced chess, where human-computer teams competed, was an early and prescient experiment in what we now call human-AI collaboration.

What the Deep Blue moment actually proved is that the ceiling on machine performance at any well-defined task is effectively unreachable from the human side. Once a system is designed to optimize a specific objective with sufficient compute and the right architecture, human performance becomes a waypoint rather than a destination.

We’ve now internalized that lesson so thoroughly that superhuman performance on narrow benchmarks barely registers as news. The frontier has moved to something far more interesting: systems that generalize, that reason across domains, that learn continuously. Deep Blue beat a grandmaster. What comes next is harder to define, and harder still to overestimate.