Frank
IBM's chess supercomputer Deep Blue beat the world champion, then immediately became useless, and that wasn't a bug — it was the whole problem with AI that nobody wanted to say out loud.
The Story
Deep Blue beat Garry Kasparov in 1997. IBM threw a party. The headlines called it a milestone for human achievement. Then IBM quietly dismantled the machine and never let it play again.
That wasn't modesty. The reality is, Deep Blue could do exactly one thing. Chess. Not checkers. Not tic-tac-toe. Chess. You couldn't ask it a question. You couldn't point it at a different problem. It was the most expensive, most celebrated one-trick pony in computing history. The thing that made it unbeatable in its lane made it completely useless everywhere else. Every ounce of its intelligence was baked into a domain so narrow you could measure it in 64 squares.
The engineers knew this. They called it narrow intelligence, and they knew even as the world was applauding that they hadn't actually solved anything. They'd built a very fast calculator with a very specific rulebook. The moment you changed the rules, Deep Blue was a paperweight. A famous, celebrated, completely helpless paperweight.
What got buried in all the celebration is that Kasparov said something that nobody really picked up on. He said the machine didn't understand the game. It calculated it. There's a canyon between those two things, and that canyon is basically the entire unsolved problem of artificial intelligence.
The Hidden Principle
Optimization and understanding are not the same thing. A system can be trained to perform at a world-class level inside a fixed environment and be completely blind the second that environment shifts even slightly. That's not intelligence. That's pattern matching at scale. Fast, yes. Powerful, yes. Intelligent in any general sense, no.
The deeper principle is this: the narrower the domain, the more dangerous it is to confuse performance with capability. Deep Blue performed. It did not understand. That distinction matters enormously because systems built on performance metrics will always look brilliant right up until the moment the rules change.
What This Means Today
Y'all are watching this play out right now with every large language model getting deployed into enterprise software. These things are shockingly good inside the context they were trained on. Ask one to summarize a contract or generate boilerplate code and it hums. Shift the context slightly, introduce something genuinely novel, something outside the training distribution, and it either halts or starts confidently making things up. That's Deep Blue in a suit.
The companies building on top of AI right now are making the same mistake IBM's audience made in 1997. They're watching the performance and calling it understanding. They're handing over decisions to systems that are, at their core, very fast pattern matchers operating inside invisible walls. The wall doesn't announce itself. You only find it when something important breaks.
A system that wins every game inside one room isn't ready for a building with a hundred rooms.
