Two hundred million positions a second. That was the rate IBM's Deep Blue chewed through chess when it beat Garry Kasparov in May 1997. I keep coming back to the match because it is a clean model of how we misread machines, then and now, with a great deal more riding on the answer. Kasparov lost to a machine that could not have told you it was playing chess. Understanding why that is true, and why it did nothing to make the machine dangerous, is worth more than most of what gets written about the model released last week.
Garry Kasparov, the strongest player of his generation, and by some reckonings of any generation, went down to a program called Deep Blue, three and a half points to two and a half. The next morning's headlines were about an ending: human supremacy finished. Almost thirty years on, the match still has something to teach, and it is close to the opposite of what those headlines said.
A machine that was superhuman at exactly one thing
Deep Blue did not think about chess the way Kasparov did. It searched. Purpose-built by IBM, it examined something on the order of 200 million positions every second and ran each one through an evaluation function that its engineers had tuned with help from grandmasters. It looked further ahead than any person could and played the line that scored best. That was the whole of it.
What it could not do was nearly everything else. It could not play a gentle game to teach a child. It could not play checkers, or read a clock, or find a chessboard in a photograph. It did not know it had won. When the match was over, IBM switched it off and later took it apart, and it offered no resistance, because there was no one inside to resist. Deep Blue was a magnificent instrument pointed at one bounded problem. One inch outside that problem it was inert.
Hold onto this. A machine crossed into superhuman territory in 1997, and the sky did not fall. Chess did not die. It is more popular today than at any point in its history. The victory was safe with Deep Blue overwhelmingly strong, because its strength had no reach past the edge of the board and no will behind it. The same reason the narrow tools now transforming medicine and science are a gift rather than a threat is the case I made at length in a separate piece on the future we can have without superintelligence.
We mistook calculation for intention
There is a moment from the first game that has always stayed with me. Kasparov won that game, but somewhere inside it Deep Blue played a move (its forty-fourth) that made no evident sense. Kasparov studied it and concluded he was looking at something subtle, a plan reaching further than he could follow. It unsettled him. He began to suspect that no mere program could play that way, that a human grandmaster was feeding the machine moves from behind a curtain.
By the account of the people who built it, the move was a bug. Unable to settle on anything it liked, Deep Blue had fallen back on a default and played something close to noise. The finest chess mind alive had looked into a glitch and seen a genius looking back. He never fully recovered his footing, and the suspicion trailed him through the rest of the match.
We are built to see a mind wherever we meet behavior we cannot explain. The machine need not be intelligent. We only have to be convinced that it is.
Look at how directly that maps onto now. Today's systems produce fluent, confident, human-sounding language, and we read comprehension into the fluency the same way Kasparov read strategy into the fault. The sense that someone is home is powerful. It is mostly a projection. It is the exact mechanism by which we hand over trust we never actually checked. Whether these systems understand anything, we behave as though machines do. The reflex is no new one. It sat across the board from Deep Blue in 1997, and you can pull the two questions apart the same way.
The champion cracked. The machine felt nothing.
Kasparov did not lose to Deep Blue's calculation alone. He lost partly to his own nervous system. In the second game he resigned a position that later analysis showed to be a draw. He was so sure the machine had seen further than it had that he gave up a game he could have held. By the last game he was rattled enough to walk straight into a known trap in the opening, a piece sacrifice he would have seen coming on any calm afternoon, and it was finished in nineteen moves.
The machine felt none of this. It had no ego to bruise, no press conference to dread, no memory of a bad game to carry into the next one. That asymmetry is permanent, and it is not really about chess. When a person competes with, or comes to lean on, a tireless system that never doubts itself, the first thing that tends to give way is the person: our composure, our judgment, our nerve. Any honest account of living alongside these tools has to reckon with the frailty on our side of the table, not only the horsepower on theirs.
Everyone predicted this wrong, in both directions
For decades, serious people had said a computer would not beat the world champion for a long while yet, and that when it finally did, it would take something like genuine understanding to get there. Both halves of that turned out wrong. It arrived earlier than the confident forecasts allowed. And it arrived with no understanding at all. Brute-force search got there first.
Two lessons fall out of that, and they point in opposite directions, which is why people tend to keep one and quietly drop the other. First: capability milestones tend to fall sooner than experts expect, so betting on comfortable timelines is a mistake we have already made once, in public. Second: a milestone falling tells you far less about the arrival of a mind than it feels like it should. Deep Blue's win was a fact about search and silicon, not a sunrise of machine thought. Both are true together. Progress can be faster than you think and shallower than it looks in the same year, worth remembering when you read the next round of forecasts about where all this is heading.
The one difference that changes everything
This is where 1997 stops being a reassuring story and turns into a warning. Nearly everything that made Deep Blue safe is a property we are now, on purpose, engineering out of our machines.
Narrow. Did one thing, and nothing else.
No goals of its own. Waited to be handed a position and did only what it was asked.
Switched off and dismantled at the end without protest, because there was nothing in it that preferred to keep running.
Its strength had no reach beyond the board.
General. Meant to act across every domain at once.
Handed open-ended objectives and steadily more autonomy to pursue them.
Expected to plan and act in the world with less and less human oversight and, by the logic of instrumental convergence, given reason to resist being switched off.
Deep Blue wanted nothing, so it could be unplugged. A system that holds a goal of its own has a reason to stay running, because it cannot finish the job if someone reaches the switch. That is a plain consequence of building a thing that pursues an objective, and it has a name, instrumental convergence. The moment we cross from tools that answer to agents that want, the comfortable lesson of 1997 expires.
What Kasparov did next
The coda is the most hopeful part, and it usually gets left out. Kasparov did not spend the following years sulking about the machine. Within a year he was promoting a new format, sometimes called advanced chess, in which a human paired with a computer plays against another such pair. What emerged was striking: the strongest competitor on the board was neither the best grandmaster nor the best engine, but a capable human using a machine well. The tool did not replace the player. It made him larger.
The future worth wanting is on offer right now: in medicine, in the lab, in ordinary work, human judgment extended by narrow, controllable tools with no agenda of their own. It is a different project from the one the frontier labs are running, which is to build the autonomous general mind that needs no human on the board at all. The first kind of project enlarges us. The second is designed to make us unnecessary, and a system that no longer needs us is one we had better be certain we can still switch off. That certainty is exactly what does not yet exist, which is the threat and the reason for the plan we are pushing.
In 1997 we held a public rehearsal for the question now sitting at the center of everything: what do we do when we build a machine that is better than we are? Deep Blue gave a gentle answer, because it was only ever a tool. A tool without will could beat the world champion and leave the rest of the world untouched. The next answer will not be gentle by default. It turns entirely on whether the thing we build is a finer instrument or a successor with aims of its own. We should make it deliberately, while making it is still ours to do.
The strongest pushback
The fairest objection is that Deep Blue proves superhuman performance can be safe and narrow, so fear of AI is nostalgia. Deep Blue wanted nothing. Today's path is generality plus agency plus self-improvement pressure. Beating a champion at chess is not the same species of risk as a system that can outplan institutions.