Kasparov on Deep Blue, AlphaZero, and Why Machines Win Closed Systems but Humans Still Choose the Questions

Lex Fridman opens by calling Garry Kasparov's 1997 defeat by IBM's Deep Blue one of the most seminal moments in the history of civilization. For centuries chess had been treated as the peak of human intellectual mastery, and the moment a machine could beat a human was, in the host's words, inspiring to an entire generation of AI researchers. Yet for Kasparov the moment read as a tragedy, something he described as physically painful.

Kasparov does not dispute the historical framing, but he corrects the personal one: the 1997 match was not the first match he lost to a machine, it was the first match he lost, period. That distinction, he says, made all the difference. What made him angry was the suspicion that the defeat was not purely the result of bad play, that factors unrelated to chess had contributed. With 22 years of distance he concedes that he played poorly and that the games, examined today, contain a large number of mistakes — but the anger came from believing that other forces were at work.

He accepts the romantic view of chess as a pinnacle of intellectual activity, a game where only the most gifted could see what ordinary mortals could not. But he reframes what the match actually proved: the machine did not have to solve chess, only to make fewer mistakes. That distinction, he argues, is the real lesson of 1997.

Kasparov invokes Claude Shannon's count of roughly 10 to the 46th power legal chess positions to show that no computer can solve the game by brute force within any relevant timeframe. The machine does not need to. It only needs to make fewer mistakes than its opponent, which is exactly what happened in 1997 and what has since happened in Go, shogi and video games.

From this he draws a general principle: in a closed system — chess, Go, any game whose rules, actions and state space are fully bounded — machines will prevail, no matter how the system is named. They do not outplay humans by being more intelligent; they outplay them by capitalizing on human error. The quantity the machine brings down is the number of mistakes.

He illustrates the resulting gap with ratings: today's strongest engines sit around 3400–3500 while Magnus Carlsen is around 2800–2815, a separation Kasparov calls insane by chess standards — comparable to the distance between a top grandmaster and an ordinary player at an open international tournament. The engine, he insists, does not understand chess better in any human sense. It simply has a steady hand.

Deep Blue was far from the first machine Kasparov faced. He played the prototype Deep Thought in two rapid games in New York in 1989 and won handily, later faced engines such as Fritz and other programs, lost games at blitz, and dropped a game in a 1994 rapid match against an engine. The 1997 match therefore arrived at the end of a long series of encounters.

He reminds listeners that he won the 1996 match against Deep Blue, a fact people tend to forget. The real mistake he and his team made was psychological: they assumed the blitz and rapid losses happened only because humans had too little time on the clock. If the match were longer, they reasoned, humans would stop making those errors.

That reasoning turned out to be nonsense. With more time, humans still made mistakes — and the machine, given more time too, stayed steady. The asymmetry is not speed but instability: humans are inconsistent, machines are not. That realization is what eventually moved the conversation away from human versus machine.

After losing the rematch, Kasparov spent a year ruminating on what had happened and concluded in 1998 that the framing had to change: it was no longer human versus machine but human working with machine. He still believed humans could play against machines, and he demonstrated it by playing two more matches in 2003 against Deep Fritz and Deep Junior, both of which ended in draws.

Those opponents were at least as strong as Deep Blue, and probably stronger. Today, he notes, an app on a mobile phone is probably stronger than Deep Blue, and dedicated chess engines are far stronger again. He adds wryly that anyone who runs the 1997 games through a modern engine will find the machine laughing at them.

Within 30 seconds, he says, Stockfish, Houdini or Komodo will show how many mistakes both Kasparov and Deep Blue made in a game that commentators at the time praised as a great chess match. The reputation of the match does not survive contact with a modern engine.

The host pushes back: if chess was wrongly treated as the epitome of intellect, what about domains such as language and conversation that still seem out of reach for computers? Kasparov does not answer with a list of protected human skills. His answer is structural — the distinction between closed and open-ended systems.

The size of the model is not the issue, he says; the issue is whether the system is closed. In a board game the rules, the actions and the state space are all constrained. Open the system up and the machine's effectiveness changes fundamentally, because it has no sense of when it is reaching territory of diminishing returns.

Put differently, the machine does not know how to ask the right questions. It can ask questions, but it will never tell you which questions are relevant. That is a matter of direction rather than horsepower, and it marks out the territory that remains human.

Kasparov acknowledges that many people are uncomfortable watching the territory that belongs to humans shrink. He expects the human contribution to be pushed into the last few decimal points, and offers an analogy: a very powerful gun where the human can only alter the direction of the bullet by a tenth of a degree — which, a mile away, still moves the impact by ten meters.

Recognizing which human qualities machines cannot reproduce in the foreseeable future is therefore essential, because the effectiveness of the collaboration depends on knowing exactly what each side brings. His warning is that the greatest danger arises when humans try to interfere with a machine's superior knowledge. He gives radiology as an example: in reading scans you may prefer an experienced nurse over a top professor, precisely because the nurse will not try to override what the machine already understands better.

His rule of thumb is that if machines do something better across 95 or 96 percent of the territory, humans should not touch that territory at all. The task is to recognize where human judgment still makes the difference, as chess players have had to learn.

Turning to AlphaZero, Kasparov argues that almost everything called AI today is a variation of what Claude Shannon characterized as brute force. Whether it is Deep Blue or the modern technologies presented as AI, it is optimization: machines finding better ways to process human-generated data.

AlphaZero is a first step into something different — machine-produced knowledge. He finds it ironic that IBM had already been there: an early-1990s IBM backgammon program contained an AlphaZero-like element, a system trying to come up with its own strategies. Because Deep Blue succeeded, that project was quietly put on hold rather than pursued.

He calls machine-generated knowledge revolutionary while insisting that many open-ended questions remain. What struck him about AlphaZero's chess was watching it handle all the pieces at once and play positions with a broken material balance, sacrificing material and staying a move or two ahead of other programs.

Kasparov describes AlphaZero's play with evident enthusiasm: it juggles the pieces, plays positions with a broken material balance, sacrifices material, and stays one or two moves ahead of rival programs. The advantage is not calculation, since the other engines calculate at least as well. It is knowledge built from patterns discovered across roughly 60 million games, and it looks like intuition.

Then comes the problem. Suppose AlphaZero faced a stronger opponent, or a powerful computer assisted by a human who helps discover specific weaknesses. If those weaknesses were exposed, how many games would it take AlphaZero to correct them? The answer, Kasparov says, is hundreds of thousands — even while it keeps losing — because the entire system rests on massive self-play rather than targeted repair.

A human, by contrast, can be helped by just a few tweaks. That is where humans remain more flexible, and it is why the most valuable part of the collaboration lies in identifying what each side can contribute. Understanding that division of labor, Kasparov concludes, is what will define the next steps in human-machine collaboration.