Deep Thinking: Where Machine Intelligence Ends and Human Creativity Begins

by Garry Kasparov

Trial and error in programming techniques and the relentlessness of Moore’s law produced chess machines that played at the level of the top 5 percent of human players by 1977, expert level. They still played terrible chess, full of illogical moves even a weak human would never consider. But they were becoming fast enough to cover up these occasional blunders with accurate defense and sharp tactics while playing against humans.

Ferrucci’s interjection during my lecture cut to the core of the matter as effectively as Picasso and Douglas Adams. He said, “Computers do know how to ask questions. They just don’t know which ones are important.”

White moves first in chess and, at least at the expert level, this confers an advantage similar to that of serving in tennis. White wins about twice as often as black at the professional level, although half of all games finish drawn. White can usually define the battleground and I used it to offer Deep Thought a “poisoned pawn” in the opening, a tempting offer of material that computers were still too eager to grab. Sure enough, the machine took the bait and was soon in great difficulties as my pieces swarmed over the board. My attack on its king forced it to give up its queen on move seventeen and after that it was a mop-up operation. Any human would have resigned against me in good conscience at that point, but machines don’t have to worry about that. Their operators usually figure they have nothing to lose by playing on even when the machine’s evaluation shows that it knows it is completely lost. Considering how tricky computers can be against humans, this isn’t unreasonable, only annoying.

KEN THOMPSON designed the revolutionary chess machine Belle, whose chips Deep Blue’s were based on, while working at Bell Laboratories in New Jersey, the famous “Idea Factory” that did pioneering work on breakthroughs in everything from solar cells and lasers to transistors and cell phones. Thompson was also the principal inventor of the ubiquitous Unix operating system while there, which is the basis for what runs Apple Macs, Google Android, and the billions of devices and servers running Linux.

So, when we say a GM has a rating of 2700, that is the balance of his performance over hundreds of games. It’s a very small margin for error, with the exception of very young players and a tiny handful of wildly inconsistent GMs. Chess machines aren’t like this at all. When I was asked after the first match about Deep Blue’s strength, estimated by the result at 2700, I said, “Yes, 2700 maybe, but 3100 in some positions and 2300 in others.” In sharp, tactical play, Deep Blue could be counted on to perform far above even my level 2800+ level. This was true even of the PC engines that were still relatively weak at the time. In closed maneuvering positions, where Deep Blue’s powers of calculation were muted, it could make strange and pointless moves that even a weak human master would never consider on general principles. Its evaluation ability was weak overall and, in some areas, such as a few I exploited in our match, it was terrible.

Carlsen is comfortable with his own strengths and sees the machine appropriately, as a tool, not an oracle. This helps him in training because he builds up critical mental problem-solving muscles instead of simply having the answer handed to him by an engine. It also helps him at the board, because when he needs to solve a tough problem he doesn’t mentally reach for the mouse.

play. I would find inspiration not at the board, however, but in my team’s apartment in lower Manhattan. I came up with an amazing piece sacrifice to use against Anand’s preferred defense against the king’s pawn, the Open Ruy Lopez. My team and I spent the entire weekend hammering out the incredibly complex tactics that followed the sacrifice, and here the machine was quite useful, even as relatively weak as engines were back then.

After eight straight draws to start the match followed by a win, Anand lost four of the next five, essentially ending the match with six games still remaining. Anand did not suddenly become a much worse player after game ten, nor did I become a much stronger one. And as much credit as I would like to give my team and me for our opening surprises, they weren’t what made Anand play far below his usual level for that horrible stretch. He lost one game to a strong novelty, was faced with a surprising defense in the next game and blundered, and was never able to recover the composure required for consistency. In a way, it was lucky I hadn’t found that strong novelty in my pre-match preparation. Had the fateful tenth game instead been the second game, he would have had time to recover. This isn’t a critique of Vishy Anand; it’s a critique of humanness. A similar syndrome would envelope me eighteen months later in my rematch with Deep Blue, and being aware of it was useless in countering its effects.

Some players actually seem to play better when they are on the ropes. They dig in and defend like tigers, taking it as a challenge to rise to. Viktor Korchnoi was like this; he enjoyed grabbing a pawn even if it meant weathering a brutal attack. A man who had survived the Siege of Leningrad as a boy was not going to be intimidated at the chessboard. This sort of mental robustness is rare, even among elite Grandmasters. Mistakes almost never walk alone.

Intuition is the product of experience and confidence. And here I mean “product” in the mathematical sense, as the equation intuition = experience x confidence. It is the ability to act reflexively on knowledge that has been deeply absorbed and understood. Depression short-circuits intuition by inhibiting the confidence required to turn that processed experience into action.

All the data in the world won’t help you overcome your biases if you don’t listen to it. Stop making excuses and rationalizations that are only your mind tricking you into doing what it wants to do. It can be hard to let the data speak for itself. After all, we aren’t machines.

If you remember Moravec’s paradox, it says that what machines are good at is where humans are weak, and vice versa. This is well illustrated by chess, and this gave me an idea for an experiment.

A clever process beat superior knowledge and superior technology. It didn’t render knowledge and technology obsolete, of course, but it illustrated the power of efficiency and coordination to dramatically improve results. I represented my conclusion like this: weak human + machine + better process was superior to a strong computer alone and, more remarkably, superior to a strong human + machine + inferior process.

I have argued that our technology can make us more human by freeing us to be more creative, but there is more to being human than creativity. We have other qualities the machines cannot match. They have instructions while we have purpose. Machines cannot dream, not even in sleep mode. Humans can, and we will need our intelligent machines in order to turn our grandest dreams into reality. If we stop dreaming big dreams, if we stop looking for a greater purpose, then we may as well be machines ourselves.