To the extent that chess is a prototypical domain for AI, AI researchers are seriously deceived [compute power isn’t enough]

While thinking about GPT-3 the last two weeks I’ve also been thinking about AI and computational linguistics in general. Computational linguistics (CL) had machine translation as its first problem domain. That’s a very different kind of problem domain from chess, which has been important to AI from the beginning.

To the extent that the AI world is built on intuitions about intelligence that are based on domains like chess, those intuitions are seriously misleading.

It’s in chess that AI had its first major success, by which I mean that computers have consistently out-performed humans. AI and CL have yet to achieve that level of success with machine translation. To be sure, MT programs using machine learning and neural network techniques similar to those used so successfully with chess are the best available MT programs. They’re certainly better than the hand-coded symbolic systems of an earlier generation. But they’re not equivalent to the performance the best human translators routinely achieve or even, for that matter, second and third tier translators. The computerized system will give you a rough sense of what’s going on in a document, but no more. You wouldn’t use these systems for legal purposes or for serious translation (e.g. literary texts).

As domains, chess and MT are very different. Physically the chess world is very simple, six kinds of pieces (pawn, knight, rook, bishop, queen, king) and an eight-by-eight board. The physical appearance of the pieces is irrelevant; one could easily get by with letters: P, K, R, B, Q, K. The board is a simple matrix. The fundamental rules of play are simple as well.

The fact is, from an abstract point of view, chess is no more difficult and complex than tic-tac-toe. Given a rule for halting games play when no pieces have been exchanged, chess is a finite game of perfect information, just like tic-tac-toe. The tic-tac-toe tree is relatively small while the chess tree is huge. It’s the size of the chess tree that makes the game challenging.

As large as the chess tree is, given enough computational power, computers are able to beat humans, the very best humans, in playing the game. But when you throw the equivalent amount of computer power at translation you do not get nearly the same level of success. Why not?

There is nothing finite about the domain of machine translation. Let’s posit that words in language are equivalent to pieces in chess. While the number of words in any language is finite at any given time, it is considerably larger than the number of different kinds of pieces in chess, 10s or 100s of thousand vs. 6. While the proper use of chess pieces is simply and sharply defined, the same is not at all true for words. As for the board on which the pieces are played, for chess the board is a simple 8 by 8 matrix. For natural language the board is the whole world throughout time and space.

Sheer computational horsepower, albeit cleverly deployed, has proved successful in chess. What’s sheer computational horsepower going to get you up against natural language? Is that even a meaningful question?

No, I fear that AI’s hopes for artificial general intelligence (AGI) may well be nurtured on domains like chess. That’s a mistake and not merely a huge mistake. It’s a category mistake. The reasoning that applies to the finite chess world simply doesn’t work in the radially open domain of natural language.

The same is true for fears about malevolent computer overlords. If the world is like a chess game, then, yes, a rampaging computer chess master is a thing to be feared. But that’s not what the world is like. Not at all. In the real world, as chip designer Jim Keller has suggested, why should a superior computational intelligence even bother with us? They’ve got better things to do.