Bad predictions about chess
If you want to know where AI is headed, it’s helpful to look back to previous episodes when most of us were taken off-guard by other innovations. In particular, I think that it’s very informative to look back on what we used to believe about artificial intelligence and chess.
It may seem odd these days to bring up chess in a discussion of artificial intelligence, because chess-playing computers are not held up as examples of intelligence. But it’s interesting to remember that this was not always the case. I’m old enough to remember when we all assumed that chess is the quintessential example of an activity requiring intelligence. We quite rightly think of chess, when it’s played at the highest levels, as a game that is only played well by highly intelligent people. The game requires calculation, but also creativity, subtle pattern-matching abilities, and nuanced judgment. If that’s not "intelligence", then what is?
Indeed, it used to be a perfectly respectable view to hold that chess computers would literally never — regardless of how powerful computers might become in the future — be able to play even a competent game of chess. And the idea that a computer would be able to play at the highest levels, or even beat a world champion, was often considered to be laughably implausible.
It’s easy to forget that we had seemingly strong justification for this pessimism. It was often noted that the number of possible chess games was so astronomically huge that brute-force calculation would never succeed in powering a competent chess computer. And it was also common for people to point out that human chess experts would consider only a very small number of possible moves when they played, and that they were guided more by an ineffable "intuition" rather than the ability to calculate many moves in advance. Top chess players displayed "creativity" when they played well, and computers performing brute-force calculation would never be creative.
Now of course, we live in a world in which it is trivial to install a free chess app on your phone that plays so well, it’s capable of trouncing the best chess players in the world every single time. When Kasparov lost a match to Deep Blue, thereby demonstrating that the naysayers were all wrong, a lot of interesting things happened.
The reactions of experts in the field and the lessons learned were very surprising. First, we learned that a computer could behave "intelligently" despite not implementing the same procedures that were used by human experts. Deep Blue succeeded in defeating Kasparov by performing a huge number of calculations, whereas Kasparov and other grandmasters did nothing of the sort as they played. So intelligence need not be human-like in its implementation in order to work.
But another reaction from experts was to move the goalposts on what counts as "intelligence". The word "just" started appearing in discussions a lot, as in "the computer isn’t intelligent, it’s just performing a lot of calculations". The word "just" is doing a lot of heavy lifting when it’s used in this way. It’s being used to discount the significance of Deep Blue’s victory, and implicitly redefining what we mean by "intelligence". The mere fact that a computer is capable of playing chess was taken as decisive proof that chess does not ipso facto require intelligence. This started the trend of thinking of artificial intelligence as "whatever computers can’t do yet".
People started talking as if it had been a foregone conclusion that of course computers would eventually be able to play a very strong game of chess. But in fact, if anything, it was a minority (even a crackpot) view that computers would ever succeed in this domain.
Then many people started to worry that the game of chess would die out. After all, now that a computer could beat the best chess players in the world, what’s the point of studying to become a stronger chess player? By this line of thinking, Deep Blue would destroy anyone’s motivation to play chess. It would drop in popularity, and eventually die out.
All of these reactions were totally wrong. It’s unreasonable to redefine "intelligence" merely because a computer is programmed to do something. The right response is to say that the computer is intelligent, but only in a narrow domain. Chess remains an important example of an activity requiring intelligence, but narrowly construed.
Second, it’s not right to denigrate "mere calculation". Calculation is very powerful, and if the machine’s ability to calculate is powerful enough, and the machine is programmed the right way, calculation will buy you intelligence.
And third, what’s especially interesting and a point I’ll return to later, is that chess only became more popular following the Kasparov-Deep Blue match. In fact, the popularity of chess is at an all-time high now, even though your phone can soundly defeat Kasparov and Deep Blue and every other chess player in the world every time.
Chess in the age of deep learning
Fast-forward to December 2017. The strongest existing chess engine is Stockfish. The team at DeepMind, however, created AlphaZero, which used reinforcement deep learning (previously known as "neural networks" before the so-called "AI Winter") to learn chess. In a match between Stockfish and AlphaZero, AlphaZero won decisively, winning twenty-eight games out of one-hundred, and drawing the remaining seventy-two. Keeping in mind that Stockfish was already strong enough to beat every human chess player in the world every time, this was rather stunning.
AlphaZero was never open-sourced, but soon an open-source project called "Leela Chess Zero" was started, which aimed to use the same reinforcement learning techniques as AlphaZero to learn chess strategy. Although AlphaZero and Leela were enormously strong chess players, their strong performance wasn’t the most interesting aspect of those projects. To my mind, the most interesting thing about the games that they play is that they both seem to be deploying novel chess strategies that human players did not use. And this caused chess experts to sit up and take note.
I should make a disclaimer that I’m not an expert chess player, but I’ve been a "coffee shop level" chess player for a very long time. So I know enough about chess and chess theory to get into trouble. But I can’t resist mentioning a specific example of a novelty that seems to have been discovered independently by AlphaZero and Leela.
There are many principles, or heuristics, that guide chess players. Basic principles such as "try to control the center", "a knight on the edge of the board is weak", "don’t move the same piece twice in a row in the opening", are all good examples of simple principles that provide decent guidance to chess players (although they all have their exceptions, and the best chess players in the world will routinely violate them because their understanding of the game is so nuanced). But AlphaZero and Leela, it was noticed, started playing a very odd sequence of moves. Especially when playing with the white pieces, they’d move their so-called "h-pawn" — the pawn on the right-hand edge of the board — two squares. Sometimes, they’d follow that move by moving it again, at least once, and sometimes twice in a row.
If you’re not a chess player, take my word for it that this is a bizarre move. It doesn’t control the center, it doesn’t develop one’s pieces, it doesn’t attack the opponent’s position, it doesn’t open up possibilities for development. It’s just an obviously terrible move; if a human played it, that would be taken as proof that the person doesn’t understand chess. But strangely, it would inevitably turn out, perhaps twenty or thirty moves later, that having chosen those moves gave AlphaZero or Leela a decisive advantage. What gives?
Chess experts descended upon those games and started analyzing them to figure out why the machines had chosen to move that pawn. Without getting into a lot of unnecessary detail here, it turned out that there is an understandable rationale for the move (it restricted the opponent’s king’s mobility later in the game, among other things). So now, some human players move their h-pawn in this way.
This is just one example of the many novelties that have been discovered about chess by analyzing the games of AlphaZero and Leela (the other major novelty is that these systems place more value on piece activity than on material advantage). There’s a lesson here, and it’s an important one, in my opinion. We can use AI systems to learn about the world. That is, we can train an AI system to become superhuman at some task, and then instead of studying that task directly, we can study the AI’s output and learn about the task indirectly. This is exactly what happened with chess: chess experts have basically reverse-engineered AlphaZero and Leela in order to learn more about the game. And they have, indeed, learned a lot.
This ability of chess experts to gain a deeper understanding of the game by studying the behavior of AI systems has not diminished the pleasure of playing chess at all. Quite the opposite is true: if you already enjoyed chess, the game is even more enjoyable having learned from the AI systems. We have new appreciation for the richness of chess because of what we’ve learned recently from AlphaZero and Leela.
Better predictions about AI
I believe that these episodes tell us something about what we should expect from this new wave of AI systems.
First, the people out there who are saying that ChatGPT and its cousins are just "stochastic parrots" that are trained only to predict the next word in a sequence have totally missed the boat. Worse yet are the people who minimize this wave of AI systems by pointing out that they’re "just" doing matrix multiplication. We know from chess that mere calculation can produce abilities that should amaze us. So you’re not allowed to denigrate large language models like ChatGPT by pointing out the truism that they, in a sense, "just" do statistics or matrix multiplication. The fact is, they’re behaving in strikingly intelligent ways, even though they’re certainly not perfect. Don’t assume that there’s an obvious limit to how powerful they can become on the basis of the cheap observation that there’s matrix math at the heart of large language models. As of right now, we simply don’t know how far these techniques can be pushed. Maybe we’re at a plateau right now, and maybe the plateau is very far away; we simply don’t know yet, and anyone who claims otherwise is a candidate for the Dunning-Kruger prize.
Second, there’s a potential use for these systems which is very exciting. We can easily imagine a scenario in which an AI system is better at physics or math than any human. If this were to happen, physicists could study, for example, quantum mechanics by back-engineering the AI system in the same way that chess experts learned about chess by back-engineering AlphaZero and Leela.
It’s for this reason that I believe AI will never replace physicists. It’s not because they’ll never be as good as human physicists. I don’t know if that’s the case or not. But even if an AI were to become a vastly superior physicist, it is still fun and exciting to learn how the universe works, and people will still be drawn to physics. The same can be said of artistic endeavors; we can easily imagine a world in which the very best artists and composers are AI systems. But this won’t replace human artists and composers because art and music are very satisfying fields to engage in. It might eliminate the need to have artists who produce clip art or commercial jingles, but it won’t eliminate artists. For that matter, AI could also eliminate the need for engineers who design ball-point pens and portable sheds. But we might be in a position to actually learn new artistic techniques, principles of musical theory, engineering concepts, and so on by examining the work produced by AI.
In short, two things can be true simultaneously. First, AI could become superior to humans in all sorts of different fields, eliminating the need for many jobs. Second, there could still be plenty of people engaged in that work (either paid or not), who routinely use AI to gain a deeper understanding and appreciation for it.
Conclusion
I don’t know where these new models will plateau. My guess is that they won’t achieve so-called "general intelligence" (whatever the heck that is!) by themselves, but that they’re destined to become an important part of a larger, more complex system that does achieve general intelligence. They will, to put it another way, have a role similar to the language centers of our own human brains — that is, they will integrate, organize, and communicate complex information, and they’ll be embedded into a larger structure that’s truly intelligent.
But I’m confident that we are careening toward a world that is very different from the one we are in now. Of course, jobs will be displaced and the nature of work will be changed irrevocably. That much is obvious. But alongside those trends, the nature of learning and research will change. We will have new AI experts that teach us about a variety of domains. If I had to make a bet, I’d say that fairly soon, the very best physicists, mathematicians, composers, artists, and writers will be machines. But we will still have plenty of human physicists, mathematicians, composers, artists, and writers. They will be lifelong learners, and they will be students with AI teachers.
Leave a Reply