Checkmate:
Can Artificial Intelligence Win the Long Game?

 

edward tenner, a frequent contributor to the Review, is a research affiliate of the Smithsonian Institution and Rutgers University. Illustrator stuart briers created the illustrations for this article using the AI software Midjourney.

Published July 24, 2023

 

Are we at the dawn of a new age in which artificial intelligence will transform not only production, commerce and communication but also domains of thought once considered limited to human geniuses? One case study offers some hints.

In the history of computing, perhaps the most widely tracked benchmark of excellence has been in an ancient game with an unrivaled searchable historic database of 9.2 million matches (and counting). It’s a contest of pure skill with no element of chance, a game both of global understanding and of passionate nationalism. I’m talking about chess, of course.

Some Ancient History

Chess has been linked to computer science ever since the latter’s birth as a discipline. Alan Turing, the mathematician whose PhD dissertation built the foundation of modern computer theory, prized the game as a test of the power of algorithms even before technology existed that could execute them. In 1951, he used a printout of his code to play a virtual computer against a human opponent. It took 30 minutes for each move and the algorithm lost — but Turing had proved that a program could play a game to the end. (Decades later, world champion Garry Kasparov defeated the program in just 16 moves but called the program “an incredible achievement.”) Within two decades, though, hardware was beginning to catch up with Turing’s vision and that of later artificial intelligence pioneers.

Enter Gordon Moore, cofounder of Intel, who predicted in 1965 that the number of electronic switches (transistors) that could be packed onto a microprocessor would double every 24 months with commensurate improvement in execution speed. He created one of the most formidable and longest-running challenges in the history of technology. And while Moore was hardly the first tech entrepreneur to make bold predictions about the wonders awaiting us, his landmark paper was the first to foresee growth in exponential terms — and on a timetable previously unimagined.

Napoleon’s armies, after all, had covered no more distance per day than Caesar’s almost two millennia earlier. And in the modern era, the pace of rail and auto travel rapidly reached practical plateaus. Even jet airliners’ speeds hit a wall within decades of the first flight, as the commercial failure of the supersonic Concorde in 2003 would demonstrate.

Intentionally or not, Moore’s prediction became what has often been called, in the phrase of the sociologist Robert Merton, a self-fulfilling prophecy. It implied that there was a solution for every physical and chemical obstacle that skeptics raised about it over the years. As one barrier after another fell to the ingenuity of Intel and its rivals, it seemed reasonable to conclude that if you didn’t see a way forward you weren’t thinking hard enough. It’s only fitting, then, that Intel became the first corporation to sponsor a world championship chess match, when Garry Kasparov and Nigel Short founded a shortlived Professional Chess Association as a rival to the traditional International Chess Federation in 1993.

 
IBM proved a natural successor in chess sponsorship since it excelled at the design of single purpose hardware along with the software to animate it. IBM’s program Deep Blue, running on a dedicated IBM-built supercomputer, made history when it defeated Kasparov in 1997.
 

Chess became even more important for the branding of Intel’s iconic competitor. IBM proved a natural successor in chess sponsorship since it excelled at the design of singlepurpose hardware along with the software to animate it. IBM’s program Deep Blue, running on a dedicated IBM-built supercomputer, made history when it defeated Kasparov in 1997. While the foray into chess was largely a public relations project, the mathematics of Moore’s Law suggested that, before too long, Deep Blue’s muscle would trickle down to PCs.

Every so often in the years since, there have been hints that the engineers were falling behind Moore’s Law. Yet Moore’s prediction survived its apparent obstacles. Indeed, an homage to the 75th anniversary of the transistor in Science magazine pointed to new ways forward in speeding computation that no longer depended on increasing the number of processors that could dance on the head of a silicon chip — for example, designing processors to work faster and offloading tasks to specialized companion chips like the graphics cards that have proved so useful in everything from video to cryptocurrency mining.

Game of Kings and Transatlantic Nerds

The 21st century has proved the point. More efficient code multiplies the impact of processor speed in enabling computer power, and chess once more has been a landmark test bed. The most important force in this movement emerged 6,000 miles from Silicon Valley in a British artificial intelligence startup called DeepMind. Its algorithm, unlike that of Alan Turing’s original program and its successors, was modeled after human learning: the neural network.

DeepMind was still deep in red ink when Google (now Alphabet) acquired it in 2014. By 2020, it was turning a profit. Meanwhile it had created a sensation by mastering the game of Go.

In the 1980s, one of my advisors on acquisitions at Princeton University Press was the Nobel Laureate physicist Philip Anderson. A Go enthusiast, he believed correctly that the techniques making progress in chess would be ineffective in mastering Go, which Deep- Mind’s website points out is 1010 (you got it: a googol) more complex. When Deep Blue defeated the world chess champion in 1997, the most powerful Go program was far from ready for prime time.

The secret sauce of AlphaGo, DeepMind’s Go program, was its ability to learn by playing billions of games against itself to learn winning strategies. Under Alphabet auspices, the program rapidly achieved master level. By 2019, Lee Se-dol, the 18-time world Go champion from Korea, announced his retirement in the face of “an entity that cannot be defeated.”

The conventional open-source program Stockfish was already beyond human grandmasters’ skills when neural networks appeared on the chess scene. Magnus Carlsen, the current world chess champion, and Gary Kasparov, the former world champion, are two of the strongest players who have ever lived, with ratings of 2882 and 2851, respectively, on the ELO scale, which is based on results of tournaments. Stockfish surpassed them at ELO 3531, and it has plenty of company. There are now some 125 chess programs (aka “engines”) that have higher ratings than the flesh-and-blood champs.

Thus the triumph of a spinoff of AlphaGo, AlphaZero, attracted more than a little attention by defeating Stockfish, the gold standard of chess play in December 2017. Garry Kasparov told the website Chess.com that Alpha Zero “approaches the ‘Type B,’ human-like approach to machine chess dreamt of by Claude Shannon and Alan Turing instead of brute force.”

In the Machine Age

Where is chess now, 25 years since Big Blue’s defeat of Kasparov and five years since AlphaZero’s rout of Stockfish? For answers, I turned to a uniquely qualified expert: Jon Edwards, a former Princeton information technology administrator who has taught and coached 5,000 players and who in 2022 won the world championship of the International Correspondence Chess Federation, the first American in 40 years to hold the title.

Once dependent on postal services, correspondence chess came into its own with email and the web in the 1990s. It is a deliberative subculture within the game, without conventional time pressure but with an honor code of independent play that many felt had become impossible to enforce as chess engines fit for PCs and middle-class budgets gained prowess.

 
Face-to-face contests at tournaments still bar the use of digital paraphernalia in real time — AlphaZero is not allowed on the premises. But now that engines and databases of games are so fully integrated into the chess ecology, over-the-board players can only remain at the top by using digital tools in their preparation.
 

Before the turn of the millennium, the ICCF’s rules were firm: using electronic aids was cheating. But the dam broke. When Jon Edwards returned to tournament play after a five-year hiatus, he lost game after game to competitors who had adapted to the newly permitted digital tools before he decided to join the crowd.

Face-to-face contests at tournaments still bar the use of digital paraphernalia in real time — AlphaZero is not allowed on the premises. But now that engines and databases of games are so fully integrated into the chess ecology, over-the-board players can only remain at the top by using digital tools in their preparation. The leading pros’ teams thus include technology experts as well as master sparring partners.

One consequence of the tsunami of computer resources available for over-the-board play, according to Edwards, is that far fewer standardized sequences of opening moves remain competitive. Of course, some openings were always considered stronger than others. But the best players were still able to win with weaker ones by using online databases to find promising follow-ups. Now those opportunities are largely blocked, at least at the highest levels of play where human and machine work hand in hand.

Over-the-board players spend hours with chess engines before each tournament. They study likely opponents’ games, looking for Achilles heels. Computerization lets both sides prepare much more systematically than they could before the turn of the millennium, so there are few surprises in the openings. Chess engines have also studied endgames exhaustively — I’ll say more about that later — so the middle-game has become the most contested ground.

A brief digression: do higher ELO scores imply that chess engines know the game of chess better than the human grandmasters they outrank? Edwards cites a key difference between human and machine. Each of the giants profiled by Kasparov in his five-volume analysis of chess greats had a clear set of strategic principles and an ideal position in mind. While chess engines could defeat them, the engines’ intelligence is savant-like, lacking an understanding of how they achieve their victories.

There are implications here for the use of chatbots in education and business. It is now possible for chatbots like ChatGPT to generate impressive performances — especially if the inevitable blunders are edited out — without being able to explain the principles leading to successes. AI researchers have a name for this: the interpretability problem. Edwards and other top correspondence chess players do have this understanding. And while the machines make it easier to become a good player online, the same doesn’t apply to great play.

Edwards says the latest chess engines “have basically created a very high barrier to entry.” Only the most affluent (or dedicated) correspondence players can invest in the hardware and software that can sustain chess play at nosebleed levels. Pre-2000 correspondence chess was egalitarian. One of the greatest appeals of the game — no expensive equipment necessary — is thus in decline.

Chess on Steroids

As in other sports, elite chess players (whether correspondence or over-the-board) are a tiny minority of players. The positive side of chess computerization is that it has broadened interest in the game rather than narrowing it, as some had feared when computers first toppled the grandmasters.

For starters, the new environment has increased access to databases and instruction that accelerate the development of human skills. It has long been possible to gain experience online and to consult databases that evaluate options at every stage of the opening — in particular, how often a move has resulted in a win, a loss or a draw. Moreover, while championship-grade chess computers may carry five- or six-figure price tags, the costs of digital assistance for all but elite tournament players are among the lowest of any sport.

Consider the fact that one popular website, ChessBase, offers access to the historical record of almost 10 million games — and for an annual fee of only $50. In December 2022, its competitor, Chess.com, claimed 100 million members worldwide.

Moreover, current and would-be grandmasters, both over-the-board and online, need not worry that Moore’s Law and the stunning software that complements the advanced chips will come close to “solving” the game anytime soon. For while processor power continues to compound and database software becomes ever more efficient, the benchmarks for chess dominance become more challenging at an even faster exponential rate.

 
But neural networks will not be a perfect substitute for human teaching and coaching, which depend on understanding the psychology of individual players. By no coincidence, the human competitive chess scene has not collapsed. On the amateur side, technology is producing mixed results.
 

To see why, consider digital data resources for chess called tablebases. Imagine any six pieces placed on a chessboard in any legal configuration. There are a mind-boggling number of sequences of moves that could be taken to end the game. But there are USB drives capable of storing the computed outcomes of every such position. So once a game is down to six occupied squares, it is in effect over as far as the computers are concerned.

Here’s the rub. Edwards notes that it took from 1977 to 2005 to expand the tablebase exhaustively from five to six pieces. A sevenpiece tablebase is available, but it is a workaround originally posted on the website of Moscow State University that builds on the six-piece model rather than being computed from scratch. Going from seven to eight pieces with new calculations, even given new processors created by the relentless march of Moore’s Law, is likely to take two or three decades.

In other words, the hurdles are immensely high to get to the next benchmark in which computers can claim a brute-force hammerlock on a game that is accessible to mere mortals. And similar barriers affect fields of exploration far more important than chess. This has been most evident at two data-hungry corners of contemporary science: genomics and astrophysics.

While the cost of data storage has been steadily decreasing, the thousands of billions of bytes (terabytes) of data collected from DNA sequencing on the one hand and by instruments like the James Webb Space Telescope on the other have been overwhelming server storage. The ironic result: a surge in use of tape drives, not long ago disdained as relics of the 20th century because, while their storage potential is immense, access is much slower.

Technology and Chess of the Future

In coming decades, we will undoubtedly see programs that can explain their decisions. Edwards showed me an example of how ChessBase, which uses its own version of neural networks, can analyze a position in a way a flesh-and-blood grandmaster could — a still-rare example of machine intelligence that can articulate the reasons behind its decisions.

But neural networks will not be a perfect substitute for human teaching and coaching, which depend on understanding the psychology of individual players. Just as medical diagnostic software might be more accurate than physicians’ unassisted judgment, an empathetic doctor may still have the edge in determining the treatment that best fits the personality as well as the pathologies of a particular patient.

By no coincidence, the human competitive chess scene has not collapsed. On the amateur side, technology is producing mixed results. Edwards points out that speed chess, augmented by computers, is booming online. But the outlook for correspondence chess is more problematic. Already 99 percent of correspondence games played under 21st-century rules end in draws as the computers virtually eliminate head-slapping errors — and, more subtly, induce stalemates as competitors iterate toward common strategies.

Edwards became a champion, he observes, because the last 40 games in the tournament he won ended in draws. Will aspiring Generation Z tournament players have the requisite patience?

The most serious long-term challenge to correspondence chess and to advanced analysis of chess may be the stagnation of discovery. Much of the history of the game has been marked by the detection of clever ideas called TNs — theoretical novelties. Edwards reports that grandmasters used to come upon a new TN roughly weekly. Today, even with teams of expert researchers in support, such insights appear far less often — and when they do, they are the product of neural net analysis.

Note, too, that the repertory of truly competitive openings has been reduced. Alpha Zero, Edwards recalls, found the French opening to be “suboptimal” in two hours — a discovery it had taken him 15 years to make on his own.

* * *

The most important lesson of computer chess may be the contrast between its uncanny consistency and the results of chatbots like the notorious ChatGPT. Modern chess engines are incapable of what appears often in generative software like ChatGPT: an eruption of nonsense amid otherwise impressive text.

To be safe, the chatbot user needs to know enough to verify the answer, just as an airline pilot must be able to operate controls when autopilot fails.

In this light, the fallibility of even the greatest human grandmasters is not a bug but a feature. Over-the-board chess is still a physical as well as a mental contest that can last five hours or longer. And chess engines allow spectators of televised matches to evaluate the giants’ moves in real time. Much of chess may have been mapped, and the frontier may be closing — but the rodeos can be as thrilling as ever.

Finally, we may not have heard the last of human versus computer. People are beginning to train software to detect the flaws of seemingly unbeatable programs. Aided by automated analysis, an amateur was able to defeat one of the leading Go programs in 14 of 15 games in early 2023, playing without live electronic help. When the victor created distractions elsewhere on the board, the Financial Times reported, “the bot did not notice its vulnerability.” The ultimate human skill may be our ability to use machines to fool machines.

main topic: Tech & Telecoms