AI Has Cracked the Most Diabolical Problems in Math. Why Can’t It Solve Chess?

The world has spent recent weeks coming to grips with the unfathomable power of artificial intelligence.

AI left academia’s brightest mathematicians slack-jawed when it found solutions to problems that once melted their brains. And it produced an international freakout after researchers sounded the alarm on technology that could bring about the end of mankind. At this point, AI seems capable of doing practically anything—and quickly.

So why hasn’t it cracked the code of the world’s most famous board game?

The starting math of chess is relatively straightforward. Thirty-two pieces on 64 squares with perfect information for both players—no one is bluffing or hiding any cards. Yet chess hasn’t been mathematically solved, because the astronomical number of possibilities eventually exceeds the number of atoms in the universe. And even though a cellphone can now checkmate Garry Kasparov, artificial intelligence hasn’t managed to concoct a foolproof plan to win every time.

But as it turns out, the story of chess and AI isn’t one of imminent doom. Instead of simply sinking human players with a single answer, AI is now helping them navigate the board by continuously unlocking secrets of the ancient game.

“You know what you’re getting with chess, and yet it remains timeless and unsolvable,” said Chess.com’s chief chess officer Danny Rensch. “The geometry within the geometry is a fascinating infinite problem.”

In checkers, for instance, computers have reached the level where perfect play by both sides will always result in a draw. And in even simpler games, they are now unplayable: A computer that’s given the opening move in a game of Connect Four cannot be beaten or even tied.

But oddly enough, the unprecedented computing power now pointed at chess hasn’t winnowed the number of strong moves—it has actually expanded them. Rather than finding one solution, artificial intelligence has discovered that there are more ways to play sound chess than previously believed. In other words, instead of squashing creativity, AI is inspiring it.

In a game where the decision tree expands frighteningly fast, both man and machine are only beginning to explore the full extent of a game that has been around for centuries.

“Humans don’t have a good intuition for what huge numbers mean, but there’s a huge difference between a billion and then the number of possible chess positions,” said Boris Power, the head of OpenAI’s applied research team. “It’s like comparing one second to the lifetime of the universe.”

It was nearly 30 years ago when chess helped mankind properly understand that it was being outsmarted by machines. After Kasparov, then the world champion, beat IBM’s Deep Blue in 1996, the computer earned its robot revenge in a 1997 rematch. It was a watershed moment, even for those who don’t know the first thing about fianchetti.

The years that followed were a dark age for chess. As top grandmasters began studying with computers, their tactics became stale and mechanical. Some feared that a game that had been long associated with intellect and creativity was losing its soul.

But those computers, which were trained on libraries of grandmaster-level games, were cave men compared to what was coming. Three decades on, the cutting edge of chess technology is artificial intelligence programmed to rediscover the wide world of possibilities from first principles. This time, a human wasn’t teaching a machine to play chess—the machine would teach itself. Leela Chess Zero, an open-source engine based on Google’s AlphaZero, has played over four billion games in order to become one of the most powerful engines around.

“No human heuristics were injected into it,” said Folkert Huizinga, a researcher who worked on Leela. “It learned everything from scratch.”

Leela even learned to play with attitude. All by itself, the computer combined the skills of a super grandmaster with the gamesmanship of a Washington Square Park hustler. Because Leela had, in a sense, learned to extend games as long as it wasn’t losing, it didn’t always take the most direct path to checkmate. Opponents interpreted this as showboating.

But when the engines aren’t trolling humanity, they’re changing how humans understand the game. Openings that had long been considered dead because players had concluded they were so tactically weak have been revived because the engines saw sequences that world champions couldn’t. One defense, commonly known as the Damiano Petrov, had been viewed for centuries as a recipe for disaster. Beginners are taught to avoid it from the time they learn how the knight moves.

AI, however, is here to rehabilitate the Damiano name. Upon further review, the machines have found ways to turn this slow-motion trap into a perfectly sound strategy.

“There’s a lot of richness in the games that you can reach with very intelligent play,” Huizinga said. “It’s a vast game space…we don’t even know what percent is explored.”

Power, the OpenAI researcher, said that even if there are simply too many combinations for a formal mathematical proof solving chess, there may come a day when optimal play is discovered and the game has been solved on a practical level. For now, he uses chess to understand artificial intelligence—and not the other way around.

When the company comes out with new versions of ChatGPT, Power likes to use chess as a benchmark for its capabilities. He has found that if he gives the latest models enough time to think, they create a chess engine from scratch that can reach grandmaster quality. But forced to calculate on their own, without coding, they’re substantially weaker.

“We want to see: Do the models get better at reasoning and being able to think in context?” Power said. “So, can the models spend more time thinking to come up with better solutions—a bit like a human does?”