Those are the same twenty pieces of information, but over the course of your life, you’ve learned patterns of words that allow you to instantly make sense of the second arrangement, and to remember it much more easily. Your restaurant server doesn’t just happen to have a miraculous memory; like musicians and quarterbacks, they’ve learned to group recurring information into chunks.
Studying an enormous number of repetitive patterns is so important in chess that early specialization in technical practice is critical. Psychologists Fernand Gobet (an international master) and Guillermo Campitelli (coach to future grandmasters) found that the chances of a competitive chess player reaching international master status (a level down from grandmaster) dropped from one in four to one in fifty-five if rigorous training had not begun by age twelve. Chunking can seem like magic, but it comes from extensive, repetitive practice. Laszlo Polgar was right to believe in it. His daughters don’t even constitute the most extreme evidence.
For more than fifty years, psychiatrist Darold Treffert studied savants, individuals with an insatiable drive to practice in one domain, and ability in that area that far outstrips their abilities in other areas. “Islands of genius,” Treffert calls it.* Treffert documented the almost unbelievable feats of savants like pianist Leslie Lemke, who can play thousands of songs from memory. Because Lemke and other savants have seemingly limitless retrieval capacity, Treffert initially attributed their abilities to perfect memories; they are human tape recorders. Except, when they are tested after hearing a piece of music for the first time, musical savants reproduce “tonal” music—the genre of nearly all pop and most classical music—more easily than “atonal” music, in which successive notes do not follow familiar harmonic structures. If savants were human tape recorders playing notes back, it would make no difference whether they were asked to re-create music that follows popular rules of composition or not. But in practice, it makes an enormous difference. In one study of a savant pianist, the researcher, who had heard the man play hundreds of songs flawlessly, was dumbstruck when the savant could not re-create an atonal piece even after a practice session with it. “What I heard seemed so unlikely that I felt obliged to check that the keyboard had not somehow slipped into transposing mode,” the researcher recorded. “But he really had made a mistake, and the errors continued.” Patterns and familiar structures were critical to the savant’s extraordinary recall ability. Similarly, when artistic savants are briefly shown pictures and asked to reproduce them, they do much better with images of real-life objects than with more abstract depictions.
It took Treffert decades to realize he had been wrong, and that savants have more in common with prodigies like the Polgar sisters than he thought. They do not merely regurgitate. Their brilliance, just like the Polgar brilliance, relies on repetitive structures, which is precisely what made the Polgars’ skill so easy to automate.
* * *
• • •
With the advances made by the AlphaZero chess program (owned by an AI arm of Google’s parent company), perhaps even the top centaurs would be vanquished in a freestyle tournament. Unlike previous chess programs, which used brute processing force to calculate an enormous number of possible moves and rate them according to criteria set by programmers, AlphaZero actually taught itself to play. It needed only the rules, and then to play itself a gargantuan number of times, keeping track of what tends to work and what doesn’t, and using that to improve. In short order, it beat the best chess programs. It did the same with the game of Go, which has many more possible positions. But the centaur lesson remains: the more a task shifts to an open world of big-picture strategy, the more humans have to add.
AlphaZero programmers touted their impressive feat by declaring that their creation had gone from “tabula rasa” (blank slate) to master on its own. But starting with a game is anything but a blank slate. The program is still operating in a constrained, rule-bound world. Even in video games that are less bound by tactical patterns, computers have faced a greater challenge.
The latest video game challenge for artificial intelligence is StarCraft, a franchise of real-time strategy games in which fictional species go to war for supremacy in some distant reach of the Milky Way. It requires much more complex decision making than chess. There are battles to manage, infrastructure to plan, spying to do, geography to explore, and resources to collect, all of which inform one another. Computers struggled to win at StarCraft, Julian Togelius, an NYU professor who studies gaming AI, told me in 2017. Even when they did beat humans in individual games, human players adjusted with “long-term adaptive strategy” and started winning. “There are so many layers of thinking,” he said. “We humans sort of suck at all of them individually, but we have some kind of very approximate idea about each of them and can combine them and be somewhat adaptive. That seems to be what the trick is.”
In 2019, in a limited version of StarCraft, AI beat a pro for the first time. (The pro adapted and earned a win after a string of losses.) But the game’s strategic complexity provides a lesson: the bigger the picture, the more unique the potential human contribution. Our greatest strength is the exact opposite of narrow specialization. It is the ability to integrate broadly. According to Gary Marcus, a psychology and neural science professor who sold his machine learning company to Uber, “In narrow enough worlds, humans may not have much to contribute much longer. In more open-ended games, I think they certainly will. Not just games, in open ended real-world problems we’re still crushing the machines.”
The progress of AI in the closed and orderly world of chess, with instant feedback and bottomless data, has been exponential. In the rule-bound but messier world of driving, AI has made tremendous progress, but challenges remain. In a truly open-world problem devoid of rigid rules and reams of perfect historical data, AI has been disastrous. IBM’s Watson destroyed at Jeopardy! and was subsequently pitched as a revolution in cancer care, where it flopped so spectacularly that several AI experts told me they worried its reputation would taint AI research in health-related fields. As one oncologist put it, “The difference between winning at Jeopardy! and curing all cancer is that we know the answer to Jeopardy! questions.” With cancer, we’re still working on posing the right questions in the first place.
In 2009, a report in the esteemed journal Nature announced that Google Flu Trends could use search query patterns to predict the winter spread of flu more rapidly than and just as accurately as the Centers for Disease Control and Prevention. But Google Flu Trends soon got shakier, and in the winter of 2013 it predicted more than double the prevalence of flu that actually occurred in the United States. Today, Google Flu Trends is no longer publishing estimates, and just has a holding page saying that “it is still early days” for this kind of forecasting. Tellingly, Marcus gave me this analogy for the current limits of expert machines: “AI systems are like savants.” They need stable structures and narrow worlds.
When we know the rules and answers, and they don’t change over time—chess, golf, playing classical music—an argument can be made for savant-like hyperspecialized practice from day one. But those are poor models of most things humans want to learn.