Gaming the System

01 January 2024
By Gwen Weerts

The book Seven Games, by Oliver Roeder, devotes a chapter each to checkers, chess, Go, backgammon, poker, Scrabble, and bridge, and describes in detail the way each game has led to advancements in artificial intelligence. Roeder, who holds a PhD in economics with a focus on game theory, postulates that games are attractive to AI researchers because they are a microcosm of the real world. By mastering games, “computers master aspects of the human world.”

Each chapter of Roeder’s book describes dramatic human-versus-machine contests, inflection points in history when a computer learned to play one of these games better than the human masters. In the chapter about Go (which, I confess, was the only game in this list I had never heard of), Roeder describes a high-profile tournament, where Go Master Lee Sodol played a tournament against AlphaGo, the first robust AI program trained on the game.

In game five of the tournament, after AlphaGo had won three games and Lee one, it looked like Lee had the edge. AlphaGo made an attack on Lee’s game pieces (called stones), leaving its own stones vulnerable to attack. It looked like a blunder—a human master never would have done that—but it wasn’t a mistake at all. “Instead, we humans simply didn’t understand Go well enough to judge what the machine was up to,” says Roeder. He claims that AlphaGo taught humans something new about this ancient game, and subsequently changed the way we play it.

After the tournament, Sodol reflected how he had expected that AlphaGo would be a mere calculation, a machine. But after seeing a particularly unusual and effective Go move, he changed his mind. “Surely AlphaGo is creative,” he says. “This move was really creative and beautiful.”

This issue of Photonics Focus is about automation, which, at the beginning of 2024, means we must spend considerable time on AI in all of its creative, hallucinatory, ethically murky, and transformative dimensions. A feature article explores the shrinking gap between AI research and clinical translation, while another describes the “virtuous circle” of AI and photonics. A third feature investigates the many ways photonics has contributed to the automation of agriculture, and a Bandwidth article weighs the potential and danger of large-language models in scientific publishing. Many of these stories are optimistic about AI, and, just as in the aftermath of AlphaGo, expect that it will help humans become smarter, more creative, and more innovative in the way we approach problems.

But of course, many of us still have concerns. In Seven Games, Roeder describes the concept of “value collapse,” which is when easy, quantified metrics can stand in for ideas about success in subjects that really deserve more nuance, such as measuring journalism by clicks, or fitness by steps taken. In games, this is no problem, because success is easy to measure—you win or lose. But AI will approach real-world tasks the exact same way it approaches a game, without the nuance the real world often requires. Roeder cautions, “One can easily imagine sinister instances of value collapse, arising, for example, in the fields of government, medicine, or engineering.”

This tension is played out in these pages. Automation comes with risks, some of them dangerous. But it can also be creative, even beautiful. It takes a human to know the difference.


Gwen Weerts, Editor-in-chief


Recent News
Sign in to read the full article
Create a free SPIE account to get access to
premium articles and original research