8 Google Employees Invented Modern AI

EIGHT NAMES ARE listed as authors on “Attention Is All You Need,” a scientific paper written in the spring of 2017. They were all Google researchers, though by then one had left the company. When the most tenured contributor, Noam Shazeer, saw an early draft, he was surprised that his name appeared first, suggesting his contribution was paramount. “I wasn’t thinking about it,” he says.

It’s always a delicate balancing act to figure out how to list names—who gets the coveted lead position, who’s shunted to the rear. Especially in a case like this one, where each participant left a distinct mark in a true group effort. As the researchers hurried to finish their paper, they ultimately decided to “sabotage” the convention of ranking contributors. They added an asterisk to each name and a footnote: “Equal contributor,” it read. “Listing order is random.” The writers sent the paper off to a prestigious artificial intelligence conference just before the deadline—and kicked off a revolution.

Approaching its seventh anniversary, the “Attention” paper has attained legendary status. The authors started with a thriving and improving technology—a variety of AI called neural networks—and made it into something else: a digital system so powerful that its output can feel like the product of an alien intelligence. Called transformers, this architecture is the not-so-secret sauce behind all those mind-blowing AI products, including ChatGPT and graphic generators such as Dall-E and Midjourney. Shazeer now jokes that if he knew how famous the paper would become, he “might have worried more about the author order.” All eight of the signers are now microcelebrities. “I have people asking me for selfies—because I’m on a paper!” says Llion Jones, who is (randomly, of course) name number five.

“Without transformers I don’t think we’d be here now,” says Geoffrey Hinton, who is not one of the authors but is perhaps the world’s most prominent AI scientist. He’s referring to the ground-shifting times we live in, as OpenAI and other companies build systems that rival and in some cases surpass human output.

All eight authors have since left Google. Like millions of others, they are now working in some way with systems powered by what they created in 2017. I talked to the Transformer Eight to piece together the anatomy of a breakthrough, a gathering of human minds to create a machine that might well save the last word for itself.

THE STORY OF transformers begins with the fourth of the eight names: Jakob Uszkoreit.

Uszkoreit is the son of Hans Uszkoreit, a well-known computational linguist. As a high school student in the late 1960s, Hans was imprisoned for 15 months in his native East Germany for protesting the Soviet invasion of Czechoslovakia. After his release, he escaped to West Germany and studied computers and linguistics in Berlin. He made his way to the US and was working in an artificial intelligence lab at SRI, a research institute in Menlo Park, California, when Jakob was born. The family eventually returned to Germany, where Jakob went to university. He didn’t intend to focus on language, but as he was embarking on graduate studies, he took an internship at Google in its Mountain View office, where he landed in the company’s translation group. He was in the family business. He abandoned his PhD plans and, in 2012, decided to join a team at Google that was working on a system that could respond to users’ questions on the search page itself without diverting them to other websites.

Apple had just announced Siri, a virtual assistant that promised to deliver one-shot answers in casual conversation, and the Google brass smelled a huge competitive threat: Siri could eat up their search traffic. They started paying a lot more attention to Uszkoreit’s new group.

“It was a false panic,” Uszkoreit says. Siri never really threatened Google. But he welcomed the chance to dive into systems where computers could engage in a kind of dialog with us. At the time, recurrent neural networks—once an academic backwater—had suddenly started outperforming other methods of AI engineering. The networks consist of many layers, and information is passed and repassed through those layers to identify the best responses. Neural nets were racking up huge wins in fields such as image recognition, and an AI renaissance was suddenly underway. Google was frantically rearranging its workforce to adopt the techniques. The company wanted systems that could churn out humanlike responses—to auto-complete sentences in emails or create relatively simple customer service chatbots.