Noam Shazeer

there's a version of the story of modern AI that you could tell almost entirely through one person's decisions about where to work.

Noam Shazeer grew up steeped in the kind of mathematics-adjacent thinking that tends to find its way to competitive programming before it finds its way to anything else — he was a three-time Putnam Fellow, which is the kind of credential that means you were very quietly exceptional long before the field you'd help define even existed in its current form. he joined Google in the early 2000s, when the company was still figuring out what it wanted to be beyond a search engine, and he stayed for almost two decades. that's a long time to be anywhere, and it matters.

what he built during those years is genuinely hard to overstate. in 2017, Shazeer was one of the eight authors on "Attention Is All You Need" — the paper that introduced the transformer architecture, the technical backbone on which essentially every major language model running today is built. GPT, Gemini, Claude — the transformer is underneath all of it. that paper is now one of the most cited in the history of computer science. and yet for a long time, Shazeer was the kind of figure you had to already be paying attention to in order to know about.

the reason he's at the center of conversation right now is partly biographical and partly structural. he left Google in 2021, frustrated — by most accounts — with the pace at which the company was willing to actually ship things. he co-founded Character.AI, which became one of the more surprising consumer AI successes of the early wave, built around the idea that people wanted to talk to AI in ways that felt personal and open-ended rather than task-oriented. Google eventually acquired the Character.AI team in a deal that brought Shazeer back inside the building — a notable reversal, and one that said something about how the calculus of talent had shifted.

and that calculus is very much what's being discussed right now. Google is in a complicated moment: it has deep research infrastructure, it has Gemini, it has resources that no startup can match — and it keeps losing people. senior researchers have been moving toward Anthropic and OpenAI in a pattern that's consistent enough to be a story rather than a coincidence. the reasons vary — pre-IPO equity is part of it, a sense of urgency and mission is part of it — but the underlying question keeps circling back to the same thing: which institution is actually going to matter, and when? Shazeer's original departure from Google, and the terms of his return, sit at the center of that question. he walked away once because the pace felt wrong. the fact that Google paid significant money to get him back suggests they understood, maybe late, what it costs to let that kind of judgment walk out the door.

what makes Shazeer interesting isn't just the résumé, though the résumé is extraordinary. it's that he keeps turning out to be a reliable signal. the transformer paper. the early bet that consumer-facing AI products had a real market. the Google exit at exactly the moment when the field was about to shift into a higher gear. he's not someone who explains what's happening — he tends to just do the next thing, and then the field catches up.

at some point you stop calling that luck.