Google DeepMind Distinguished Engineer: To Land a Top AI Lab Job, You Need to 'Work Like a Dog'

If you’re aiming for a research role at OpenAI, Anthropic, or Google DeepMind, you’d better be ready to sacrifice your nights, weekends, and possibly your social life. That’s the candid message from Vladimir Feinberg, a Distinguished Engineer at Google DeepMind and the head of Gemini pretraining, who told aspiring AI researchers they’ll need to “work like a dog” to have a shot.

Feinberg recently published a blog post titled How to Get a Position at a Frontier Lab, in which he lays out the unvarnished reality of competing for the most coveted jobs in artificial intelligence. Speaking to Business Insider, he broke down exactly what it takes to break through.

The competition, Feinberg explains, is exceptionally fierce. “The top universities always have a cohort of the strongest undergraduates and PhD students,” he wrote. “They are already doing machine learning research at top-tier conferences, competing in math and programming contests, and they’ve made connections with these labs through upperclassmen or friends.” These students, he says, succeed because they exhibit three traits that reliably predict future achievement: a sense of purpose, mathematical maturity, and perseverance.

Feinberg is blunt about what he would do if starting college today: “I would do everything I could to join that group. Choose difficult, proof-based math courses — and, of course, write code.” On the use of AI coding tools, his stance is nuanced: use them freely, but only for tasks you already know how to do yourself.

There are no shortcuts to mathematical maturity, Feinberg warns. But beyond that foundational requirement, “the most direct way to land a lab position is to prove you’ve mastered a specific skill that the lab needs.”

The experience paradox

Feinberg acknowledges the frustrating catch-22 that defines the field: you can’t get frontier model experience without a lab job, and you can’t get a lab job without frontier model experience. His advice is to attack from the edges.

“Frontier labs pour most of their energy into developing large language models,” he explains. “So ask yourself: what does the model need to run? What do its outputs plug into? Those are the areas where frontier labs will expand next. There are also a handful of domains that are critical to the business yet don’t require you to personally train a model.”

Be someone worth rooting for

Beyond the technical grind, Feinberg also offered a piece of advice that cuts across disciplines: “Be the kind of person your colleagues want to see succeed.” He urges researchers to identify opportunities where complementary team skills can shine, to clearly credit collaborators’ contributions when speaking to management, and to gravitate toward projects where your success lifts others as well.

It’s a recipe that demands both intensity and generosity — a combination that, in Feinberg’s view, is what separates those who merely apply to frontier labs from those who actually get in.