How to land a job at a frontier AI lab — without a PhD

Vladimir Feinberg leads pre-training on Gemini at Google DeepMind — about as deep inside a frontier lab as you can get. So when he writes a piece titled, roughly, “how to land a job at a frontier lab,” the first thing worth noticing is what he doesn't say. He doesn't say get a PhD. He doesn't say publish at NeurIPS. He says something far more useful for everyone who isn't already on that track.

The short version: you don't need a doctorate. You need to be scarce. And he names two specific skills that are scarce right now — two doors that are genuinely open if you're willing to go narrow.

The mistake almost everyone makes

The instinct, when you want into AI, is to get broadly good — learn a bit of everything, do the famous courses, build the same three portfolio projects everyone else built. The problem is that “broadly competent at ML” is now one of the most crowded talent pools on earth. You're one of ten thousand identical-looking résumés.

Trade “general and crowded” for “specific and scarce.” That swap is the whole strategy.

Feinberg's framing flips it. Labs don't have a shortage of generally-smart people. They have a shortage of people who can do particular hard things. Find one of those things, get genuinely good at it, and you stop competing with the crowd entirely.

Door one: kernels (going low)

The first scarce skill is kernel engineering — the low-level work of making GPUs run fast. This is the world of CUDA and Triton, of squeezing every last bit of throughput out of the hardware that training and inference actually run on.

It's unglamorous and genuinely difficult, which is exactly why it's scarce. Most ML people stay up at the model-and-math level and never touch the metal. But a frontier lab lives or dies on how efficiently it uses tens of thousands of accelerators — a kernel that's 20% faster is, at that scale, worth a fortune. People who can reliably write those kernels are rare, and labs hunt for them. No PhD asked for; the GitHub repo and the benchmarks speak.

Door two: agentic systems (going high)

The second scarce skill sits at the opposite end of the stack: building agentic systems — the high-level engineering of AI agents that plan, use tools, and carry out multi-step tasks on their own rather than answering a single prompt.

This is newer, messier, and less codified, which again is the point. There's no fifty-year textbook on how to make a reliable tool-using agent — the people who are good at it mostly learned by building things that broke and fixing them. If you can take a model and turn it into a system that actually does a job end to end, you have a skill that labs and serious companies are short of right now.

Why these two, and not the middle

Notice the shape: one door is as low-level as software gets, the other as high-level. The crowded middle — “I fine-tuned a model on a dataset” — is where everyone already is. The edges are where the scarcity lives. Pick an edge that fits how your brain works: if you love performance, systems and getting close to the hardware, go low. If you love product, orchestration and making messy real-world tasks work, go high.

The honest caveat

None of this is a hack. Feinberg is clear-eyed that real depth in either skill takes months to years, not a weekend bootcamp. That's not the bad news — it's the entire reason the strategy works. If it were quick, it wouldn't be scarce, and we'd be back in the crowd. The time cost is the moat. Pick the door, then commit to it for long enough that “rare” becomes true of you.

And here's the part I'd add as someone who hires for applied work, not research: the frontier labs aren't the only buyers. Every company now deploying AI seriously — the ones putting agents into real workflows, not demos — needs exactly these two skills. The agentic-systems engineer who can't get into DeepMind on day one can build genuinely valuable things for businesses in the meantime, and arrive at the lab later with a portfolio instead of a transcript.