Vlad Feinberg, Google DeepMind's pre-training area lead for Gemini, shares what it takes to land a job at a frontier AI lab. Key skills in demand include low-level kernel development, distributed systems engineering, mathematical maturity to read ML papers, and familiarity with scaling laws, distillation, reinforcement learning, and quantization. He explains the difference between software engineering and AI research — framing research as navigating a stochastic MDP versus a deterministic DAG. Vlad also covers how to signal your skills (e.g., contributing to vLLM or SGLang), how internal transfers work, and offers concrete exercises (implementing a transformer, working through scaling law problems) to get an interview. He shares war stories from Flash 2.0 development, including the MoE + pipeline prefill breakthrough that enabled Gemini 2.0, and 40 days of around-the-clock training shifts.
