The bottom line: Vlad Feinberg outlines a practical career path in AI labs with a focus on kernel optimization and pretraining, emphasizing that the key skills—kernel-level performance tuning and translating abstract concepts into practical implementations—are learnable and immediately deployable.
Vlad Feinberg from Google has published practical advice on how to prepare for positions at leading AI laboratories. The focus is on kernel optimization and pretraining techniques, which are considered key competencies for aspiring AI engineers.
Just before Google’s I/O conference, where new Gemini versions are expected, Anthropic and OpenAI have made headlines with smaller updates. Cursor released a SpaceX AI model with detailed insights into synthetic data and continuous pretraining. But the real story of the day comes from Vlad Feinberg’s contribution on job preparation in AI labs.
Feinberg emphasizes that kernel optimization represents the critical bottleneck in all LLM work: “The core of all LLM projects lies in performance work that makes abstract, logical changes practically implementable. Every project needs people who can optimize LLMs at the kernel level.” This is a learnable skill and the most direct path into the major AI labs.
Surprisingly, Feinberg also mentions Domain-Specific Languages (DSLs) for kernel development and agent-based work such as AutoResearch and AlphaEvolve. The practical learning path concludes with a concrete task: deriving Chinchilla scaling laws for different architectures, implementing them in JAX, and writing a Pallas kernel that achieves performance improvements.
Feinberg signals his willingness to present outstanding workshop leaders as guest speakers—a clear signal that practical technical expertise is highly valued in the AI community.