How to Land a Role in a Cutting‑Edge AI Lab Amid DeepSeek’s Hiring Surge

DeepSeek announced a hiring drive that aims to double the size of every department, opening positions ranging from full‑stack development and AI core system engineering to research scientist roles and even administrative functions. The author highlights that roles such as deep‑learning researcher, AI algorithm engineer, AI infra, and Agent Harness are especially coveted by aspiring AI talent.

Vlad Feinberg, the lead of DeepMind’s music‑model pre‑training, shared a unique recruitment exercise: candidates must solve every problem in the book How to Scale Your Model on paper, record the process, hand‑write a Transformer implementation, and derive the scaling law, then submit a video of the work for a potential interview in New York.

Feinberg identifies three underlying traits that make top AI talent valuable: intent (choosing high‑impact directions early), mathematical maturity (solving ill‑defined problems with general problem‑solving ability), and grit (persevering through rigorous proofs and heavy workloads). He argues that candidates lacking these traits should seek work on the periphery of a lab’s core activities.

The periphery strategy focuses on the low‑level components that enable large‑model training: kernels and operators. These are carefully crafted code pieces that execute neural‑network kernels efficiently. Feinberg notes that mastering kernel development—especially adapting CUDA kernels to domestic hardware—is a self‑taught skill that directly maps to lab needs. Optimizing open‑source large‑model implementations (e.g., making a model run faster on limited GPUs/TPUs) provides concrete, measurable results, with benchmark speedups serving as a clear performance transcript.

Beyond kernels, Feinberg points to emerging work on agents (e.g., moving from harness to loop designs) and quantization, where reducing weight precision from 32‑bit float to 4‑bit dramatically cuts GPU power consumption while preserving quality—a critical trade‑off given that AI hardware operating costs are dominated by electricity.

He stresses that interview impact comes from demonstrable artifacts: contributions to open‑source inference frameworks such as vLLM or SGLang, or performance‑boosting patches that are merged upstream. Such tangible outputs outweigh a polished résumé.

Feinberg also contrasts engineering and research mindsets. Engineering follows a relatively deterministic path—building services step by step—whereas research traverses a stochastic graph of ideas, requiring “research taste” to estimate success probabilities and prioritize experiments. This intuition, coupled with deep paper‑reading skills, is essential for research roles.

Reading foundational scaling‑law papers (Kaplan 2020; Chinchilla 2022) is presented as a prerequisite. Feinberg recommends reproducing the scaling‑law exercises: solving every problem in the scaling book, completing a JAX tutorial, implementing a ~10‑million‑parameter Transformer on free Colab TPU that learns simple arithmetic, deriving the Chinchilla law, comparing dense versus MoE architectures, and finally writing a custom kernel that outperforms existing operators while documenting speedup reasons.

While the exact New York DeepMind offer may be unattainable for most, the methodology is transferable: focus on edge‑case contributions, optimize open‑source models for specific scenarios, and submit pull‑requests that demonstrate functional, performant code. The author cites the example of Su Jianlin’s rotational position encoding (RoPE) in Kimi, which serves as a strong portfolio piece for any AI lab.

Finally, Feinberg advises cultivating collaborative teammates who genuinely support each other’s success, as large projects require diverse skill sets and mutual assistance beyond individual technical prowess.