Vlad Feinberg, a Distinguished Engineer-level practitioner at a frontier lab, published a tactical hiring guide in May that has been circulating steadily on X since a recent repost surfaced it to a wider audience.
The central argument: research and trench engineering have merged. The person writing the paper is often the person who built the training infrastructure. The shared underlying skill is making progress without a recipe, whether that means a novel research direction or debugging a distributed training run where standard tooling does not apply.
The most direct practical path, and the one least discussed elsewhere, is performance work: CUDA kernels, custom distributed training primitives, low-level optimization. The labs always need more of it and the skill is hard to fabricate in an interview.
Public artifacts beat credentials. Working open-source code and technical blog posts signal demonstrated ability more clearly than a prestigious institution.
One piece of advice drew pushback: Feinberg suggested sacrificing weekends to build skills fast. Aidan Clark, on X, explicitly distanced himself from that framing and argued the “in-demand skill” angle carries without the burnout cost.
Based on Vlad Feinberg’s essay published at vladfeinberg.com on 2026-05-10, currently resurfacing via X.
