The Case for Craft

Thu, 24 Sep 2026 00:00:00 +0000

Markus Hartikainen runs a team at AMD that scales LLM inference on vLLM, and last week he asked a question I have been answering, in one form or another, for most of my career:

After a morning of writing PRs to vLLM, I want to reflect on how many of the PRs I write should sit on the critical path. I still write them. I think I have to reduce how many are mine. Otherwise, I become the blocker. Also, I think more of my time should go to public roadmap issues, RFCs, and public comments on other people’s PRs, so the work still moves. I do not know if that is enough to remain technical. I am worried about both credibility and skills/knowledge. How do I keep both of those? If you have gone from writing the kernels and deeply technical work to more directing people who write them, how did you keep contributing to discussions about KV layout, traces, and HBM traffic?

In the comments he put the fear more sharply: “I am not sure I can still see the critical issues in that path if I am less hands-on.”

It is a real fear, and it is aimed at the wrong thing. Reaching for the keyboard is natural, and the keyboard was never the craft. I learned that on a ladder that took my hands off the equipment one rung at a time, in a place where losing the craft is a plant problem, not a career problem.

Career on Andrew Hunter — Systems Architect & Builder

The Case for Craft