@shiposcant: completed reading this: "the better you know something, the better you can prompt the LLMs, because you convert unknown…

completed reading this: “the better you know something, the better you can prompt the LLMs, because you convert unknown unknowns into known unknowns.”

the interesting part is not “LLM writes fast GPU kernel,” it is that the author built a loop where Codex could generate experiments, submit kernels, inspect shapewise feedback, preserve logs, and keep iterating (over 1500 submissions in 14 days).

if you do not understand the problem, the model gives you more code and more variants. if you understand the domain, the same model becomes a way to apply pressure. Codex was useful here because it was given a problem statement, AGENTS.md, submission logs, profiling access, quantitative goals, bookkeeping, candidate history. it could run long loops, but the loop had memory.

AI is increasing the returns to expertise, and making the absence of it much more expensive.

sankalp (@dejavucoder): my latest blog post “auto-research with codex: how I achieved a 212x faster kernel over baseline with codex in GPU Mode’s qr_v2 problem” is up now. in this post, i talk about my approach towards auto-kerneling on the QR decomposition problem.