LLMs work better in codebases with less tech debt.
Especially as a codebase grows, LLMs can get lost in dead code and red herrings, so I don’t think the fundamental assumption of this article holds true, atleast not in the size of codebases I’m working with.
Also, I think most tech debt doesn’t take the form of coding errors, but decisions with external dependencies, and difficult migrations - changing your api schema, self-hosted instead of managed databases, unscalable database architecture. The hard hard problems. Not that you shouldn’t take in the tech debt early on; a startup shouldn’t start being infinitely scalable.