[gamblor956]

I think that presumes that an LLM is capable of "unblocking a tricky C# problem" that a group of humans cannot with research. LLMs don't understand the code they output; they just regurgitate code that's already in their training set (the proof is in the pudding; see the many posts on HN about these LLM coding assistants outputting copyrighted code token for token).

So if the tricky C# problem isn't already in their data set, the output of the LLM is, at best, random crap. Even the worst human effort would exceed the output of the LLM, and that is the average case for any "tricky" problem. LLMs are fundamentally only useful on the most common types of problems that are can better be addressed by using frameworks, plugins, or APIs.

(And on that note: every programmer I've met who says that LLM coding agents 10x'd their output is the type of programmer that would have been PIP'd or fired 10 years ago for incompetence. We used to call them "code monkeys" for obvious reasons. Junior programmers think that LLM coding agents are awesome because they don't have the experience or skill to understand just how bad the output of LLM coding agents is, and the few that survive in the industry long enough to become senior programmers will laugh at their younger selves at how much of an unmaintainable mess they made vibe coding.)