"if one person knows how to code, they don't need an LLM. I've tried using all of them on my job, and I will give them a second chance time by time (to see if the claims made by these big tech hold up) and ALL of the time I've didn't noticed any speedup compared to coding manually."
I mean, have you truly tried? do others thing this is true? I'm truly astonished if that is majority experience, I think we need to get surveys on this.
The entire draw of coding with LLMs is that you can generate programs without knowing how to code. That's literally what "vibe coding" is. Why would it surprise you that people who do know how to code, and don't hate the process the way a lot of AI enthusiasts seem to, might not find them useful?
You'll find plenty of people on HN who have this experience,but they just tend to be ridiculed or told they just haven't used AI correctly or enough. Admitting you don't use LLMs for everything, or god forbid at all, is like telling Christians you don't want to be saved.
Imo it depends on the type of change required. If it involves churning out 1000 line of new boilerplate, an LLM will win every time. If it means tracing a subtle edge-case bug and carefully planning a deep one-line change in a 100k codebase that leaves all other behaviour unaffected, I would still expect an experienced dev who's deeply familiar with the code to do it at least as fast and more correctly first time.