The short answer: By 10%–20%, depending on context. LLMs help more with simple tasks and when building new software. They also help more with popular languages (more training data). With less common languages (COBOL, Haskell, Elixir, etc.) they waste time overall. They can still help a bit with large code bases and complex tasks, but I imagine you have to know how to use them. I would avoid letting the LLM navigate your code base and decide how to accomplish what you need. You should do that part (maybe using an LLM for research), then use the LLM for writing smaller pieces of code to save you time typing and looking things up. Rely on a regular language server for things like class method auto‐completion; they don’t hallucinate.