On 23 Jul 2026, at 12:56, Marc Haber <[email protected]> wrote: > > On Thu, Jul 23, 2026 at 12:30:40PM +0100, Jessica Clarke wrote: >> LLMs and compilers are very different in their nature, and as someone >> whose day job is working on compilers quite frankly it’s insulting how >> much people want to say LLMs are the same for code. Compilers are >> engineered to follow input and output specifications and generally any >> miscompilation is regarded as a serious bug. LLMs are a roll of the >> dice where you pray you get something sensible out, and if you don’t, >> that’s just accepted as a thing that happens, and you have to pull the >> slot machine arm again after another prayer. > > An LLM in software development will easily bring you to 90 from 100, while > for a hand-written solution of the task you will be at 80 after spending ten > times the time. And probably stop there ("good enough") and release. > > Starting with the LLM output at 90, you still have energy to bring the > project to 95 of 100 and then release and STILL be quicker than 100 % hand > coded, 80-of-100-code.
I am not arguing, and do not want to argue here, about the effectiveness of LLMs in practice. My point was solely to highlight these are very different approaches to “solving” similar problems, with very different properties regarding aspects such as determinism, predictability, and trustworthiness, and that it is inappropriate to treat them as the same. Please do not drag me into discussions on -vote about the *merits* of LLMs. I do not want to be a part of that. Jessica
