Damn though the first argument is some LLM is bad pitchfork level comment… it’s not even about functionalities but about how it’s written.
Damn though the first argument is some LLM is bad pitchfork level comment… it’s not even about functionalities but about how it’s written.
AllNewTypeFace@leminal.space 2 weeks ago
That makes a huge difference to something that’s meant to be an ongoing service or product rather than a throwaway script or something. Writing code is a very small part of the effort: maintaining the codebase, fixing bugs/security issues, adding features/making any changes that are needed dwarfs that, and for LLM-generated apps, that part is vastly harder. LLMs are great at generating superhuman quantities of code which works (for some definition of working), but the resulting code, being the process not of a mind with intentions and their own understanding of the nuances of the problem and the tradeoffs of solutions, is much more difficult to understand and maintain. Individual LLM edits may look sensible, but the complete codebase is incoherent, with parts mismatching, which often leads to bugs. (This is assuming that the project was developed by someone who at least reviewed the LLM changes before merging them. If it’s “vibe-coded” by using the LLM like a slot machine and not looking at the code, the result is going to be obviously garbage.) In any case, a shiny new vibe-coded web app is unlikely to last in the real world.