Le jeudi 15 janvier 2026, 17:02:05 heure normale d’Europe centrale Lucas Nussbaum a écrit : >[…] > If you haven't already, I would recommend getting first hand experience > with LLM-assisted coding. Using it to modify existing code is very > different from using it to do "Vibe coding", where you generate large > amounts of code without even reading it.
Okay, but you said: “I let the agent write an initial version of the code, and then I review it.” “Initial version,” even if that’s not “pure vibe coding,” it very much smells like it. > In general the code produced by the LLM is very specific to the context : > - the existing code being modified > - the technologies in use and the general context (Debian) > - the change being asked > > And it's also mostly about incremental modifications to a codebase. > So the amount of lines added at a time is low. It's almost never about > adding many lines of code from (likely) the same source. It's more like > copy-pasting three different 2-lines snippets from documentations or > stackoverflow, adjusting them to your context, and assembling them > together, except the LLM does it for you. Two things: 1. As said above, I wasn’t responding about incremental changes on existing code. Nonetheless, if you ask the tool to add a function to one of your files… maybe it’s incremental… but maybe it’s just regurgitating almost verbatim someone else’s code (changing their tab indentation and camel case notation to your code’s 4-space indentation and snake case). Maybe you won’t agree on “regurgitating,” maybe you’d call that “reinventing”… but if a human studied a lot of codes (and was known to have done that) and managed to write a (non-trivial) function exactly as it was in that corpus (modulo cosmestics), they would be suspected of plagiarism. 2. In any case, and more importantly, most of these tools have been trained on other people’s work, regardless (and stripping them) of their copyrights and licences. They are poisoned. -- Sylvain L. Sauvage
