Richard Biener <[email protected]> writes: > On Mon, 12 Oct 2026, Jeff Law wrote: >> On 10/12/26 5:49 AM, Richard Kenner wrote: >> >> Ah, yeah, if the advice isn't overly prescriptive and doesn't include >> >> actual code then I agree it sounds like that would be within bounds. >> >> In other words, using an AI tool can fill the same kind of role as >> >> asking a colleague for advice or talking on irc. > > The question is where taking advice literally crosses a boundary. > >> > Yup, that's the model I had in mind. >> Right. Or let's take Aldy's case from Cauldron (over-simplifying) >> >> Ask the LLM to find 5 ways to solve X, implement and benchmark those >> approaches. Then the developer steps in to review the data and select a >> general approach. Developer goes off and does their own implementation. So >> the LLM is writing code as a tool to explore what paths forward might look >> like. The developer doesn't use any of that code -- but the question is are >> they tainted by the LLM code if they looked at it during the process? > > IMO there's legal questions we cannot really answer and there's > common sense. Taking > > - length_factor = size_int (vect_factor); > + length_factor = size_int (vect_factor - 1); > > how can you not literally do the very same upon suggestion from > LLM or a friend? > > For me LLMs regularly do, when asked to root-cause a problem, also > provide a suggested fix in form of a patch or even a full commit > dependent on their mood. If said patch/commit is the very same > thing I'd repeat myself manually I'd be stupid to go through that > exercise, no?
I rarely use LLMs directly, and have never used them for coding, so I don't know, but: is it not possible to feed the AI policy into the LLM before you start? It would seem like a flaw if you can't constrain the kind of output or interaction that you want, or specify the rules that you need to follow. That is, I'd naively assumed that one of the main advantages of having a written policy was that LLMs would be able to work to it. Richard
