On Fri, 17 Apr 2026 at 19:22, Jerry D via Gcc <[email protected]> wrote: > > First I apologize for the broken thread, my email client, operator error. > > I sense a lot of contention from the text of this thread which I read off of > the > archives. > > Since I started this thread, I thought I should step in. I think folks may be > missing an important aspect of this.
I think you might be too. > Yes, an LLM set loose to automatically go forth without human interaction is > going to produce garbage. Garbage in garbage out. Don't do this. We are > adults. > > The key to these LLMs is the prompting. If the LLM is prompted with gcc > copyrighted code chunks or files, then everything it produces is "derived > from". > The use of the word "generated" is a false presumption. It does not generate. > It > goes through a complex process of probabilities. This is why it is so good at > identifying details and corner cases. > > Yes it can "hallucinate", a horrible use of the word. Go back to statistics. > Statistics do lie. This is why it is critical to have the iterative human > interaction which is the process used by Albert and that is coupled with > multiple real reviewers doing real testing behind the scenes. The gfortran > folks > use both email and our MatterMost channels for this interaction. The topic of the thread has not been about whether the output is good quality or not. It's about the legal status of the generated output, and whether that is something the project can accept. > I have been seeing folks here using language like "shocked" or "absolutely". > Truth is we are all on the bleeding edge. I started this thread because I did > not see enough discussion in the open. I see high potential here to do the > greater good. This must always by our goal. > > It is not my decision on any of this. Please everyone, lets not shoot our > selves > in the foot. > > As always, best regards to all. > > Jerry > >
