Lucas Nussbaum <[email protected]> wrote on 23/07/2026 at 10:00:13+0200:
> On 23/07/26 at 09:26 +0200, Pierre-Elliott Bécue wrote: >> De : Lucas Nussbaum <[email protected]> >> >> + all Signed-off-by tags and GPG signatures on anything submitted to >> >> Debian infrastructure (package, commit, mail, …) have been produced >> >> by the submitter, not by an AI assistant; >> >> + commits have been pushed and packages uploaded by the submitter, >> >> not by an AI assistant. >> > >> > This adds arbitrary restrictions to personal workflows. What is the >> > problem with asking a local agent to commit and push to a branch (so >> > that it can later fetch and analyze salsa CI results)? >> >> I don't consider reasonable to allow an LLM to push on a repo from >> which work can be extracted into a distro used by millions and core >> to a big chunk of the world ecosystem. >> >> You are of course entitled to a different opinion. > > Removing AI from the equation: > > If I push some work-in-progress, experimental code to a branch > just to see how it would go through CI, and someone else decides to > extract that crappy code from my branch and ship it, I think the problem > is with someone else, not me. Nowhere did I state that humans can't do crap. It doesn't me that we should push automated systems to do more crap based on these grounds. > I think that "The submitter is solely responsible for the submitted > work" is sufficient in that case: I would be solely responsible for > pushing WIP stuff to a branch (and that's fine), and someone else would > be solely responsible for deciding that that code should be shipped. Until you (or your agent) decides not to push in WIP but rather directly on main. > Also, when you write: > >> I don't consider reasonable to allow an LLM to push on a repo > > This is not what happens in practice. What happens in practice is that > the LLM instructs a local agent (e.g. OpenCode) to execute a tool > (such as 'in a shell, run git push'). The local agent may ask the user > for explicit confirmation. The user of the local agent could also > configure it so that it does not have access to required credentials > (typically isolating it with bubblewrap). If developers fail to ensure > that the tools they use don't do crap, I think it's on them. I'll leave aside the fact that this classic behaviour brings to ~5x tokens consumption just because people are too lazy to reread and do prefer the models to fix their own crap based on CI outcome (or the fact that sometimes they modify the tests to make them pass rather than fix the code). We have a right to expect better from people. Just because a ton of them do trust enough their agents to upload on their github/gitlab/whatever account is no excuse to decide that it should be the standard we adhere to. I'm sorry, but what I see here is what I would bluntly call accelerationist hype. I know you (and some others) are totally hyped by LLMs, to the point that you might maybe believe they achieve greater than what they do. I'm not. I'm no doomer either, I use these tools on a daily basis. But I believe the more automated and fast-paced a tool is, the more safety nets are required, and do not allow any LLM to sign or upload stuff for me. -- PEB
signature.asc
Description: PGP signature
