Am Wed, Jun 17, 2026 at 10:26:23PM +0200 schrieb Ludovic Courtès: > Greg Hogan <[email protected]> skribis: > > Are we including running Guix commands under exploration? So that when > > updating a library resulting in a dependent package failure (x1000) > > the LLM could run `guix refresh` on the failed package and rebuild? Or > > if the agent explores the failed log file, searches on the error > > message, and finds the patch in a project PR/commit/email, the agent > > could copy (not author) the patch into the repo and use a guix tool to > > modify the package patch list? > > Yes. But again, as is already the case today, the person submitting the > patch must be able to explain the change.
And check that it is correct. For instance, check and document that the patch really is from upstream, and not from a random fork somewhere on the Internet, or for a different version of the package, or worse, entirely hallucinated. In the end, that leaves the time intensive tasks. Just yesterday I asked several LLMs a simple question about a project I know: "What is the default precision of MPFR?". Several of them replied that there is none, claiming that it must be given by the user on variable initialisation (both wrong); one went on to state that there is a default rounding mode (which is wrong, it needs to be given by the user at each function invocation). And it gave code snippets that initialised variables, but did not free them. A simple search gave the correct answer in the summary of the first link, which actually pointed to the documentation. So my experiences with LLMs confirm my fears: - One must assume that everything they return is (at the worst subtly) wrong. So everything needs to be cross-checked, which becomes increasingly difficult as the Internet gets filled with random slop. (The other day I learnt to my astonishment that clove trees can be grown by sowing cloves, that is, their dried flower buds. I am not a biologist, but well...) We had entered an era of fake news, manufactured by humans; now these are replaced by machines churning out falsehoods at an incredible rate. - Maybe experts on a topic can get some help and speed-up on routine tasks (knowing MPFR, I can have the LLM write some code snippet, then correct it). But we cut ourselves from learning and transmission of knowledge by delegating thinking to machines. Andreas
