Emacs freezes LLM package fight pending GNU standards update

Emacs maintainers have put a stop to a renewed fight over packages that wire the editor to large language models, saying nothing constructive can be decided until Richard Stallman revises the GNU Coding Standards.

The flashpoint is packages such as gptel, already on GNU servers, which can talk to local inference engines but ship with ChatGPT configured by default and include dedicated modules for a long list of commercial APIs. Critics led by Jean Louis call those modules pure clients for Service as a Software Substitute and want them stripped so the package no longer steers users toward non-free remote computing. Defenders reply that removing them would collapse the user base and that Emacs is otherwise losing programmers to VS Code and Copilot.

Cecilio Pardo opened the latest round with a pragmatic case: prompts and Org-mode output already make Emacs a strong LLM environment, remote model access is closer to rented hardware than a replacement for a local program, and keeping people inside a free client keeps them closer to free software than pushing them to proprietary IDEs. Stallman, answering on the list, drew a sharp line: GNU does not reject LLMs as such, only freedom-denying ones; a useful free model would face no a priori objection. Temporary distrust of LLM-written patches, he added, is about copyleft provenance, not a moral ban on the technology.

Jim Porter raised a separate practical worry that LLMs flood maintainers with fluent text their authors do not fully understand and erode the human community that sustains Emacs. Eli Zaretskii, the lead maintainer, rejected that thread as off-topic social policy for emacs-devel and later told the whole renewed packaging argument to wait. Once Stallman publishes updated standards text, he said, the project can read it carefully and only then decide what it means for Emacs and ELPA.

Side points remain unresolved: whether any current open-weight model meets GNU's bar given training-data licenses, and how far local CPU inference changes the SaaSS analysis. Those questions are on hold with the rest.