Emacs performance on macOS drew months of instrumentation and benchmarking. Repeated runs showed rendering bottlenecks and aggressive malloc churn dominate latency, while missing memory compaction inflates virtual memory and hurts cache locality. Regexp handling, pervasive in the codebase, remained a recurring hotspot the author kept revisiting.

Using a friend’s Max‑plan access, he ran the open‑weight GLM 5.2 model on the code. After three hours the model produced two concrete suggestions; the author refined the stronger one into a 92‑line patch targeting regexp allocation. He posted it to the emacs-devel list, explicitly noting LLM assistance, but the GNU policy bars acceptance of any LLM‑generated contribution.

Frustrated, the contributor announced he will no longer invest time in Emacs, citing the policy’s punitive effect on transparency. He retains roughly forty performance patches, some verified, and has released the few that show measurable gains. His next efforts will target other open‑source projects where contribution rules are less restrictive.

The episode fuels a wider discussion about LLM‑generated code in free software. The author argues that machine‑assisted edits deserve extra review, not blanket bans, pointing to commercial sectors that ship LLM‑derived binaries daily. As the GNU maintainers grapple with legal and ethical concerns, the Emacs project will need to articulate a clearer policy for future contributors.