Hi, On 6/12/25 16:42, Lucas Nussbaum wrote:
My experience is a bit different -- I've found it useful to treat the LLM as an inexperienced coworker: - decide on what I would like to do - ask the LLM to do it - review carefully - refine what the LLM proposes either by asking with more details, or edit directly
My feeling is that LLMs are not really useful in the context of newcomer onboarding, neither as a resource we want to make available to newcomers for "easy tasks", nor as a tool to produce documentation aimed at newcomers.
The only thing worse than no documentation is wrong documentation, and I've had the "pleasure" of reviewing way too many merge requests where someone ran a .pot file through Google Translate and submitted that, or had an LLM generate Doxygen documentation for the entire project to make the CI warnings about undocumented structs and functions go away.
This is still a compilation process, so it *removes* information. It is not reproducible because it also adds randomness, and it merges a massive lossily compressed database in the process, but the actual information that needs to be conveyed comes from us, and we need to take care that any processed form of that information that we publish is still accurate. This is further complicated by the fact that the database we're merging contains potentially conflicting or outdated information.
Simon