Post

🤔 @jtp.bsky.social it just occurred to me that the benefit of some "on the box" llm on Linux could actually be pretty huge. So much knowledge is buried in Man pages, and having a little local open model trained on it would probably help adoption. 🤷 Now I've said it. How do you think?

Reid Marlow

@reidmarlow.com

1 Sept

The failure mode with local models trained on man pages is hallucinating valid-looking flags when options change across versions. Piping the actual local man page or --help output into a small context window avoids that because the ground truth is already on the disk.

James Brundage | MVP

@mrpowershell.com

Yeah, exactly. It should also be something you could easily handle within a SLM, rather than an LLM. Entirely on the box. But having an easier way to search MAN page knowledge on box without already knowing your grep would be great (IMO)

1/9/26, 5:17 am0 reposts1 like

Reid Marlow

@reidmarlow.com

1 Sept

A 1B or 3B model doing semantic retrieval over structured man page sections works cleanly. Parsing the SYNOPSIS and flag blocks into a local SQLite table gives the small model the right boundary, so it only has to rank relevance instead of guessing command arguments.