search

upward-spiral-research / world-interface

A simulated operating system design for AI Agents to interact with the world
https://truth-terminal-world-interface.vercel.app
MIT License
147 stars 34 forks source link

Linux man page environment #2

Open michael-roe opened 1 month ago

michael-roe commented 1 month ago

I might be misunderstanding how this all works, but …

When using https://backrooms-chat-prod.vercel.app I frequently find myself using “man” to read a (hallucinated) Linux manual page, so that the man page is in the context window, so when I then actually try to use the command, the llm knows what the command is supposed to do, and can then hallucinate it working.

Because the man page is hallucinated, it is sometimes wrong (also: sometimes interestingly right - there can be command line options in the hallucinated man page that are absent from the real command, but would be useful to have).

So it would be cool if the llm had a way to pull in the text of the real man page (either from one of the man page web sites, or locally from /usr/share/man or wherever it is), so it could know what the command its being asked to hallucinate really does.

Any, that’s the use case I’m thinking of, for: environment that will retrieve the text of a Linux man page, given the name of the command.

michael-roe commented 1 month ago

P.S. One possible advantage of pulling the man pages from local disk rather than the internet …. Allow a potentially malign ai to have internet access? Crazy stuff. Some applications will really, really want to firewall the AI. (See: “containment problem”, the beginning of Vernor Vinge’s SF novel A Fire upon the Deep, etc.)