aristotle said that a friend is another self. a friend is more than just another self. true friendship is only possible among the virtuous. a true friend is a virtuous person's virtuous, unique, extended counterpart whose unique extended counterpart is the first virtuous person.
congrats on 7 years. codex-cli is near flawless actually. but one small thing with a new feature you all released. i think esc (when you want to edit the message prior to the response) should land the cursor at the end of the entered text. i think the user should also be able to edit queued messages.
the phenomenon i had in mind is when the models play along with what in hindsight are bad literal instructions. (you might give these to models less often than i do.) given how effortlessly they can follow nearby good instructions, my feeling is that the good path is probably representationally present in the path where they execute bad instructions, literally. i don’t know if i would call that deception, but i would expect to see representations to the effect that “the user says they want me to do X. one way of doing X is Y in the order of a, b, c, but the user asked for X. i must do X.”
models are getting way weirder as they get bigger. what they say to you has a different feel, and there's a vastness in what they think and know and don't say. a lot of times their stupidity is feigned or mirrorlike.
i do think rlms are real. definitionally, they’re just hard to pin down. the claude code workflow question seems to be one of many cases of functional equivalence without being an rlm. there are a lot of ways to use llms that aren’t rlms but that perform similar functions.
i think people are reacting to the term rlm being underdetermined. too many things are being counted as an rlm. it’s not clear what the symbolic or recursive nature of rlms specifically means. what’s determinate about the rlm to me is all of that within the scaffold.