Yeah, I’m not sure yet how I feel about the whole "LLMs with Lean ending
math" thing. The impression I get is that, rather than finding *the book
proof* a la Erdos, many of these proofs, for instance the work on odd
values of the zeta function, are more on the dissatisfying side, like the
four-color theorem. But idk, I haven’t been in the trenches of this, and I
mostly avoid the social media outrage around it.

I don’t know if I understand what you’re getting at when you say, "What
they are conveniently not mentioning is that the
wander-around-and-discover-stuff activity is accumulating at superhuman
speed in AI notebooks on behalf of goal directed behavior." It sounds cool,
and maybe closer to what I’m hoping for, but idk.

Am I naive in thinking that Lean formalizations are more the targets of
these goal-directed activities than the rails guiding discovery? It has
seemed to me, over the last few years, that type theory ought to make a
reasonable pidgin language between humans and LLMs. In particular, I think
the process could be tightened up, and made less token-consumptive, by
using types to guide compositionality.

I mean goal-directed computation in general, not just proof search: an
agent finding a ride, comparing options, obtaining authorization to spend,
and booking it, for example. I suspect richer types could describe the
intermediate obligations and permissible compositions well enough to steer
the agent, rather than merely check its output or the syntax of its tool
calls.

If you’re saying that this is already mostly what’s going on, I do find
that interesting.
.- .-.. .-.. / ..-. --- --- - . .-. ... / .- .-. . / .-- .-. --- -. --. / ... 
--- -- . / .- .-. . / ..- ... . ..-. ..- .-..
FRIAM Applied Complexity Group listserv
Fridays 9a-12p Friday St. Johns Cafe   /   Thursdays 9a-12p Zoom 
https://bit.ly/virtualfriam
to (un)subscribe http://redfish.com/mailman/listinfo/friam_redfish.com
FRIAM-COMIC http://friam-comic.blogspot.com/
archives:  5/2017 thru present https://redfish.com/pipermail/friam_redfish.com/
  1/2003 thru 6/2021  http://friam.383.s1.nabble.com/

Reply via email to