Post

david barsky
@davidbarsky.com

wild to see the replies to this post. i just had sol-5.6 write a fuzzer for incremental compilation in rustc almost autonomously. one of my cofounders had claude spend the night profiling/benchmarking our storage system and came back to a stack of changes in the following morning

P(aul) Frazee@pfrazee.com

This is the crux of the debate, right? Whether it's doing mimicry or reasoning. I struggle to see how you could view it as purely mimicry. It has to tie together multiple novel thoughts in a logical construction to create its responses.

Solly
@directxman12.dev

one the more interesting parts of llms (to me) is what their results imply about the relationship between language and reasoning (particularly, say, the recent discoveries in math). i also kinda get the feeling that a lotta folks tied their foundational sense of self around the idea ...

0 reposts12 likes

Solly
@directxman12.dev

that human reasoning is some sort of ineffable foundational concept, and so the idea that we made a language machine and reasoning fell out is off-putting to them because it calls into question the exact nature of reasoning.

randomanybody.bsky.social
@randomanybody.bsky.social

Fwiw, years ago i was really moved to read that, in human evolution, the brain size grew after the troat changes allowing to modulate sounds. We didn't start talking because we had stg clever to say. We got the tooling for formulating thoughts after being able to talk

Arrdem
@arrdem.tirefireind.us

I feel like what's happened is that we've gotten an existence proof that substantial amounts of "reasoning" are formulaic (which we've known forever) and that thus within the algebraic context or whatever you want to call it of a language CULTURE, it's possible to IMPLEMENT reasoning as evaluation

Arrdem
@arrdem.tirefireind.us

LLMs are functions from (next token, culture environment, activation/attention state over context window) -> next token so they're this interesting double-conditioned thing on the language culture generally AND the running context, and they sure seem to be able to use fuzzy matching to bridge logic

Arrdem
@arrdem.tirefireind.us

I think the surprise is that language is turning out to be quasicomputable? At least a field within which you can define computation on or something like that. But we've been doing natural language direction passing forever and you have to teach people to reason in those terms so...

Solly
@directxman12.dev

is it exactly reasoning? who knows, likely depends on your definition. but it's clearly _something_, and that _something_ has fascinating implications.