at the /bare/ minimum to even be able to claim real "reasoning" ability i think people would have to figure out how to move inference compute from a language-based chain/tree of thought to purely doing that computation "in its head" like we do, what that actually looks like in practice though idk - ThreadSky

nocab.lol • 26 days ago

at the /bare/ minimum to even be able to claim real "reasoning" ability i think people would have to figure out how to move inference compute from a language-based chain/tree of thought to purely doing that computation "in its head" like we do, what that actually looks like in practice though idk

Comments

segyges.bsky.social•26 days ago

in some sense thinking tokens are this in that they aren't final output that is under constraint to be coherent and sane. there's a few designs that allocate arbitrary amounts of compute without generating more output though

segyges.bsky.social•26 days ago

in a funny way though this is just an artifact of the decoder transformer where the way we allocate compute is by sequence length so you have to have moar sequence length to have moar flops and ram

nocab.lol•26 days ago

ooh do you have any info or papers or smth about allocating arbitrary amounts of compute without extra output? and yea i guess it just feels weird to be trying to teach problem solving in a way that we don't even do, like how often do u rly have to think step by step in spoken language to do a task

segyges.bsky.social•26 days ago

i think google had a paper or three, you want "early exiting" as a topic, i think i've seen it called "dynamic allocation". this is also kind of trivial with a diffusion model because different sampling strategies let you run fewer or more times

nocab.lol•26 days ago

ohh yeah i almost forgot about text-diffusion, that will be exciting if it catches on. i applied to test gemini diffusion but haven't been accepted yet ;(

segyges.bsky.social•26 days ago

diffusion is in theory strictly better than what we do but it's a very different design space and ordinarily instead of "start at standard next token predictor, switch it so it's technically diffusion but actually a next token predictor, search design space starting there" we do something else

segyges.bsky.social•26 days ago

a few papers have pointed out that next token prediction is mathematically a special case for diffusion but they then search the design space for diffusion models starting at some arbitrary point and not that one

nocab.lol•26 days ago

and with the way reasoning models seem to like mixing different languages during their reasoning steps makes it seem to me like they're grasping for the ability to fit more nuance and context into their output any way they can, so like surely taking the word->vec->word conversion out could help?

segyges.bsky.social•26 days ago

we actually usually do penalize them to try to make them roughly monolingual. (also performance improves if you stop doing this, lmao, but then the reasoning trace is unreadable) imho it's actually fine to have words in there, vector storage gives you continuous storage and words give you discrete

nocab.lol•26 days ago

oh interesting i didn't know they're generally trained out of that, figured people would prefer increased performance over readability since you can't be sure it isn't (probably unintentionally) "hiding" anything from you during reasoning anyways lol

segyges.bsky.social•26 days ago

the impression i got was sometimes the internal reasoning trace became a bizarre eldritch horror but i am unsure if they weren't just upset that it was mixing languages and jargon and stuff

kdkeck.bsky.social•26 days ago

Now you've lost me, what does inference compute have to do with consciousness? 🤨

Comments

Posting Rules

Reply