"reasoning model" is kind of like "gourmet deli" - if it was true you wouldn't need to put it in the name, because it would be self evident. It's only there as a counter argument against the evidence of your eyes.
Comments
Log in with your Bluesky account to leave a comment
I call bad product/service qualifier terms like "luxury", "sexy" and "classy" - "BADJECTIVES". I need to come up with another term for nouns and modifier nouns.
But surely logic is just an emergent property of statistical analysis and if we do enough stats we can proc the immaculate conception of logic from therein
What's weird is that we know what the LLMs are doing from their basic architecture as laid out in Attention Is All You Need, and they're not reasoning, by construction. They manipulate strings of text based on rules about text - there's no access to ideas at all.
What's weird is that you make claims that are completely unfounded. Nothing in the the "construction" of LLMs mean they can't reason. Humans "manipulate strings of text" too. Please provide a formal definition of "no access to ideas" that applies to LLMs that doesn't apply to humans.
No, sweetie, with humans it goes:
Words -> ideas -> reasoning about ideas -> new ideas -> words out.
Whereas LLMs do
Words-> rules about word association -> words out.
As you should know.
Sweetie, LLMs do the same thing and you have utterly failed to prove your point.
First: what is an "idea". What are "rules about word association". Without precise definitions, you can't prove that LLMs are doing anything different than humans.
You don't know how the transformer architecture works? The weights in the model are concerned *entirely* with the likelihood of words (tokens) being associated based on the training corpus - which is unrelated to how the associated *ideas* are related when thinking. Really this is very basic stuff.
Comments
Words -> ideas -> reasoning about ideas -> new ideas -> words out.
Whereas LLMs do
Words-> rules about word association -> words out.
As you should know.
First: what is an "idea". What are "rules about word association". Without precise definitions, you can't prove that LLMs are doing anything different than humans.
It's laughable to say they don't reason.
Humans associate likelihood of words too. Your claim that word associations are "unrelated to ideas" is baseless.
And we are no longer in the era of base models approximating the corpus. RL changes the output distribution.