I've always wanted to ask @petemandik if he thinks it more likely that LLMs are closer to doing "real" reasoning than many think, because of his endorsement of proposition-centric conceptions of cognition.
(I happen to think that if LLMs are not doing real reasoning including reasoned inferences, then what they are doing will do until the real reasoning gets here. However, I think this has more to do with something like the brain equivalent of sematic vectors in LLMs rather than propositions being largely sufficient for cognition.)
Professor Mandik thinks propositions are pretty cognitively and metaphysically powerful. (I think. I could be confused or not doing a good job of stating his view).
For example, I am pretty sure the professor thinks that if one can conceive of a proposition about a round square then that is sufficient to entail the metaphysical possibility of a round square (or a circle with corners). I am only mentioning this metaphysical idea/view because it goes to his emphasis on the centricity of propositions in both metaphysics and cognition. (I think that's an emphasis/view he owns. I could be confused.)
(I'm not convinced about the proposition-heavy metaphysical-possibility-entailment idea, but that's no knock-down argument. I just have trouble getting from a proposition about a circle with corners to a metaphysical possibility thereof. Others don't have a problem with it. Overall, it's probably not really very important.)
The idea is that if Prof Mandik is right about proposition-centric (either natural language or mentalese?) cognition/reason, which removes the need for mental imagery to play a major/necessary role, then statements/propositions are exclusively doing more of the work in human cognition and reasoning than I think they are.
Since LLMs are all about language and therefore propositions, maybe the Prof thinks they're pretty close to doing real reasoning but for proposition-centric reasons?
Another way of putting it is that with an emphasis on (cognitive processing of) propositions as (either mostly or completely) sufficient for cognition, AGI might not require what I think it requires: some kind of human-emulating synthesis of mental imagery, the phonological loop, mentalese, the visuospatial sketchpad, and limbic signals among other things.
(Sufficient and necessary conditions in logic are generally all or nothing on an alethic and bivalent basis, but I am being the paraconsistent Devil's advocate because I'm lazy and not completely sure how much work propositions have to be doing in cognition to call it propositional. I am not even going to try to wrestle with the nature of propositions like some philosophers have. Down that path madness may lay waiting.)
I also wonder if LLMs might be much better emulations of human cognition than many suspect because the semantic vectors they use are very close to the way the brain works. One could have a kind of physicalist, identity-theoretic conception of mind and endorse eliminativism about qualia, but retain the concept of semantic vectors (reductively) sustained by neural circuits and assemblies of Hebbian-clustered neurons.
