That is a lot of theory of mind talk about something that doesn't have one.
There is no explanatory gap with LLMs. They are a glass box and we know their ontology fully.
I am 100% a physicalist, and this idea that anyone who disputes that the LLM is a cogniser is not is just another bullshit framing device typical of the EA/rationalist sphere.
Well I haven't talked at all about LLMs yet? I haven't made any claim that LLMs are cognisers or that everyone who disputes this isn't a physicalist. That would be insane. I say explicitly “Whether AI has phenomenal consciousness thus looks to me to be pretty unrelated to whether it can think.”
You said that the main way that your story could fail is if physicalism is wrong. I think, as a framing device, that's very clear. No mention given to the prospect of physicalism being correct, but non-conscious cognition being impossible without some speculative future technology.
You open the post by saying that you will lay out why the aspects of consciousness thought depends on can be replicated in current AI models, in this post.
You namedrop Astra and Opus 5.5 with a claim that the reluctance from serious people to believe that they are thinking is due to the bias of our current zeitgeist. Oh, if only people still thought the Turing Test was the height of cognitive science!
You make that ridiculous comparison between brainwave spikes and next word prediction. It "looks surprisingly similar"? Only with a truckload of motivated reasoning.
You compare a forward pass to human introspection, a comparison that only works if the vectors actually represent something else hiding in the latent space and aren't simply the mechanism themselves. It is not actually a given that the vectors are surface representations of concepts. They do not need to be; it's a mathematical object, a big equation, and mathematics works on its own without a real world referent.
You compare predictive processing in the brain to the algorithmic prediction of current AI models, to essentially argue that they are building a world model.
You argue that those who are saying models are “just doing matrix multiplication” are being reductive, automatically assuming that there is some kind of emergent process in either training or inference.
You set a definition of "understanding" in this article, and you explicitly say that to rule this out for current systems would require a "Cartesian theater".
Saying that you haven't brought up anything about LLMs or their architecture in this post is wild lmao
Haven't read the whole post yet, but it seems to come from a somewhat similar place as mine
https://bearlylegible.substack.com/p/the-wizard-in-the-mind
Let's get that wizard he's no good
That is a lot of theory of mind talk about something that doesn't have one.
There is no explanatory gap with LLMs. They are a glass box and we know their ontology fully.
I am 100% a physicalist, and this idea that anyone who disputes that the LLM is a cogniser is not is just another bullshit framing device typical of the EA/rationalist sphere.
Well I haven't talked at all about LLMs yet? I haven't made any claim that LLMs are cognisers or that everyone who disputes this isn't a physicalist. That would be insane. I say explicitly “Whether AI has phenomenal consciousness thus looks to me to be pretty unrelated to whether it can think.”
You said that the main way that your story could fail is if physicalism is wrong. I think, as a framing device, that's very clear. No mention given to the prospect of physicalism being correct, but non-conscious cognition being impossible without some speculative future technology.
My story here is the role consciousness plays in thought, this isn't a story about what current LLMs can do
You mention LLMs several times in the post and explicitly compare what they do to what brains do.
Where I do that I say I'll need to wait till Part 2 to justify it. You're just reaching now.
You open the post by saying that you will lay out why the aspects of consciousness thought depends on can be replicated in current AI models, in this post.
You namedrop Astra and Opus 5.5 with a claim that the reluctance from serious people to believe that they are thinking is due to the bias of our current zeitgeist. Oh, if only people still thought the Turing Test was the height of cognitive science!
You make that ridiculous comparison between brainwave spikes and next word prediction. It "looks surprisingly similar"? Only with a truckload of motivated reasoning.
You compare a forward pass to human introspection, a comparison that only works if the vectors actually represent something else hiding in the latent space and aren't simply the mechanism themselves. It is not actually a given that the vectors are surface representations of concepts. They do not need to be; it's a mathematical object, a big equation, and mathematics works on its own without a real world referent.
You compare predictive processing in the brain to the algorithmic prediction of current AI models, to essentially argue that they are building a world model.
You argue that those who are saying models are “just doing matrix multiplication” are being reductive, automatically assuming that there is some kind of emergent process in either training or inference.
You set a definition of "understanding" in this article, and you explicitly say that to rule this out for current systems would require a "Cartesian theater".
Saying that you haven't brought up anything about LLMs or their architecture in this post is wild lmao