I’ve participated in a discussion under one post by each of those three people, I believe. And every single time you’ve started arguing with me in a fairly rude manner.

(By which I mean the complexity of the learning problem for each of these systems)

Humans…in the mathematically precise computational complexity sense

The authors do not (and it would be absurd to) claim that the learning problem is intractable for, say, c elegans. If you accept the authors claim that the problem is intractable for humans then the problem is more *complex* for humans in a very clear sense

In the sense that the authors of the Dr. Ingenia proof do not claim to have shown that the AI-by-Learning problem is intractable for such systems, since the property of human cognition that makes the learning problem hard is its “domain generality.” Read the paper!

Not the point

Right, I’ve read your reply and do agree that the paper doesn’t show what it claims to show. I’m just making the point that even if it did, by their own lights it can’t be used to deny LLMs properties that do not require “human-level cognition”

And the divide-and-conquer point in the Dr. Ingenia paper doesn’t apply here since it concerns subcomponents that could be composed into human-level cognition, and you can’t obviously compose simpler whole systems together to get human-level cognition.

The Dr. Ingenia result doesn’t apply here since it concerns “human-level cognition.” To say you could use it to deny AI systems mental states, consciousness, pain states, etc. would require denying those states in simpler systems (say, a fish).

Yes obviously. The conclusion that LLMs are not currently conscious etc is correct but the arguments being used here are not great and aren’t specific to LLMs. They show too much (that a computer could never be conscious) and are just…Searle.

See also Dennett's "The Milk of Human Intentionality"

This idea that there's a map-territory confusion here just begs the question against the computational theory of mind (see passage below from Russell and Norvig's AI textbook)

Did I ever say that human cognition works the same way as current AI systems?

Anyway yes I think you've put your finger on the impasse and I appreciate the context for your general view. I do think reasonable people can disagree here.

I’ll also have to take a closer look at Birhane’s dissertation for the relevant bits when I get the chance. And the slow science manifesto also looks interesting, thanks for sending. I'm certainly not against slow science or embracing new forms of science.

That's obviously too quick a reply to the paper, and I'll give it a reread when I get the chance. (I also am familiar with the Reclaiming AI paper and the How Hard Is Cognitive Science? paper—I disagree with the conclusions but getting into that is a whole other discussion.)

Your paper about the tension between computationalism and reverse engineering is interesting, although I don't think it shows that you can't reverse engineer an evolutionarily designed system.

(Cao’s “Multiple Realizability and the Spirit of Functionalism”) but I also think it’s not necessary to reproduce all such functional detail and that there’s a level of abstraction above that lowest level that is feasibly multiply realized.

I also acknowledge of course that functionalism is not without worries but any substantive position here will have problems. For example I do think a perfectly detailed functional description of human cognition is likely not nomologically possible to multiply realize in a different medium