Response from Scott Aaronson:
a) We have experiences, like being hungry, tasting a strawberry, seeing blue, etc.
b) It's possible to imagine a being which located food when hungry, ate when necessary, used colours to navigate the world, etc, but did not have these conscious experiences. To put it another way, what we've learnt so far about how the brain works gives us great insight into how we would eat, navigate, etc, but does not give us any insight into either how or why we would have conscious experiences.
c) Therefore conscious experiences are not explicable by physical brain processes.
This is a belief arising from an appeal to intuition and does not present a testable or falsifiable proposition.
Integrated Information Theory (which I am not a vigorous proponent of) posits that the experience of consciousness is related to the level of "integration" of a system. However, if you come to this while believing in the "Hard Problem", that cannot possibly be true, because IIT relates consciousness to physical properties of the system such as connectivity, but the "Hard Problem" defines consciousness as something which does not arise from any physical property of the system.
Nagel states it most clearly in that science is the view from nowhere. The world doesn't feel like, taste like, look like anything on it's own, because those are creature-based sensations which depend on the kind of sensory organs and nervous systems an animal has.
This argument has played out before so I'll just link to that discussion. quoting from https://www.scottaaronson.com/response-p1.pdf
> There’s two responses to this. The easiest response is to say that φ is merely necessary for C —problem solved. GT’s response would be to challenge your intuition for things being unconscious. Here’s a historical analogy; imagine when the Kelvin temperature scale was introduced. Here Kelvin was saying that just about everything has heat in it. In fact, even the coldest thing you’ve touched actually has substantial heat in it! Think of IIT as attempting to put a Kelvin-scale on our notions of C . I find this “Kelvin scale for C ” analogy makes the panpsychism much more palatable.
Then, Scott's response to that:
> Suppose, again, that I told you that physicists since Kelvin had gotten the definition of temperature all wrong, and that I had a new, better definition. And, when I built a Scott-thermometer that measures true temperatures, it delivered the shocking result that boiling water is actually colder than ice. You’d probably tell me where to shove my Scott-thermometer. But wait: how do you know that I’m not the Copernicus of heat, and that future generations won’t celebrate my breakthrough while scoffing at your small-mindedness?
Ok, pretty dismissive, but this doesn't actually address what the first quote mentions. Maybe phi's necessary but not sufficient for consciousness; that would still be pretty interesting. (I.e. maybe the pesky function does have high phi, but phi is not itself consciousness, because Consciousness(system) = phi(system) + Corrective_factor).
Or suppose we add an axiom like "feedback loops that affect future trajectory" to exclude such a pesky static function definition, or better still suppose IIT moves towards something that generally accounts for more subtle dynamics as well as structure. I can't help but think of the relationship between euclidean/noneuclidean geometry here, especially when "obviousness" comes up in these discussions. It seems productive to play with adding/discarding axioms and looking for richness/consistency. And isn't lots of modern physics about exploring model parameter-space ( https://www.quantamagazine.org/using-the-bootstrap-physicist... ) to zero in on "the" model given "a" model? Like 1D-Ising won't phase-transition, maybe current 1D-IIT won't cut it but has some fruitful generalization.
Would it be interesting? Why?
Coming up with conditions that are neccesary but not sufficient is pretty easy. Sure they are interesting if the conditions are non-intuitive and you've actually proven they are neccesary. I'm not sure either criteria is met here.
> suggest that it's boring/trivial to answer the smallest LLM that speaks english ( https://arxiv.org/abs/2305.07759 ), etc?
That's kind of the opposite question. There's a difference between asking "What is the smallest LLM that can speak english" and "What size must an LLM at least be in order for it to be able to speak english". The former is an interesting question. The latter is trivial as it has an answer of "1". I believe the parent was talking about the latter type of lower bound not the former.
If proponents of IIT object that it is only an assumption that these systems are not conscious, they should be pleased that Aaronson has given them a way to empirically evaluate their thesis!
We are not an automaton that simply eats a strawberry because it is edible. We are decision making organisms that can adjust the composition of our diet based on the chemical properties of the food. We can presume that there is an evolutionary advantage to being able to taste and therefore select from multiple dietary options.
It is clear that the conscious experiences as described are extremely subtle forms of information that allow us to plan and make decisions based on the information they provide us, and not simply blindly react to the world, and I think it is pretty obvious that that is a massive advantage and also more in line with my experience as a conscious being.
that said i'm still keen to see more/better discussion of IIT, and/or more modern extensions. IIT is certainly quantitative, and arguably elegant, despite problems. so it puzzles me how eager some people are to just junk it rather than repairing it.
For purposes of comparison, here's another approach that's super interesting ( https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4168033/ ). So both frameworks are looking at system structure/dynamics and then attempting to deduce/quantify some aspect of them that (hopefully) gives us some insight into "Mind" or "Consciousness". But to me, grounding such inquiries directly in neuroscience seems awkward, because brain only runs Mind. Uncovering a bunch of implementation details about that from FMRIs probably won't give us a lot of interesting insight at the right level. To an extent some implementation details are important and might shed light on the architecture/algorithm of Consciousness/Mind, but evolved systems are also going to be cluttered up with total hacks that evolution brute-forced where implementation is kinda arbitrary.
All of which is to say, IIT being "platform agnostic" and proposing stuff like a thermostat is more conscious than a rock, a dog moreso than a thermostat, and a human more than a dog is not just a cool trick. Being able to approach this kind of intuitive problem in a somewhat rigorous way seems like a necessary requirement for a serious scientific theory of mind. I like the Ising and graph theory modeling approach, but the implementation details of {neuron-counting, FMRI, lesions on functional areas, etc} is probably a distraction if you're trying to understand mind rather than medicine.