Few philosophical definitions actually involve qualia or handwaving, because serious philosophy is quite careful about noting vagueness (handwaving) and circular defintions (qualia).

Also, saying "I define consciousness as X" and then going on to show that, because I decree consciousness to be X, we don't need to worry about Y, Z, and so on, isn't pragmatic at all, unless you pick an X that everyone agrees on. The definition you give is something hardly anyone would agree on.

Social consensus probably isn't even an aspect of consciousness, but at least one can make that argument coherently, if you rephrase it as "part of the decision to call something 'conscious' or not involves social consensus". But this is also trivial and boring and not why anyone actually cares about the "is AI conscious" debate.

▲

orbital-decay 5 hours ago | parent [-]

Meanwhile that's exactly what we see now, do we not? Whether AI is conscious will be determined by the social acceptance of that idea, because how can you prove or disprove it without using a ton of questionable assumptions? How can you prove that I'm conscious, for that matter? Call me ignorant, but I see the entire consciousness debate as pretty irrelevant to what the thing is.

I agree that the article is all over the place, though. (not that my posts are much better)

▲

D-Machine 5 hours ago | parent [-]

Whether AI is conscious or not is independent of the social acceptance of the idea, unless you believe in thoroughly discredited and quite radical forms of social constructionism.

In fact, if you stop talking about consciousness, and go through all the various aspects of consciousness, the whole debate gets very interesting. There are various ways in which AI is very clearly conscious or exhibits / involves phenomena that are clearly involved in certain aspects of conscious, other ways in which AI already very clearly is missing other aspects, and then again some remaining ways which are unclear, and which may forever remain unclear.

Being clear about the different ways in which AI is, isn't, and might be conscious is a useful discussion and we still need to be having it.

EDIT: An analogy would be to recent shifts in whether or not lobsters feel pain, and so by extension, the ways in and degree to which lobsters can be considered consciousness. Research and debate has shifted opinions enough for parts of Europe to regulate certain methods of preparing them. But the ways in which lobsters are conscious and/or feel pain is an objective fact, regardless of what we decide.

▲

orbital-decay 5 hours ago | parent [-]

>Whether AI is conscious or not is independent of the social acceptance of the idea, unless you believe in thoroughly discredited and quite radical forms of social constructionism.

Are you saying I need to have certain beliefs (discredited by someone) for something objective to be true or false? That sounds like social acceptance of an idea.

>There are various ways in which AI is very clearly conscious or exhibits / involves phenomena that are clearly involved in certain aspects of conscious

Which ways and how can it be demonstrated experimentally?

>An analogy would be to recent shifts in whether or not lobsters feel pain, and so by extension, the ways in and degree to which lobsters can be considered consciousness. Research and debate has shifted opinions enough for parts of Europe to regulate certain methods of preparing them. But the ways in which lobsters are conscious and/or feel pain is an objective fact, regardless of what we decide.

That's not what the consciousness debate I see is about. The signal propagates and causes the reaction, but do they feel pain? What is the threshold? Is it just the complexity of the state? How complex it needs to be?

▲

D-Machine 5 hours ago | parent [-]

> Are you saying I need to have certain beliefs (discredited by someone) for something objective to be true or false?

I am saying that, if you reject radical social constructionism (or solipsisms of various flavours, i.e. we are all just plugged into the matrix or some kind of simulation), then you must believe there is a real world out there independent of our consciousness / awareness of it. There must also be patterns / structure in that world, some of those patterns and structures being essentially constant, and those are what we call "true" (objective) facts about reality. BUT if you are a radical social constructionist or solipsist of some kind, then it can be perfectly coherent to reject any idea of objectivity or truth, except as contingent on social consensus (or personal beliefs, in the solipsistic case). Radical social constructionism being "discredited" is just a recognition that it basically forces you rapidly into incoherence.

> Which ways and how can it be demonstrated experimentally?

Try reading through the SEP article I linked above (https://plato.stanford.edu/entries/consciousness/#ConCon), particularly sections 2, and then especially section 4. You'll see you don't even need to run experiments to answer certain questions, you can just logically prove there are certain aspects of consciousness LLMs can't have, or clearly already have. Others you'll see immediately need qualifying or more information to answer. EDIT: Or just ask any AI to list the ways in which LLMs can, cannot, and might be able to exhibit the various aspects of consciousness. It will give you a decent answer, likely in a pretty little table.

See e.g. 4.8, Dynamic Flow: LLMs basically can't have this, as their weights are frozen. You'd have to claim that the activations during inference are the dynamic component, but then that means LLMs are conscious only in the few seconds or even milliseconds during inference. Likewise, they don't have memory, so every inference is this consciousness being born anew, as it were, at least in current models.

You can of course go much further on even this specific point, but, the point is, there is lots of useful discussion and thinking to be had when you go through the aspects of consciousness in this way, and definitely at least some clear answers.

EDIT:

> That's not what the consciousness debate I see is about.

I just mean there are some similarities between the lobster consciousness debates and AI consciousness debates. Obviously the stakes and many aspects are different, I was just trying to illustrate why and how discussion and experiments can be useful and consequential here.

▲

orbital-decay 5 hours ago | parent [-]

>See e.g. 4.8, Dynamic Flow: LLMs basically can't have this, as their weights are frozen.

So consciousness = neuroplasticity? Requires neuroplasticity? Sounds pretty arbitrary, just like the entire link is an arbitrary set of ill-defined criteria. Besides, LLM weights are far from frozen if you look at the entire process in its specific time projection.

In fact the entire definition of a dynamic flow in that article is so vague that you can apply it to an industrial arm guided by a few dynamic equations, as it puts no definite threshold on the complexity of the process or behavior.

▲

D-Machine 4 hours ago | parent [-]

You are still stuck in thinking consciousness is one unified thing, but it is about one of the clearest examples of a family-resemblance category as you can think of.

Dynamicness / change is one aspect that is definitely part of what we tend to think of as being quite fundamental to consciousness, and there are a lot of ways in which LLMs can't and/or don't really change, and this strongly constrains how similar LLM consciousness can/could be to ours.

If the SEP isn't enough to trigger the thinking needed for you, just ask an LLM what I said, to "go through the ways in which LLMs do/don't/may exhibit phenomena that are generally what we consider fundamental aspects of consciousness". You'll see there are some clear answers, and some uncertainties and vagueness.

▲

orbital-decay 3 hours ago | parent [-]

Then I suggest you do the same for rigorous definitions that don't rely on consensus, and maybe also addressing my point about non-frozen weights directly.

	▲	D-Machine 2 hours ago \| parent [-]
		LLM weights are frozen once the model is trained and a specific version is released, or this is true at least currently for almost all major models. You can ask interesting consciousness questions about LLMs, but you most definitely need to separate LLMs during training (whether this is pre-training, later fine-tuning, or general instruction tuning approaches) and after, during inference. None of this has any relevance to the core points I was making re: useful vs. useless discussions of LLMs and consciousness, because the weights point was just a minor example to illustrate a larger point about how narrowing the scope of the conversation is necessary for productive discourse. I doubt we agree on what "rigorous definitions" means, given that you are happy to continue using a useless, vague family resemblance concept like "consciousness", but complain that pointing out dynamicness and/or change is involved is too vague. I believe one cannot give a rigorous definition of consciousness precisely because it is a family resemblance concept, and there is no actual essence. This is why I suggest shifting to talking about aspects, since then productive insights and truths can be discovered.