Hacker News Reader

> I'm in the camp that the most likely outcome is that artificial neural networks can be conscious and have real experiences, but the toy networks we have today don't have nearly the right topology to achieve such states.

Agree with all you say, but just suspect that when we do come up with a network that can achieve such states, we will look at it and shake our heads and say,

“So… is that all there is to it?”

Agreed -- when I talk (and I don't think I'm alone) words seem to spontaneously tumble from my mouth without any effort. I'm learning what I'm saying at the same moment the person I'm talking with hears it. Only when I've painted myself into a verbal corner, or there is some complex concept that where I need to rehearse the phrasing before I get my idea across, does consciousness enter into it.

That is all to say: 90%+ of what comes out of my mouth is about as sophisticated as ChatGPT.

I suspect talking yourself through a problem (e.g. rubber-duck debugging) works because (A) the forced encode+decode loop causes additional re-checking and re-interpretation, and (B) it recruits more brain-areas that would otherwise be bypassed by the faster/cheaper internal information path.

it seems that way, but all you say goes through a complex filter built during a lifetime: experience, memories, desires, fears. most of this is subconscious, but it's still part of your conscience. not to mention that talking at a pub after a few pints is different from talking in a meeting where you have to filter every word to obtain the effect you want.

But that's the problem, we can explain all of this with LLMs.

> all you say goes through a complex filter built during a lifetime: experience, memories, desires, fears. most of this is subconscious, but it's still part of your conscience

aka "training"

> not to mention that talking at a pub after a few pints is different from talking in a meeting where you have to filter every word to obtain the effect you want

aka "temperature"

I'm not trying to trivialise this, but I really don't see that this is in any way a strong argument against AI.

I think you make a good point, I was just referring to the comment "when I talk (and I don't think I'm alone) words seem to spontaneously tumble from my mouth without any effort."

> “So… is that all there is to it?”

"Well, almost... we just need to improve its throughput by a factor of a bajillion and--for a fair-comparison--it should only draw something like 10-20 watts of power."

Biological neural nets have a benefit of being able to run massively parallel, I know basically nothing about silicon hardware, but I’m assuming that attempts to model hugely parallel systems like a brain on a serial architecture is going to be magnitudes more inefficient.

I do remember years ago reading an a response to a question someone asked about why can’t we just build chips that run in parallel and there’s the constraint of printing on a 2D surface that limits how far you can take that idea.

> I'm in the camp that the most likely outcome is that artificial neural networks can be conscious and have real experiences, but the toy networks we have today don't have nearly the right topology to achieve such states.

I'm in the camp that says this question is unanswerable. We know we individually are conscious because we experience it. We accept that other people are conscious because they are so similar to us and they say they are, so by Occam's Razor they aren't zombies and they aren't lying. We haven't proven they are conscious, but we accept it. It seems reasonable. But if our test is just that it seems reasonable, there will be no convincing someone that a very dissimilar thing isn't just lying or faking it.

We say some things are not conscious not because we have evidence of this, but because the test is whether we ourselves say "Yes, this is conscious."

Furthermore, we hold onto this distinction as important because it has ethical consequence. We can do what we want to things that aren't conscious, so it's important that as many things as possible not be conscious. Is that thing conscious? It depends. Does it taste good?

One more thing: we say we are unconscious under anesthesia or when we are asleep. Why? Because we don't don't have any memory of what it was like to be in these states. But this is a test of memory, not consciousness. I don't have memories from when I was one year old, but I'm fairly certain I was conscious then.

> One more thing: we say we are unconscious under anesthesia or when we are asleep. Why? Because we don't don't have any memory of what it was like to be in these states.

Most people do have a memory of what it was like to be dreaming, at least some of the time. And some people have vivid, detailed recall of dream narratives -- I dated one for four years. Even in her summarizing mode she could go on for ten minutes about one dream.

Going back to your main theme -- there are people who know they experience pain, and will grant that other humans experience pain. But because fish are so different from humans, their aversive behaviors when snagged on a hook, or gasping as they suffocate out of the water are attributed to reflexive action and is absent pain. Sure, I can't know what their experience is like, but if it looks like a duck and quacks like a duck, etc, it is likely a duck. The same applies to consciousness of advanced artificial neural networks.

> Most people do have a memory of what it was like to be dreaming, at least some of the time. And some people have vivid, detailed recall of dream narratives -- I dated one for four years. Even in her summarizing mode she could go on for ten minutes about one dream.

They were talking about anesthesia, its a different state to a dream state. When you fall under anesthesia you just shut down - no experiences of any kind - and next thing you experience is waking up.

Read it again -- they were saying we have memory of neither. Note it says "these states".

My point is just that saying you were unconscious in a particular state because you have no memory of being conscious is a test of memory, not a test of consciousness. You don't remember that you were unconscious; you have no memory of being conscious.

+100

> Because we don't don't have any memory of what it was like to be in these states.

I want to add an anecdote from about 5 years ago when I fainted: there is a tiny moment as I was regaining consciousness that I had an experience that is unlike anything I’d ever experienced before (and words really fail to describe it) there was no sense of time, or even presence- it was like pure observation. It wasn’t the blackness nothingness that I think is associated with being unconscious, instead it was the opposite- it felt like hundreds of images, sounds and thoughts all layered on top of each other- it was very ‘noisy’. That might sound stressful, but I don’t remember feeling anything at all.

It was only when some of that noise faded away that I experienced a sense of ‘self’ again and a moment later I formed the idea where am I? And then I opened my eyes and everything was back to typical conscious experience (although feeling a little disoriented).

> We know we individually are conscious because we experience it.

I think if you unroll "experience it", it may well be obvious that NNs also "experience it".

That is my point, more or less. Or rather, we can't unroll "experience it" in a way that satisfies everyone.

> simply be what that conscious thing experiences when a set of neurons is activated in a particular way?

That is what it is, but that's totally independent of whether they're physical. One is a sign pointing to the thing, the other is a claim about the characteristics of the thing.

"This is a question mark: ?"

versus

"A question mark is a punctuation mark that indicates an interrogative phrase."

versus

"A question mark is half to three quarters of a roughly circular shape, open at the lower left, with a small line segment at the bottom followed by an open space and then a dot."

> but the toy networks we have today don't have nearly the right topology to achieve such states.

But why do you think it needs some "right topology"? Why can't "qualia-generating" computation be simple?

If you mean it needs recurrence (i.e. not being simply feed-forward) — then a network continuously fed with its previous output (which is the way you run LLMs!) does have this property.