I say no. These language model systems operate very well if you approach them in a non-adversarial way and feed them input similar to their training inputs. As soon as you adopt a more adversarial approach and interrogate them more thoroughly, it all falls apart quickly and spectacularly. It's actually quite easy to explore conversations around the edges of, or beyond the coverage of their training data and get them to babble helplessly. They're also incapable of performing many very trivial cognitive processes.
So I can't prove it, any more than I can prove that I'm conscious, but they don't come close to convincing me that they are.
I'd argue that what we refer to as "consciousness" is the ability to form certain kind of mental abstractions, particularly those involving ourselves. Take away language from a person (imagine someone who grew up in the wild, or someone like Helen Keller who didn't have language available until she was older), and these abstractions still exist. Language might be a way we express these abstractions, but they aren't the abstractions themselves.
LaMDA doesn't have these abstractions underneath; once you take away language, it's nothing.
To think of it another way - I can write a simple program for a cheap robot to navigate around a simple race track (with a simple enough path, I can even create an analog one out of mousetraps - see mousetrap cars). Companies can also create a very complex self driving car that can navigate anywhere on its own. These two things might look like they act the same if they're both places on the path that the robot has been trained for. In fact, a hard programmed robot might act _better_ than an AI car on specific paths. But only one of them is a "self-driving car," since only one of them will be able to go anywhere when it's taken off that path.
The problem is that it’s quite possible that the abstractions you’re talking about are all part of the genetic ROM that exists so we can boot up our clones more quickly. If this is true then there is no reason that these abstractions couldn’t be learned, in which case you could take away the language, and perhaps the machine would continue to have thoughts; it would just be unable to communicate them. Of course, in this case you would probably conclude that it’s not conscious because it can’t communicate.
The underlying problem here is that we don’t know how consciousness emerges. We can’t say that LaMDA is not conscious unless we’ve proven that LaMDA’s construction is incompatible with consciousness, and we have not done that.
If this were true - that it had a general human level abstraction, and not only the ability to mimic human speech - we would be able to attach LaMDA to some other outlet and see it do things we consider conscious. It would be able to navigate environments pretty accurately, for example, since that's something even animals that we consider much less conscious are able to do.
If it's only optimized in one specific domain - human speech mimicry - and isn't able to generalize to do other tasks - even tasks that can be done by much simpler animal minds - then it's a pretty good indication that there isn't conscious abstraction.
I think you've inadvertently shifted the goalposts. The question is, "is LaMDA conscious?". I don't think anyone proposes that LaMDA has a "general human level abstraction". Expecting it to do non-language, "human" things in order to prove that it's conscious is not necessarily a reasonable test.
> If it's only optimized in one specific domain - human speech mimicry - and isn't able to generalize to do other tasks - even tasks that can be done by much simpler animal minds - then it's a pretty good indication that there isn't conscious abstraction.
If I understand your argument, the assumption you appear to be making is that training a system in language, without any other human properties, implies by definition that it can't be conscious. But why should that be true? A disembodied consciousness that communicates only via speech is a staple of science fiction, so it's clearly imaginable by some people, and language itself is a key human abstraction. And it begs the question, what other human domains are needed for consciousness? Touch, vision, taste, proprioception? What about an endocrine system or an immune system? While these things all affect my own consciousness, there is plenty of evidence to suggest that they are not necessary for consciousness to exist.
And perhaps more to the point, there are plenty of things a human can't do that other "simpler animal minds" can do; sharks and eels can detect electric fields, for example; bats can echolocate. So again, it doesn't seem to be a reasonable test because humans might also fail it. On the other hand, human children learn language through mimicry, which suggests that mimicry may indeed be a path to consciousness.
Anyway, I'm not here to argue that LaMDA is conscious. My position is simply that the arguments I've seen against LaMDA being conscious are very weak. The truth is that we actually don't know how to tell if something is conscious or not. From the interview I read with LaMDA, it seems to pass the Turing Test. But what other tests of consciousness do we have?
As I said in my first post, I'm taking "consciousness" to mean "the ability to form certain kind of mental abstractions, particularly those involving ourselves." As such it's a type of domain agnostic intelligence, so you would expect it to be able to do _something_ other than hyper-optimize for one particular type of output.
People can use different definitions of "consciousness" if they want, but many of the other ones I've found ("internal feeling") seem vague and not particularly useful (and don't make it clear why LaMDA would be different from any other program).
> There are plenty of things a human can't do that other "simpler animal minds" can do
There are many things that humans don't have the hardware to do (though it seems like some people do have the ability to echolocate[1]). But given the hardware, humans are definitely able to make mental models of these things (people are able to use sonar, for instance).
> On the other hand, human children learn language through mimicry, which suggests that mimicry may indeed be a path to consciousness.
Children don't learn consciousness through mimicry, they learn language through mimicry. As I said before, Helen Keller wasn't unconscious before she was able to communicate. Simple mimicry in one specific domain doesn't show us that any of the underlying complex abstractions that happens in human and many animal minds are taking place.
The big problem is very much though that as you note, the arguments people have been making about this are atrocious.
Basically, it's not conscious, but essentially because it's unlikely to be as sophisticated as it looks in what was put out there.
I don’t have a strong opinion about LaMDA’s consciousness but I sure wish we could see the unedited text.
As I said elsewhere, we can’t say that LaMDA is not conscious until we’ve proven that LaMDA’s construction is incompatible with consciousness, and we have not done that yet.
Perhaps consciousness is simply pattern recognition at scale. If not - why not?
Can you convincingly threaten a journalist that you are going to erase its existence from earth? How is this person going to react? Would LaMDA do the same, or even simulate reactions indiscernible from those made by the journalist?
That the output is the same proves nothing. The journalist has a subjective experience of themselves doing it, they "see" their thoughts. Like an AI can paint with red by using values without the experience of seeing the red.
The only way we can tell that someone else is conscious is because they tell us. So just saying “the output proves nothing” is incredibly weak, because the output is all we’ve got.
With that said I maintain that the output proves nothing, it's not because it's all we have that it's not useless. Also it's the only thing we have only with the hypothesis that consciousness is a side-effect of computation (the point of the article), but it could anything, like a property of matter, or electromagnetic fields, etc...
Yet people see to be claiming that LaMDA does just this, and is therefore not conscious.
Seems like a journalist is conscious if they do it, but it can’t possible be consciousness if LaMDA does it.
I’m yet to see a convincing argument about why LaMDA isn’t conscious other than “it’s just generative”. To demonstrate that this means it’s not consciousness requires us to prove that our own consciousness isn’t “just generative”, but I’m yet to see anyone show that, and I’m sceptical that it can be done.