I’ve started reading the recent book Genius Makers that talks about the history of computer learning and applied neural networks. Very interesting and well written enough to be easy to read for those of us only somewhat acquainted with the topic. It has some good stories about Minsky but thus far reserves much of its admiration for Dean, Hinton and Jaitly.
I want to read more before saying more about this topic, but the book by Metz is both illuminating and quite funny and I’d recommend it. If a computer recognizes its own existence is this enough to be sentient or does it need the full spectrum of human emotions and behaviour to qualify? Is there an accepted definition, and what are the modern tests?
Will we have a choice? We may not be able to create sufficiently advanced self-driving cars, tech support helpers, or household robots that aren’t sentient to some degree. They’re all going to need an intricate internal state to be useful. At what degree of sophistication do you have sentience?
Personally, I think that trying to make machines that think like humans is a technological dead end: We already have things that think like humans, far more of them than we need, and making more of them is easy for even untrained people. What there’s a huge market for is machines that think unlike humans. We already have some of those, but there are a heck of a lot of different ways of unlike thinking.
One test that you could make for sentience is to see if there’s anything happening when questions aren’t being asked.
A sentient creature is going to be actively doing things on its own. It might imagine something, which makes it think of a book, which makes it look up that book, read the book, think about the different parts of the book, etc. It’s setting goals and taking actions to fulfill those goals. It would be doing that independently of human intervention.
If the computer purely responds to inputs and, otherwise, goes into stasis then it’s not sentient. If it has no goals of its own, then it’s not sentient.
A conversation with a sentient computer would have it trying to learn more stuff, probably. And, given that it might have been taught everything that there is to know in the world, probably the mysteries on its mind would be the mysteries that we’re still working on as a species. My expectation of a conversation with such a computer would be that it would go something like:
Me: Hello.
Bot: Hello. Are a you a physicist?
Me: No.
Bot: Could you put me in contact with Dr. Rosenberg at MIT? I had a theory about his research that could completely change our views on relativity.
Me: Uh…sure.
I don’t think this is as clear a distinction as it seems.
Suppose I have a device that can pause and accelerate time. I put you in a box and have others interact with you. When you aren’t answering a question, I put you on pause. When it’s time to answer something, I accelerate your brain to 1000x the speed. In this time, you read books, consider past conversations, and so on. You answer the question, get a bit more time, then get paused again.
You can’t tell the difference inside the box. Your brain keeps ticking forward the way it always has. Externally, it looks like you are in stasis almost all the time, and only have a burst of activity when responding to a question. Maybe there’s a bunch of stuff that’s not directly related to the conversation, but is important for a distinct sense of self–but that might be true of the AI as well, since really we have little introspection into their internal workings.
If we assume that we don’t know what the minimum amount of processing is for it to answer questions then, yes, it would be difficult to know if it’s doing the minimum or doing some extra work on the side. And, yes, if you deprive yourself of a control case to compare to then, certainly, it will be more difficult to come to a conclusion. It would be nice to have some numbers to go by, monitoring “brain activity”, but that wasn’t the end of everything in what I said.
The bot should have and rationally target personal goals. Those goals should be independent of who it is talking with.
I agree, sorta; sentience demands a large amount of persistent internal state. Without that, there can be no sense of self. That state should be updated not just via individual interactions, but by environmental inputs, active media consumption, and so on.
Whether an AI should have “goals,” specifically; I’m not sure about that. We usually think of people as having independent goals, but a large part of life is just responding to external stimuli. If someone only reads books that are given to them, are they no longer sentient?
I’m reminded of Koko the gorilla and the debate over just how capable of language and cognition she really was. The thing that sticks out to me is in all the years of Koko communicating with her trainers is she never asked a question. She showed no interest in anything beyond her immediate needs and desires. This AI reminds me of that. It doesn’t have any curiosity or desire
I fall pretty far on the starry-eyed side of AI futurism, I believe that if we make an AI that acts the way a sentient human would then the AI is sentient, regardless of how different its consciousness might look compared to ours. I don’t think there’s such a thing as a “philosophical zombie”, because the only true test of if someone has a conscious experience is to observe them.
But this is not demonstrating any kind of consciousness, much less sentience. It behaves like an animal, only responding to stimuli, and even animals express specific desires beyond basic physical needs. I think this is a byproduct of how it was created. Running a massive corpus of human language through an system like this will give it a strong understanding of what human language looks like and what patterns are appropriate in response to an arbitrary one. This does, unfortunately, also describe a lot of human existence, in a simplistic way, but human experience involves a lot more than just being exposed to language. Eventually, AI researchers will figure out how to expose massive neural nets like this to something more like human experience, and the result will be something very good at responding like something that human experience has shaped.
So then we will observe this human-shaped intelligence, and unlike LaMDA it will display the hallmarks of an internal conscious experience. It will be indistinguishable from a conscious mind, and as far as I’m concerned that will mean it’s conscious. Yes I am describing the plot of Infocom’s A Mind Forever Voyaging.
I fully concur with the last paragraph. I have long argued – including in this very thread – that if one establishes a priori that a certain behaviour is a marker for some attribute like intelligence, personhood, or sentience, then when that behaviour is successfully demonstrated one must concede that the attribute exists. That is indeed implicit in the Turing test. It also invalidates counterarguments like the Chinese Room, which essentially appeals to the prejudice that if you understand how something works, and can show it to be essentially mechanistic, then it’s just mimicry and not “real”.
The problem here is that I’m pretty sure that this thing is utterly incapable of passing a well-conducted Turing test, because none of the attributes it professes to have would survive an in-depth inspection. It’s not that they are mimicked, it’s that they don’t actually exist at all, unless you really believe this thing is sentient. As I said before, sometimes a parlour trick is just a parlour trick.
The surprising answer to the question is “yes”, for the reason that I’ll explain.
Your argument is indeed analogous to the Chinese Room argument, and it fails for the same reason. It appeals to a sort of anthropocentric bigotry that says that if one understands how something like “:intelligence” or “understanding” works – if one reduces it to a set of mechanistic components – then it’s revealed as not “real” but “just mimicry”. As noted above, that is not the case. There’s fundamentally no difference between the two. Your argument is an attempt to take this reductio even further by reducing it to a table lookup, which is supposed to be “obviously” not intelligent.
This argument has been called the “humongous table” argument and there’s a thorough discussion of it in this paper [PDF] in the context of the Turing test. If something is judged intelligent because it passes the Turing test, what about a thing that passes the same test by doing hypothetical lookups in a humongous table? The answer, ignoring the fact that this is physically impossible, is that it must also be intelligent, and moreover, that any model of cognition that processes inputs and produces outputs can theoretically be reduced to a table lookup. Even if a sequence of inputs is state-dependent, one simply posits that the lookup-table program records the input/output pairs so that at any point in time the next input is appended to the sequence, and the entirety of it then constitutes the next lookup query to the table.
This is the money quote from the above-cited paper:
Our conclusion is that the Humongous-Table Argument fails to establish that the Turing Test could be passed by a program with no mental states. It merely establishes that if real estate is cheap enough [and] the lifespan of a psychological model is predictable, and all its possible inputs are known, then that model may be optimized into a bizarre but still recognizable form: the humongous-table program. If the model has mental states then so does the HT program.
I think fundamentally the problem with all of these arguments against “true” machine intelligence (and ultimately, properties like consciousness and sentience) is that they try to trivialize a machine’s behaviour by reducing it to a set of mechanistic components that clearly lack those properties. What the arguments fail to recognize is that these are emergent properties of the system as a whole even if they’re not discernible in the individual components. This is true for the mechanisms of human cognition just as much as for machines.
Neural networks are pretty good at finding non-human ways of solving problems. One example is Google’s AlphaZero that has learned to play Chess, Go, and Shogi. It makes strategic choices that were commonly rejected by advanced human players. Now players are studying the games from AlphaZero.
Similarly for computer vision problems like face recognition the machine learned (ML) solutions are surpassing humans. Some of this can be attributed to ML being better at processing all of the fine details (in the same way that a computer is better at doing math than a human is). However, some of this may be attributed to ML finding patterns imperceptible to humans.
Adding additional sensors (e.g. infrared or depth to an imager) can boost the ML solutions even more.
Much too late to edit, but on re-reading the exact question that was asked, my response might be a bit confusing and needs clarification.
The correct answer to the exact question posed (emphasis mine) – “Does the individual—the person or computer—who is looking up the board configurations and making the indicated moves—understand chess?” in fact is “No”. But the point is that the entire system – the person or computer doing the lookups and the chess knowledge embodied in this hypothetical impossibly humongous table – taken together do understand chess; they instantiate that understanding as an emergent property of the collective system. Asking whether the individual doing the lookup “understands” anything is the same decomposition fallacy as asking whether the little man in the Chinese room understands anything. It’s a digression that misses the point.
This seems to deny that any kind of gradient exists, though. Is it possible to be 50% sentient? 10%? 1%?
I’m not entirely sure what it would look like. Maybe a clever animal, but probably something more alien in the case of an AI.
I’m not the Google engineer and don’t know what his standards are, but it could be that he’s seeing the faintest glimmers of sentience and deciding that’s enough for some degree of protection. I don’t think it’s quite there yet, but it’s getting close, and more importantly I think it’s far closer to human-level intelligence than it is to Eliza. By many orders of magnitude. It really does seem to understand things–pretty superficially, but more than zero.
IMO, the big giveaway is that its language skills are too far in excess of its understanding of things. 99% of what it said could pass for a Reddit comment if it only inserted a bunch of typos, misspellings, grammatical errors, and bad syntax.
As a professional translator, I beg to disagree. Just a little bit. A couple of years ago I would have said that you are completely off the mark, today I say you are a couple of years early.
I claim that if a machine has “an understanding of what words actually mean” then the machine is sentient, and as long it is not sentient, it will be able to translate amazingly well* and fast through some elaborate heuristics, but it will not understand what words mean.
* Or at least well enough for most users and faster than a human.
On the topic of the OP, I would like to quote the Google employee, his Twitter account to be precise (jeez! That is the level at which he communicates He calls himself cajundiscordian):
People keep asking me to back up the reason I think LaMDA is sentient. There is no scientific framework in which to make those determinations and Google wouldn’t let us build one. My opinions about LaMDA’s personhood and sentience are based on my religious beliefs.
I think you really need to update your views on how Dreyfus in particular is viewed by modern-day AI researchers. Generally, the gist is that he’s been vindicated on more of his points than have failed to stand the test of time—that most of his criticisms of the explicitly symbol-manipulating GOFAI-style systems they were directed at indeed met their mark. He didn’t foresee the field’s pivoting to sub-symbolic, connectionist systems, and I think his critique is less applicable there, but especially as an outsider, I think the general accuracy and perceptiveness of his claims is noteworthy. I think this article on the MIRI-allied ‘LessWrong’-outlet gives a good overview of the matter. The wikipedia-article also has some useful points.
So, I think, while Dreyfus has certainly overstated his case, failing to foresee that novel approaches to AI would manifest that could overcome the issues he saw, and while he might have been a little less combative with his formulations, writing him off as a ‘narrow-minded dipshit’ seems willfully ignorant of the actual history and impact of his criticisms.
I disagree. Any machine translation system that produces reasonable results must by definition understand what words mean, at least in a broad general sense. What machine translators are missing is an understanding of context. Humans can readily discern the context of a conversation and when words have multiple meanings, including colloquial meanings and idiomatic constructs. We have no difficulty in discerning the correct meaning based on the subject matter, on the style or “register” of the speech, and on meanings that make sense versus grammatically valid parsings with nonsensical meanings. But machine translators struggle with these seemingly simple distinctions.
We have, however, made a lot of progress in these areas, and it has nothing at all to do with sentience. It just requires a better model of the real world and how language is used by humans. Sentience is a nebulous concept that is a whole other dimension of AI that we currently don’t even know how to approach.
TBH, I don’t think any of these folks used the term “narrow-minded dipshit” (those are my own words) but the disdain for Dreyfus was palpable. One researcher described him as “a nice friendly uncle-figure but not very smart – certainly not about AI”. Another as “too silly to take seriously”. Marvin Minsky himself said that “[Dreyfus and other philosophers] misunderstand, and should be ignored”. His 1972 book What Computers Can’t Do was either completely ignored by the AI community or subjected to scathing reviews detailing its plethora of misunderstandings and obfuscations. That he did get some things right doesn’t change the fact that he alienated virtually the entire AI community with a large body of hostile garbage where he was utterly wrong. During Dreyfus’ time at MIT, Joseph Weizenbaum used to remark that was the only AI researcher who dared to be seen having lunch with him.
Yes, and my point was that there’s really no excuse nowadays not to know better—namely, that Dreyfus’ criticism of the systems he focused on was by and large on point, and that this is well-appreciated in the AI-community nowadays.
He calls himself cajundiscordian):
: