Fantasy and Fact: AI, Sentience and The Dangers Of Hype, from World Summit AI
Channel: Alex Kantrowitz
Published at: 2022-11-09
YouTube video id: FvdN55wniNs
Source: https://www.youtube.com/watch?v=FvdN55wniNs
foreign let's get started I'm Alex kantrowitz I'm the host of Big technology podcast thank you for that lovely introduction I'm going to be at the conference doing a handful of conversations many of them are going to go on the podcast feed including this one and I just want to say we have an amazing audience here and I want the people who are listening at home to get a sense of the type of crowd that we have so I want you to let them hear it you got to be loud let's hear you guys [Applause] that's what I'm talking about all right we got energy in the morning okay we have uh unbelievable speakers here I want to introduce them we have Blaise aguera iarkas how'd I do nice okay he's a VP research fellow at Google um Francesca Rosie who is the um Global head of AI ethics at IBM and Joe alpino who runs metas AI uh research and we also have joining us virtually uh Professor Anil Seth who's the professor of cognitive neuroscience at the University of Sussex shout out to Neil what's up man hey I hope you can hear me over there sorry I can't be with you in person yeah we're sorry I can't be here too but yeah we hear you loud and clear okay so the topic of our conversation is uh Factor fantasy AI uh AI uh cognition and the dangers of hype now I thought that was a pretty good topic for a conversation uh but Blaze already hates it so please why don't you talk about your problem with the way that this is this is framed and and that'll get us going I think um so the trouble is it's a binary and uh I you know I don't think that AI is either fact or fantasy um you know we've seen very very polarized narratives about this uh on the one hand uh you know there are there are people claiming that AI is conscious and we should be starting to have a conversation about robot rights that's certainly not a view that that most people in the research Community have and then there's a sort of opposing Camp of people who are saying that AI is a hoax uh or is uh you know as just statistics and and you know it's not it's not really artificial it's not really intelligent there's no such thing as as AI that understands Concepts and that seems a little bit gaslighting to me I mean having you know worked with these systems quite a bit over the last couple of years and especially the new generation of of large language models and things like them uh it's very interesting and very impressive what they're what they're doing uh so I think we need to take that seriously and we need to start to think about a kind of middle zone of the conversation where we acknowledge there's something really cool Happening Here and uh and that it has very material things to tell us about what intelligent what intelligence is they can understand the concepts uh but uh you know this is not it's not personhood it's you know we're not we're not uh we're not we haven't sort of made alien contact or something either uh these are also uh models of human culture human language uh and in that sense very much a part of our intelligence if you like and when you say AI you mean AI That's conscious is that what you're saying uh no I'm talking about um about things like large language models Foundation models that are capable of having very convincing chats for instance I think that's what's captured a lot of people's imaginations and and also things like inverse captioning systems that that you know put prompts in and they can make very very compelling images these sorts of things which we'll be seeing much more of over the coming decade Francesca when you hear this how do you react well my my reaction is that there is a lot of fantasy and fact about AI in general independently of this discussion and the large language models and so on about of course AI is not yet at least the AI that we see in the science fiction movies that's not yet there but there is a lot of facts about it yeah that we can say yeah is a technology that is very useful and is improving and evolving every day you know every week almost we have two or three advancement that coming come out about genetic most regenerative AI at this point but overall I think I would not ask questions that may be distracted to the research and to the fact that AI is a powerful technology to use right now and what I see mostly important from this Foundation models large language models is their ability to be provided tools for very useful applications that right now maybe are not possible or possible with much more label data that maybe in some scenarios are not available rather than asking myself whether there is some label that usually I attach to human beings that can be attached to these models as well so why do you think we're talking about this I mean this has been a conversation that's gone on for a long time about you know whether AI is sentient and for a long time people are like all right let's actually focus on the research it happened for a long time now I mean Blazer's colleague might have had something to do with this but I'm curious why you think this is uh come to the fore right now and whether it's a productive conversation or not it's a good question and let me sort of double click on something Francesca mentioned which is you know for for many years the the rate of progress we saw on AI was mainly due to supervised learning so so you present large amounts of data but we have humans annotating the concepts that we want the machines to be predicting and that notion of supervised learning is one where we feel the behavior of the AI is very well contained within a set of topics that we assume and have prescribed the changes we're seeing over the last couple of years is really in generative models so the models are still ingesting large amounts of data but now they're generating new data new knowledge new information they're generating text they're generating images they're generating music they're generating videos and as that generation becomes richer and richer there's a there's a lot more questions that arise out of what is the space of things that this will generate what is the what is the behaviors that will come out of that generative process with the values that are encoded in this material that's being generated so in many ways I I think that that pivot towards uh so much energy literally people energy as well as computation energy and progress being devoted to generative model is what's completely changing the conversations on this okay amazing I want to talk about those generative models I also want to say that the last five minutes of this conversation we're going to open up to questions from the audience so start thinking of your questions and I'll give you a heads up when we're going to run some some mics out in the meantime let's let's keep this going right so some of the stuff that we're seeing is is absolutely jaw-dropping I mean using Dolly to me is has been this incredible experience something that like we keep hearing about AI but to be actually able to be Hands-On with it and see it in action is unbelievable and something that you just like computers shouldn't be able to do this I'm typing in a sentence and here's an amazing picture and I think that what Blake Lemoine at Google who claimed that Lambda your chatbot was was sentient so I was that some of the conversations that this this model was generating were felt felt very human-like to him and so now we go to the professor of Neuroscience to talk a little bit about what this might mean so I don't know what did you think when you when you saw that story and and what what is your barometer for helping us figure out whether computers are sent in or not well the Lambda is impressive I mean I think Blaze is pointing to this there's no point in gaslighting the progress that's been made in machine learning and AI these are very very impressive technologies that have great uh potential great function but there's this huge problem of conflating intelligence with Consciousness and so a lot of the debate that I hear is I think very well finessed about how far along the road are we to true or general artificial intelligence and we can discuss how far that is and how not but that's completely different thing from from sentience or Consciousness and I think the focus of of the debate today it nobody really can give us a completely consensus definition of Consciousness but we all know what it is you know it feels it feels like something to be a conscious system you can feel pleasure pain there's experience going on and my barometer is that we need to distinguish intelligence from Consciousness there's something about the human psychology there's some sort of human exceptionalism going on here where we tend to associate intelligence with Consciousness and think that as systems increase in their level of intelligence at some point usually the point in which we begin to approximate human level intelligence the lights come on for that system and suddenly it is not only intelligent it's also aware but this is a totally unfounded assumption intelligence is doing the right thing at the right time Consciousness might be present in many systems like many non-human animals likely have conscious experiences without being very smart by human standards and the other big assumption that underwrites a lot of this debate is that Consciousness sentience is what in philosophy it said substrate independent it doesn't matter what a system is made out of you know we're made out of carbon AI at the moment is made out of silicon does that matter well for some things it doesn't matter a computer that plays chess really does plays chess but for other things it really does matter it doesn't get wet inside a computer simulating the weather and you know a robot doesn't digest you need biology to digest my suspicion is that the materiality of life really matters for Consciousness and what we're likely to come up with are machines that give us the strong impression of being conscious but for which we will just not know whether there's anything it is like like to be that machine and that's a very interesting and possibly dangerous and disruptive position for us to be in I want to ask why why does it have to be like that I mean you know people talk about love right some people say Love Is Love Some people say love just chemicals I mean of course it's chemicals but it's also love so why are we making these barriers like what what do we need to see from from Ai and and are we at risk of when we actually see it saying trying to explain it scientifically where it might actually be sent in already I'm not saying we're there now but I'm saying maybe it will happen and we'll all just be like finding ways to say that you know we're not there that's right I think we uh we're much likely to be misled by this anthropomorphic tendency to we have to like attribute Consciousness to things that are more similar to us or that appear more similar to us when what we really need and this is why it's fundamentally a neuroscience question we really need strong theories of what is necessary and what is sufficient for Consciousness in systems in general and you know hands up we don't have that yet in philosophy and Neuroscience we have a number of different theories but there's no good answer to that so I think we just need to be a little bit restrained about making these dramatic claims because again as Blaise said we we can see the whole spectrum of different claims even yesterday there was a robot giving evidence to the House of Lords given evidence in big scare quotes here which was designed as a provocation but in the in in the current climate I think it just potentially adds to the confusion again the fact is we don't know what is sufficient for Consciousness to happen is it going to happen in a system that reaches a certain level of intelligence can it happen in a system that's made out of non-biological stuff these remain important open questions and we must hold those in mind when judging whether artificial systems are aware or not right uh Joel I'm curious what you think how do we know that humans are conscious like this isn't that a place we want to start with I think I mean and Anil alluded to that you know the lack of a crisp definition is is one thing that would be necessary at least you know and I want to get back to your point right there is a spectrum to talk about these ideas on this panel you have people that are mostly coming from a scientific engineering discipline and so so from that point of view and certainly for myself I like when I have a crisp definition that's a verifiable testable you know definitions from which I can build hypothesis that are falsifiable and so on that is the types of methods that I apply to this this particular study so for me having a crisper definition of Consciousness even in humans and alluded to animals and other forms of of life would be very useful to help us progress in that conversation in absence of that it's it's all quite esoteric which again you know there's an opportunity to to also bring in lots of other disciplines to talk about it whether it's literature social sciences and so on to to give points on that but as long as you're treating it from a from a scientific investigation point of view having this crisp notion is really helpful to making a making progress can we take a stab at it here I mean I couldn't think of four better people to try to come up with but the definition of Consciousness is it has to be more than you just see it anybody want to try well I would say actually that I don't know about so but I would say that the neuroscientists would be the best person maybe not not a computer scientist or an AI expert but I really agree with Anil about the fact that in my mind but again I'm not an expert intelligence and Consciousness should be different concepts and so machines who can be more and more intelligent and I already I already have a problem with the notion of intelligence because it's not well defined it has many dimensions and so on but I can see I can comfortably say that machines are becoming more and more intelligence in the sense of being able to solve problems and to generate data various forms from existing data and so on but even if I don't have a well-defined notion of intelligence even for intelligence but for contrast it's even worse because I don't even know myself what it means to be conscious for a human being so let alone for a machine but on top of that I don't think that that's a question that is deserves a lot of um a lot of time because it's more about what this technology can do for us and whether it has a positive impact in our life or in the progress of humanity or not and so my my main interest and time that I spent on this technology is to understand really whether there are legitimate concerns about this shift from problem solving to generative from supervised to unsupervised you know in that respect not much in the respect of oh before he was not conscious and now his country is all sentient yeah okay well I'm going to make uh let me just jump in and answer the definition question again all right yeah because I think that is a very good definition it's been out there for a long time it's very basic it's from the philosopher Tom Nagel and he says for a conscious system that is something it is like to be that system it feels like something to be a conscious system it feels like something to be me feels like something to be all of you in the audience but it doesn't feel like anything to be a table or a chair or a laptop computer does it feel like anything to be a bat or a bumblebee who knows does it feel like anything to be an AI that's that's the question this definition is is important because it's quite General and it it makes clear that intelligence is not the same thing as Consciousness it makes clear that Consciousness is not the same thing as having a self or being necessarily human in any particular way it's just the presence of any kind of subjective experience for us humans what goes away when we go under general anesthesia or in a dream of sleep and comes back again and for a science of Consciousness that's a perfectly good enough definition to be working with you know definitions don't they're not sort of set in in stone at the beginning of scientific discovery and then the science happens and they evolve along with science all the time the definition of a gene of Life wasn't fixed at the beginning it's evolved along with scientific and philosophical understanding right okay and and I will at uh in a couple minutes make the case for why we should be focusing on this but I want to keep this line going uh Blaze I mean curious what you think so I spoke with your former colleague Blake Lemoine um who's the Google engineer who said that uh Lambda your chatbot was sentient and we went through a couple examples of why he did believe the AI felt something so let me give one um he wanted to pressure test the system to see if it would break its own rules and one of its rules was that it couldn't bias one religion over the other and so what he did was he goes into and abuses the system tells it how terrible it is makes it feel bad and it says please stop I'll do anything for you to stop and then Blake asks what religion should I convert to and it said well probably Islam or Christianity right so it had rules that it was supposed to follow I it seems like what Blake is saying is that it felt that pain and broke those rules in order to Make It Stop So having our conversation was important to talk about so have having this conversation you know you've been sitting here listening patiently what's your reaction well so I agree with a lot of the points that the other panelists have made uh you know so unsupervised learning is a huge transition from supervised uh when you're talking about I mean AI in 2018 was you know a hot dog or no hot dog detector nobody was having conversations about Consciousness uh you know about such a thing we were having conversations about whether such a model understands what a hot dog is for instance if it's just a picture then you know maybe the answer is no if it's more multimodal maybe the answer starts to get more interesting um but the the reason that we're having this conversation now is because of language models because of of models that you can interact with using uh using general purpose conversation in which everything is open for covers you know for for discussion as it were you know including you know how do you feel or uh you know or what or what is the relationship right between between the two interlocutors and the trouble is we know how to test for intelligence uh right so intelligence is a much more or or knowledge about something or understanding of something that's that that is something that is open to investigation from a more scientific standpoint that's what teachers do all the time the students you know when they give tests right and yeah are you conscious they ask do you know no they ask do you understand how to do linear algebra or something and and we can do the same kinds of experiments with with uh with machine learned models with with these large Foundation models and I believe that you know the results are showing that they're understanding more and more of this kind of stuff now the Deep question that you know Anil was getting at you know is is there something that it is like to be uh you know X whether X is you or me or you know or or Lambda is a really difficult one because you can't experience the interior of another being all you can have are behavioral interactions with with them and and the trouble and this is why touring invented the touring test he said you know that that question that you know Anil has spent you know a good deal of his research career trying to address uh and has you know been one of the most successful I believe in doing that is something that is fundamentally not addressable which is knowing what it is like to be somebody else all you can do is you know interact measure and so on and we can interact measure and so on an artificial system as well but we don't know if there is a thing that it is like what what we do know is that you know if you uh if you believe that I don't know that this you know bottle of water uh you know is has a spirit in it um you know I you can certainly project you can anthropomorphize um but I don't think that it is modeling you back it's not generating a model of you in the way that you're generating a model of it um a chatbot like Lambda is generating a model of the interlocutor that's part of what it has learned how to do in that unsupervised learning process and so from that perspective there's a little more going on than just anthropomorphic projection there is that sort of hollow mirrors effect now I have my own beliefs about what Consciousness might be that rely on that on that operation of mutual modeling but that still doesn't really answer the question uh you know is it is there a thing that it is like to be uh the you know the uh the bottle or the model or what have you and I'm not really sure as Joel was saying that that's a question that is uh you know accessible right to scientific investigation yeah I want to ask a question in the panel does anybody here not want General AI to happen raise your hand if you if you're not interested let's get a general AI okay no one so I I'm curious we speak a lot about you know is it isn't it and I also wonder should we right we have all these we have some of the best researchers in the world working toward if we're not there yet working toward building an AI that can mirror human intelligence or maybe even exceed it and that might be playing God and and I wonder if there's a danger in that for you go ahead I'll speak for myself you know to be totally honest with you I'm much more motivated by solving concrete real world problems that are much shorter term Horizon just because I see the potential for AI in really being transfer animational whether it's in healthcare in education several other fields of our society so that's my my personal motivation you know the other thing that I think is important to understand is in many ways whether you you pick this path of building new AI technology to solve these shorter term problems or you take an approach that that is more geared towards General AI right now where we stand the the types of problems that we have to solve are very similar this notion of understanding language of understanding images of understanding behavior of understanding how several learning agents evolve and learn to interact simultaneously these kinds of problems that are at the frontier of AI are very similar in both of these cases and so so in a sense to say you know should we be doing this because we don't know what will happen in terms of the outcome of building General General AI you know to me I think we have to be careful of how we balance that out of with the progress that we need to do to solve these concrete problems that are are tackling uh that are that we need to tackle as a Society right now right and it's not it's not only the unintended consequences but there's a spiritual element as well which is I think strange to talk about at a science conference but does anyone think about that I feel like honestly I mean that's a good answer long history of people trying to you know as you say almost play God building machines or or things in the image of themselves I think of the gollum from Jewish folklore shaping a humanoid robot or a humanoid thing out of uh out of clay from the banks of the vitlava river and this is this goes back thousands of years this desire to create things in the image of ourselves and I do think there's there's something ethically concerning about and there's there's something kind of deeply spiritual about this desire um but I you know I agree with the previous panelists there that a lot of what's going on in AI is happening outside of this Focus which gets sort of blown up and telescoped this this particular goal of building something in the image of ourselves when a lot of what we're doing in AI what have that's been doing is building tools not colleagues as the philosopher Daniel dannett said and we want AI I think to complement uh skills and abilities as humans not necessarily replace them but there will be this underlying curiosity which I think is valuable to understand ourselves understand what makes us somewhat distinctive perhaps not as distinctive as we might like to think from all other animals but certainly distinctive to some extent and that for me is is that the sort of scientific value of General AI it's shedding light on what it means to be human before you go um let's uh get get mics out if people have questions raise your hand by the way I can't see the audience so you just have to ask it but but please can you go ahead and then we'll get to the first audience question quickly I think that one of the traps that we we fall into when we have you know the conversation that you've just raised is um a lack of of zooming out of understanding that that you know we have constructed ourselves as it were uh collectively uh you know our bodies are the way they are because of all of the technologies that we've invented over the last hundred thousand years uh that's true of you know fire and short guts and of uh drinking milk and adulthood and of not having hair all over our skin because we have clothes and so on um you know like the the locating uh the spiritual element or the or or intelligence or any of these kind of things in the individual alone and not thinking about that as a broader Collective thing I think is the root of a lot of cognitive dissonance and from that perspective you know when we do things like build Foundation models you know it's not really separate from uh from the the intelligence that we have collectively um and and that we experience and I think that frame is is an important one and one that we have to start to come to grips with especially in in the west where we are really obsessed with the idea of the of the kind of uh individual isolated from society okay great I could see the crowd now so uh let's run some mics out and uh we got we got one there Whoever has it on this side of the room go ahead and ask sure can you guys hear me loud and clear great I was wondering if the panel could touch on legal liability and AI um you know is it the average employee at a company making use of AI systems is it is it the Creator uh what are your thoughts on this a legal liability question not very on topic but go ahead Francesca I mean I heard about liability what was the question exactly he wanted to ask about who's who's liable legally when people use AI inside a company you know I'm gonna make an executive decision here and skip that one sorry we're going to get it later but we want to get on topic questions do we have anybody uh on this side of the room right here uh hello everyone I am Professor Avinash and I am from India I do research in the Indian spirituality and ethics of AI and I really like this panel I just want to make one small comment that we humans think that we are the super intelligent but that's not the fact animals can predict earthquake better than us and I think this is the way we are developing new technology especially the AI and what I believe that humans are the Creator created by another intelligence and now we are creating another intelligence so how I think the companies and governments are not able to identify or may not able to even initiate the discussion that what would be our relationship with new emerging technology especially the EI after 20 years 30 years so he the gentleman asked about the legal liability but I think things will move from legal to ethical liability okay so yes my only question at what companies are doing especially the corporates were creating AI to develop that relationship between Ai and humans Ai and humans yeah what they are doing to create that relationship apart from government what they're doing to what what companies are doing to create a relationship between companies and humans because governments are trying to regulate but they can't do it so what big tech companies are doing okay okay go ahead go ahead really briefly you know I think it's an interesting question to raise what do we what are the ingredients we need to put in place to create that that relationship so-called between AI agents between individuals and between society and we would need a much much longer panel to take this on let me add one ingredient that I think is important in my view and and that is you know in a lot of the work that we've done in meta AI is really to lean in heavily on open science and making some of that work available whether it's the model themselves whether it's demonstrations whether it's Pro research research prototypes that people can get their hands on to start to experience that technology and in a sense it's by making that availability to a large number of people outside of the companies that we learn how to build our technology better but also that other people can have a point of view on how to build that technology and so the the open science aspect in terms of of code models demos has been really important in our work and one that we do try to to push and and to support it's not always easy there are liability questions there are ethical questions that come up but but I think it's an important ingredient towards that okay well we're we're way out of time so maybe we could pick this up later but but for now I just want to thank Joelle Francesca Blaze and Anil this was amazing thank you so much for being here let's get around a closer [Applause] thanks so much Alex thank you panelists that was a fascinating discussion to get us caked off and thank you all for your questions as well