Intelligent Thinking About Artificial Intelligence

[Music] [Music] welcome everyone to this conversation on AI AR VR maybe some other twetter abbreviations who knows where it will go but it's an exploration of these ideas with jarn lanir let me just give a quick introduction I'm sure most people in the audience are quite familiar with who he is but he is the prime unifying scientist at Microsoft where he helps lead the company's development and implementation of artificial intelligence systems he's a leading computer scientist who's done pioneering work in the field of virtual real ity a bestselling author who has written extensively about technology and its impact on society he's a musician with a deep interest in unique and historical musical instruments and a composer of new classical music in 2018 he was named one of the 25 most influential people in the previous 25 years of Tech History by Wired Magazine which is pretty good he's also been named one of the 100 most influential people in the world by Time Magazine and one of the top 100 public intellectuals in the world World by foreign policy magazines so those are quite a range of credentials so welcome jiren it's so nice to see you hey Brian good to see you how's everything going oh it's it's great it's great good you know I was just thinking in preparation for this conversation about the first time we met I don't know if you remember it's about 25 years ago I came to your apartment in Manhattan which was very similar to what I'm seeing in the background in your shot right now chalk full of instruments that's right yeah you know I remember delicately walking around because there was barely any floor space that wasn't covered by an instrument yeah I've since learned to put them more up on the walls uh having a child taught me the value of having floor space so yeah no doubt no doubt and so I this won't be the focus of our conversation per se but where did that fascination with musical instrument come from I think it came from my mother uh what happened was uh she was a holocaust Survivor who kept me very close and before she was taken by the Nazis at age 13 she'd been a prodigy performer in Vienna and she really treasured uh teaching me to play the piano and uh some other instruments and then she died when I was quite young in a car accident um and somehow music became my continuing connection to her and in particular this experience of learning a new instrument and so I just keep on learning new ones and all I can say is I think it's still cheaper than heroin and as far as like ridiculous obsessions go I hope it's one of one of the less damaging ones but it does get a bit extreme but anyway that's who I am and I I won't pretend to be otherwise I'm an instrument fanatic that's how it go that is a that's a remarkable story um and and so do you have one instrument in particular that you consider your instrument or is it really you just continue to expand the repertoire well not really I mean um the piano a little I suppose which is this black thing covered right there but um I it always varies which one I'm playing more than others and uh I've been burning through all kinds of instruments from different cultures now for decades um who can make sense of it yeah you know we're gonna talk signs of course but I also recall that you and I were once both part of a music festival together in California Philip GL was having a a music festival and I was collaborating with him on right and you were collaborating with him too uh yeah Philip and I did a lot together he produced some of my records way back in the 90s when we were we were we Lads I guess and um we were still playing at least once a year together until recently and uh you know um he's getting up there in years and I don't I don't know what the future holds but I really love the guy yeah no he's quite wonderful I mean our my collaboration with him you know and World science fesal collaboration with him was really a highlight of Performing Arts things that we've done you know it also turns out and I suspect you're not aware of this that our lives intersected even before that at least in principle because when I was at IBM this is way back in the 1980s I work for a guy named John cooch I don't know if you know that name but he introduced me to Marvin Minsky and Ed fredkin at MIT and my understanding is that those guys especially Marvin had a profound influence on you right yeah Marvin was um with whom I absolutely disagree about everything Marvin was a uh astonishingly generous mentor to me uh he was my boss I was hired as a young researcher at MIT when I was I think still a teenager and really yeah like way back um and uh I uh I love Marvin and I I really do disagree with him very profoundly about almost everything but it's all right and Ed fredin of course was a buddy but not we never really work together um but Marvin yeah very important person in my life and so so Marvin I guess together with John McCarthy maybe others that you could fill me in on I don't know the history that well but that this is really where artificial intelligence began right ah yes well the history I there's some history I have a high degree of confidence in being able to relate to you because I was there but some of this is from before my time so I have to re re rely on what people tell me and they tell me different things so I have to do my best to to infer um it's one of those things like virtuality where if you want to find antecedence going way back in history you can find it with the ancient Greeks or whatever but right I think a nice starting place is AA love La the first programmer and aah um thought about it and decided AI wasn't a good way of thinking and it's just the human programmer we should we should concentrate on as the entity um and then about the human the human programmer the human programmer right that was A's where she where she landed on it and then um probably the next major thinker I think would be em Forster uh the the the famous novelist who wrote a vision of what the future internet and computer science might be like in I think it was 1907 I might have that slightly wrong but something like that first decade of the 20th century um which is sort of an astonishing Act of premonition I mean it's almost Supernatural and makes me question whether time is what it seems to be because it seems impossible that he could have written it and he also had a very strong was that like a white paper or because I'm not a Nolla is what we usually call it it's a somewhere in length between a short story and a novel and it's a it's the Prototype dystopian science fiction novel um 1984 and Brave New World and the Matrix movies and so many others are obviously following the plan laid out by the machine stops great they they all have a similarity in the characters and the overall themes and uh uh that one was a deeply humanist work that was critical of the idea of of of putting too much stock in the machine as a center or as an entity um where we really see the tables turn is with Alan Turing um oh and I should mention van ofar Bush who who uh thought about uh computers and networks and algorithms as a way to help people but it was still person Centric now Turing turns it around and proposes well maybe we should think of these things as entities but what I always ask people to do is think about the context in which Turing did that uh here's a guy who help helped win World War II who was declared one of the most important combatants against the Nazis he was a savior of all those who didn't want to be killed because of their arbitrary identity that they couldn't do anything about right and yet here he was after the war condemned precisely because of his identity that he couldn't do anything about because being gay was illegal yeah and he was forced into a bizarre quack medical treatment where he was getting female hormones and to understand why somebody would try to cure gayness with female hormones it sounds so Twisted but the metaphors at that time were pre-computer ones it was all about the steam engine and the idea was equalizing pressure so maybe if you the opposite hormone will equalize the pressure and I don't I don't know you know what yeah trying to make sense of something stupid only can go so far but anyway he was starting to develop secondary uh female secondary sexual characteristics he didn't want uh so far as we can tell he committed suicide in quite a poetic manner by uh eating a cyanide laced Apple in front of the first computer right so sort of an anti- eve or something like that um the Turing test his famous idea was written down just in this last phase of his life before his death and he proposes well really the only judge of whether their person's a person is another person and if that person can't really give you a rational basis for saying this one's a person that one isn't then call them all people or call them all nonp people but don't make the choice and I think there was something coming out of a deeply absurdest dysphoric place because that's where his life was and I think we have to understand it almost as a critique of what was happening to him rather than as an idea to take it face value right that's my don't take on it um you could say I don't have any right to interpret Turing to that degree but then who does really you know yeah sure sure all right so then after touring um the next major figure flips it again and this is where we get to Norbert weiner right so Norbert weiner uh says you know so there there's a standard model of what a computer is which tring developed um along with the astonishing uh mathematician physicist John V noyman who's almost like a person of surreal talents who's very hard to even Creed is real but there there he is so anyway uh the Turing Von noyman machine is a an abstract device that receives an input uh a strip of of values processes and then either gets stuck forever or comes to an end and gives you an answer now uh Norbert weiner looked at this and said mathematically this is very general this is great but it doesn't help us understand the real world where we see or organisms and we see uh complex systems interacting the weather and the oceans and all these things so let's reconceive this in a different way which I'm going to call the cybernetic way that was his term uh cyber comes from the ancient Greek for navigating and he said instead of thinking of a system that's just logic gates let's think of it as a tangled network of thermometers we're going to say we have measurement and feedback networks and the thing is always on it's always is connected so when as the environment around it changes what it gives back to the environment changes now strictly from a computability point of view they're equivalent but in terms of the way you can position them in their environment they're different um and weiner in terms of so that that's part of his technical Legacy and and at the time he was quite a celebrity scientist he's less well known now but in um he lipped the argument back away from tring he wrote for instance a book called the human use of human beings and it ends with this crazy thought experiment which is what if you could give every person in the world a little radio connected device that would give them some kind of experience based on what happens in some Central Computer somewhere and that Central Computer was observing how they behaved and changing what signal they got in a feedback loop to to put them in a behaviorist kind of experiment like have Skinner or or pavlof did with their animals is that imaginable turning the whole human world into a big cybernetic system and he says this is only a thought experiment as a scientist I want to tell you this is not possible yeah and of course we did it we are there his point is that if you start to think of people in computers as being equivalent you'll start to treat people like computers and you can destroy society and destroy everything um and then okay now here what I've heard from a few people and I absolutely am in no position to judge whether this is true is that in person vener was kind of arrogant or difficult I don't know but anyway apparently he rubbed a lot of people the wrong way and part of what happened with the early artificial intelligence movement was a bunch of people who were just uncomfortable or pissed off with them or something so McCarthy and Minsky and a bunch of others had a a famous meeting in 58 I believe in Dartmouth where they coined the term artificial intelligence which was initially designed as an alternate term to cybernetic yeah yeah yeah right I've heard and then they P they try to pull it back to the other side and there we are so that that's the briefest history as I understand it no it's a great history and and of course for for most of us it's been an abstract idea obviously I'm not in the field and for you know my entire professional life everybody who until recently everybody who spoke about artificial intelligence it was always something that might happen and then sort of of November of 2022 takes place and all of a sudden it feels like it's no longer as abstract as it once was with chat gbt and and so on do you consider in the science of the development of the subject is chat GPT a vital moment or is it a passing moment oh yeah you know um I like it and look it's it's some kind of weird irony of faith that I'd end up in the middle of this thing that I'd always been sort of cynical about that it just happened that way nobody planned it yeah see the way I interpret it is there's no AI there there's no entity I'm still with with aah and with Norbert Weiner and so on uh in that um and I I could continue the history beyond that it's always been bouncing back and forth Doug engelbart after Minsky pushed it back in the other direction who you know uh and and many other important figures there's a whole long game of tennis going on between the camps here but um from my perspective the right way to think about the large models like chat PT uh is as a collaboration between people you take in what a bunch of people have done right and you combine it in a new way which uh is very good at finding correlations and then what comes out is a collaboration of those people that is different and in many ways uh more useful than previous collaborations but there's still no entity there there's no AI there's just the people uh you know collaborating in this new way and when I think about it that way I find it much easier to come up with useful applications that'll really help I find it much easier to think about its role in society and just in general it makes more sense say it's just like a giant mashup from the internet basically you know all text that's out there and just combine it and recombine it in New Way some of it's from the internet some of it's from other sources you know like uh without going into details we've tried our best to bring in the highest quality sources that might not actually be online to try to give it um to bring collaborators in who are good at particular technical or informational things that might be AB to people but at any rate it still is a collaboration of human individuals at the end of the day and uh I think it's a good one I find it helpful I find it useful I don't think there's an entity there but of course I mean perceiving an entity is a matter of faith if you want to believe your plant is talking to you you can you know like I'm not going to go and judge you but sure this is similar to that like it's yeah POS I totally I totally agree with you on that Jaren I mean you know theory of mind or the intentional stance whatever word you know an evolutionist psychologist or otherwise might use benefited the species it was better to imagine that there were entities out there even if perhaps they weren't because if they were there they could eat you and better to anticipate their presence than to overlook them by not assigning them agency so I totally I totally get that but I too do not do not see an entity in there but I'm wondering and I don't mean to put you on the spot at all but the notion of what's in there is such an opaque one do you have a working way of describing what happens inside of a large language model okay um I believe I can help with this I haven't quite yet but just give me a little more time um I have uh a Layman's explanation of how our large models work uh coming out soon in a a major journal that I think people might find useful and I've been play testing a way of talking about that uh which I can I can tell you about if you want yeah I would love to if you were willing but the other thing is uh well okay I'll do it why don't I give the briefest version of how how this stuff works but then I also want to describe to you an additional layer I think we can add to it to make its role in society and the role of people in it much clearer yeah okay so uh how it works let us start very simply with the thing all of us were obsessed with 10 or 12 years ago which is can you get a program to tell a cat from a dog and so you can do statistical measurement of a frame and see if it's more blue or more red that's easy it doesn't give it give you whether it's a cat or dog yeah you can uh find uh fiducial points that are maybe points where lines intersect and things change and that might give you an outline of a face but that still doesn't help you because they're similar they have snouts and fur and everything yeah and I was doing that kind of stuff I in fact God I sold a company to Google with some some friends a quarter century ago uh that that did this fiducial point tracking um and you can get pretty far but it doesn't help you how is how the computer knows your face to log you in it's still it has its role but it doesn't tell a cat from a dog uh so what does work is what we call Deep learning and what that means is you have a grid of kind of really simplistic statistical measures along the ones we've just talked about and then there's another grid looking at that one and another one looking at that one and it turns into a sort of a highrise and then even that doesn't do it but then you have to train it you give it known cats and known dogs and whenever the particular weights which means how much you value the output of a particular grid on one of on each of these many stories it's divided into grid grids every time it seems to work you make those settings more valuable when it's less valuable you get rid of it yeah gradually you train it this is called gradient descent and I'm oversimplifying it it um but it's really interesting because the thing about gradient descent is it's antiviral meaning like online the way we choose things as we say well if this thing's popular it'll get even more popular and then more popular more popular and then the problem with that is that then you only have these viral things which are often the worst stupid things but to train a network you have to look at the overall combination so you're constantly batting down virality that's what makes gradient descent work it's kind of marvelous like we don't have the wisdom to do that with our society but we have to do it we're forced to do it in the training so then you end up at the top layer gives you cat versus dog as an output okay so that's only cats and dogs the way you do it for everything everything everything is you look at the whole damn internet and you assume adjacency is important you say if one word tends to follow another word or tends to be close to it maybe that means something yeah if some words are close to some image maybe they're descriptive of the image not absolutely certainly but they'll tend to be and then instead of knowing in advance that you have tagged cats and dogs you use adjacency to tag the whole internet approximately train the whole thing and it takes us astonishing computational resources and like a year to do one of these Cycles so every time you see GPT turning like from three to four it's one of these giant yeah and now what you have is this huge sort of uh you can you can think of it as a a virtual Forest of these towers that are like one I just described except they're not like delineated they're all sort of virtually there and they're all mhed up but here's the thing the magic that comes is you can call on more than one of them at a time to sort of blend the you can blend them so um there have been a number of critiques like from timu and from uh many others saying well it's stochastic parrots you know it's just regurgitating and that's true I mean that's accurate but the magic is the combination so if you say I want a picture of um a Geral uh writing um uh riding a stock car on the moon or some crazy thing and I want it in the style of Monae it'll it'll start with just random pixels put them through that combination of these trees and then if it gets a little closer to satisfying all of them all of them then it'll keep the result and if not it'll throw it away then it just keeps on cycling through until it gets closer and closer to s to addressing all of them at once and then you have this image come out that looks like that which is kind of crazy and great and then you can say write me uh write me a caption for this picture that sounds like a a pirat edit and then it'll do the same thing using adjacency in the text and it'll do it I think that's great I think that's that has a lot of utility yeah I'm I'm happy it's there but but here's um what I would suggest as an image to get a feeling for its limits are if you imagine one of these uh prompts that combines access to different uh Towers if you like if you can accept this sort of confusing set of metaphors perhaps um you have uh you have gerbal stock car was it Mars Monae whatever okay so you have the set of things in between them there was never a tower to recognize the combination and it sort of constructs a virtual tower for the combination in order to be the feedback to generate this new image in generative AI right the key issue to understand is that I can fill in the gaps but it typically only randomly will go a little bit above it doesn't go higher than the original Towers right so it's a collaboration but not arbitrary constructive intelligence now when I say that people will say well how do you know that people do more that's the wrong question I'm not talking about people I'm trying to say let's understand what these models can do and this is a good approximate intuitive way of understanding what they're good at and what they're not good at but even if it's yeah go ahead but even if it's the the wrong question I can't help but ask for your intuition on the relationship between what you just described really wonderfully well you know in in layman's language what's going on which in essence is these systems wind up having the capacity to give output based on recognizing like patterns of patterns of patterns of patterns as it goes up the tower how is that related if at all to what we human beings do because our intelligence is certainly in part about finding patterns and patterns and patterns and from that being able to extrapolate to something that perhaps wasn't IM initially there as part of the data set yeah I mean I think the cor the best ansers that we don't know at present because we have the this working thing the large language models it's very natural for people to look for something similar to that in the way that natural biological neurons tend to organize themselves in layers which might be a little reminiscent there are some important differences uh natural biological brains can learn things with far fewer examples yeah of course and every single day and I'm not exaggerating every single day there's a new paper published in some AI Journal somewhere that says we've cracked it we have a program that can now learn in as few number of examples as a person it never turns out to quite be true it just every single day every single day there's somebody has new unified theory uh that's a slight dig at your profession but never mind that and then uh ours they don't come out every single day in our field but have you looked at the archive I mean okay not every day not every day don't do it you're you're that's a wise decision but anyway um the thing is the thing is uh there is that difference and also we don't have to get it to work in in our computer programs this issue of having the Deep neural network which means a lot of these layers is really important uh and we don't seem biological systems don't seem to have that requirement and then the other thing is that the training methodology appears to be different rather than a sort of the global gradient uh descent thing it it's a different sort of mechanism so there there are there different is there how deep those differences are is really not well understood I'm all for doing the research it you know we had a conversation with Yan Lon um and his I'm wondering what you think about it and obviously I'm sure you know far better than I do what his perspective is but the point that was of interest to me was he was emphasizing that what these large language models lack compared to this thing inside of our heads is this ability to reason you know ability to have a model built up of the world into which you put the data and rather than just throw all the data in there and try to find patterns try to put the data within a template within a rubric that reflects how the world really works do you think that that is for instance the natural next step for where the technology will go well um so this reflects a bit of history in the community so uh from the time that Marvin and his cohorts started to have a Triumph against the cyberneticists and whatnot They promoted a style of AI That's often called symbolic AI where it's all model and it's in the early days when I was a kid it was described using uh formal logic you know the idea is that we're gonna really we're gonna Pretend We're russl and Whitehead and we're gonna like just describe the world from these fundamental things right so then the type of AI that started to work more uh this Century let's say in the last 10 years is different and it's this large model very statistical thing and so it's very natural that a lot of people are saying can we combine these and a great deal of effort goes into that and you'll see um once again without exaggeration every day there'll be in your report of some kind of a combination of a statistical and a and a logical uh so-called AI they're very different so I don't know why they're both called AI but but they are um and some of them have been to good effect uh for instance you can uh train a model on a bunch of examples that are coming from a a logic based system and this has been used for instance to create an AI to solve U geometry problems famously announced by Google recently and um so there there I think there are all kinds of cases where the combination works so far we don't really have a way of generalizing that approach it tends to be nichy but it seems worthwhile to pursue it we certainly are yeah and and so when you experience the current state of affairs in a field that you know you're there right at the beginning of of thinking about this kind of stuff or at least you know at the in the modern way of thinking about this kind of stuff were you did you have a wow moment in in in playing with any of these systems and saying I just never thought put enough data in and follow this procedure of you know fixing the weights through this descent mechanism that we would be able to do what we do or is it like yeah uh it's cool but I me I I feel a bit sad saying this like it's a little pathetic but I never really got to have the wow moment I did for virtual reality stuff and I still do but for AI you know like um the the generative prompt thing where you combine you combine things and have constraints uh transform random pixels into something that looks like yeah I was looking through wild stuff we had proposed that and talked all about how could work as early as the 80s in the mid 80s right we were talking about that process and in a sense it hasn't been surprising and I sort of wish it would be because everybody likes that sensation you know uh but uh whether or not it's surprising to me I don't think it's particularly important to anybody uh but no I think for me I was denied that sense of surprise about it and what about what about some of the fear that certainly some of your colleagues and certainly you know many commentators around the world describing how you know this could be the beginning of of the end does any of that have a grip on you or do you think it's all just fearmongering and I mean um I have to really emphasize that it's all about people it's all about humans and so the right question is is to ask could humans use this this stuff in such a way to bring about a species threatening Calamity yeah and I think the clear answer is yes um now I should say that I think that's also true of other Technologies and has been true for a while I mean the truth is that the better we get with Technologies the more responsible we have to be the less we are beholden to fate and the more we take charge and the natural corela to that is the more power we have to destroy ourselves and there there's no way out of that I mean the power um the power to support a large population means the power to transform the Earth which means the power to transform the climate which means the responsibility to take charge of the climate when we didn't before and there's there's no way out of that chain that leads to Greater responsibility so um I think the particular way the thing is framed based on the movies people grew up with like the Matrix or the Terminator movies with Skynet is not helpful generally because it tends to frame it as there's this other entity that will arise and that thing will be the threat it'll take over and ultimately the way to fix this is to frame it again and again as human responsibility the more we hypothesize that we're creating aliens who will come and invade the less we're taking responsibility for our own stuff yeah that makes a lot of sense to me but even framed in that way you know it feels like it's pretty hard but obviously we know it's not at all impossible for you know non-governmental individual actors to get a hold of a nuclear weapon not impossible by any means but it's it's kind of hard you know if someone challenged me to build a nuclear weapon I kind of know how they work on the inside but I'm not good with my hands I'd have difficulty putting one together you know but but when it comes to these systems it feels like they're they're more at our fingertips and that's what I think you know but but but all right so look let me let me describe the thing that I think we should do that would help with all of this that I didn't get to yet so this is called dignity so the idea is when you're training the big model if you think of it as a sort of a forest of towers using the metaphor it just gave you yeah let us imagine that you can leave breadcrumb in that forest and you can say this particular Source document from this individual person was disproportionately important to this particular output that we got from the system so when I say I want an image of a Geral well which Geral well I can actually say hey it's mostly this particular Geral who comes from a Geral breeder in New South Wales or I don't know whatever you know you can actually lose the anonymity and tie it to specific soures now why is that important I'll tell you a few reasons let us suppose uh you want to put guardrails on these things that they're not used terribly and believe me we do and we put great and I should make I'm not speaking for Microsoft here I'm I'm sure there are plenty of individuals at Microsoft who wouldn't accept everything I say so this is just me that's an arrangement I have with them but at any rate what I can say is that we meaning Microsoft open Ai and the broader community that does this at at a high level seriously work on um guard rails to keep it from being terrible and and that's the reason why nothing Terrible's happened so far in the first year of this stuff or year and a half whatever it's been and uh the uh that has involved so you can think of this as a triptic like a like the herous b triptic or something you have the input data you have the outputs and then there's this middle which is Hell which is this weird thing that's understand okay so right now because the parts that are intelligible are the input and the output we have guard rails on those two parts so input means trying to avoid destructive training so we try not to train on you know things to kill people or whatever but of course it's very hard to do because you can the system will infer that by combining things and then on the output we try to catch things so here's a very hypothetical example let's say we're having one of these conferences where we have uh kids pretending to be bad actors who try to Ed the system to show us where we're not doing a good as good a job and there's there's been a lot of that and so somebody says I want a cake recipe and then sorry what I didn't hear a cake recipe like a a wedding cake but I want the cake and they find some language I don't even know what it would be they want the cake to be plans to make a bomb maybe maybe this uh atomic bomb that that uh even so you could understand how to do it all right so this this cake is going to come out with this bomb plan in it and so by doing that it evades guard rails both on coming in and coming out but the thing is if you trace what source content was needed to make that output right it'll Trace back to something about how to make a bomb inevitably and so suddenly you can illuminate where the dangers are and catch them in a way that that's very hard to do if you're just trying to characterize inputs and outputs and how they might be combined in unforeseen ways but you consider that is is that is that a realist I guess the two things that come to mind to me on that art One technical are the levels of influence on an output sufficiently distinct that you can really point to one and say yeah that is that how it really happens this is still basic research in progress and so it's possible that everything I'm saying will not work out okay however um there indications that it can and um I'm my main research partner are not in the business world they're at universities I'm trying to do it in such a way that whatever we do is not in any particular camp like us versus the Google Camp or whatever uh but anyway um the uh the question is all right first of all how do you characterize what deserves a breadcrumb and it's not a one-dimensional thing it's not just influence it's influence plus Fung abilities it's at least two dimensions and it's it's kind of like you can think of it as representing a field that's going through uh the training process and uh the geometry of the field isn't something we totally know yet and then the next question is how can we characterize that field so that we get a really strong pero principle in the results so we can say that for this particular output they'll tend to be fewer than a dozen top inputs that matter even though in an in a broader sense everything yeah so the question is what does the Pito Peak look like and so far based on a few examples it looks like they'll be a nice healthy poo Peak almost all the time uh there's still a lot of math to do it's uh some of it's actually challenging I'm sorry to report so yeah I mean I mean that that that sounds like a promising Direction but I guess the other thing that comes to mind is what if it is in the mashup phase that the more difficult or dangerous or well the the thing is it still has to trace back certain kinds of content still has to trace back and this has to do with the limited ability so that's why I'm saying the system can fill in the blanks between trees but not go above them yeah now um the counterargument to that is if you combine it with the logical type of system or the formal system we're talking about maybe that thing can climb above but you know what um thus far um I'm not too concerned about that because the question isn't how well the system can perform but how well it can perform losing any Providence that about what allowed it to perform that way and and that's a different question and there I'm feeling pretty confident that there's a solution space right now you also Advocate I gather that this breadcrumb approach might be a way of sharing the Bounty at some point of actually those sources so just tell us a bit about you're thinking on well I mean a popular idea in Silicon Valley is or you know the AI world or whatever is um that we'll put a lot of people out of work but then we'll have this this uh Universal basic income and I I don't trust that solution and there's a bunch of reasons I don't one is that anytime there's sort of a single paay for a society it becomes really tempting for takeover by the worst actors yeah and the way I sometimes put this is you might start with Bolsheviks but then you get stalinists because there's this thing that's really tempting and and creeps are going to want it you know yeah um and there's a human nature element that is unavoidable in this well I think if if there were seop pods doing it the game theory would be the same I just think that's true I think it's a it's just a perverse incentive and it's really problematic uh and furthermore um the the systemic challenges of keeping something like even if has only the best people doing it the politics of it and the stability of it are probably going to be really unpleasant uh and um uh most of the science fiction that's tried to foresee a world like this where people aren't needed starting with the machine stops but then you can see a very nice example of it in The Matrix movies subdue and contain humans in some way because this idea is that people would become like horrible without being totally controlled and contained they have to be in little pods or cells or something and that intuition which has been repeated for well over a century in science fiction I believe is correct so I think there's a better way which is if you can trace who added the value to the AI system with their examples pay royalties pay them dividends give get them into a professional Society where they they share some kind of a fee for licensing or something like why do we have to end the idea of an economy just because we have this new tool that's better there's actually no reason why we can't continue the economy idea into this new era and I think that would be a healthier more distributed more dignified way to do it and do you feel confident that if that approach could be realized that you wouldn't run into problems of the sort that there are many like let's just talk about your gerbal example there are so many places where there so many places where you can imagine the primary gerbal image coming from and the the output would still be really good regardless of which of them gets used and yet if one is primary that individual who I don't know took that photograph or drew that picture will get that Bounty everybody else won't well you now how I was saying in the process of gradient descent that we use to train yeah a neural network we have to actively fight against virality in order to not have the system spin out of control uh so one way to put this is we have to avoid attractor I don't know I I don't know how technical your audience yeah you can go a little bit further yeah sure but at any rate the idea is that you you're interested in this subtle combination rather than letting any one component of the training uh Dynamic dominate right and so what we have to do is recognize that what works inside also works outside and do the same thing in all of our society level systems we have to avoid virality in the way the stuff is calculated uh I wish we were doing it online I think we'd have a much healthier and happier world right now yeah I mean again you've been very vocal along that direction too and things I've read I I I largely agree with where you know you talk about the virality you talk about the pernicious nature of of social media which I don't want to put words in mouth let you say it yourself but this idea that a system has as its goal to addict you and pull you in and trains Itself by virtue of what works on you this is you know exactly the antithesis of a system would allow you know freedom to Reign so so how does AI play into this because when the combination seems you know again it's human actors but it seems terrifying well you know the funding for the AI World from the end of the 90s until recently was all for algorithms to help addict people more yeah and um they were uh what I'm hoping and a hope is different than an actuality I've discovered much to my suin but what my hope is is that the Advent of AI will start to force some of the companies that rely on addiction to change their business model uh so um for instance are my good friends and colleagues at Google might find that just having a conversational AI answer somebody's question is more efficient for that person than a bunch of links to follow right um I find this all the time if I'm trying to find some really obscure detailed answer from uh owner's manuals for some equipment about how to get the things to connect together and there's different versions and different you know whatever um I can use the correlations in our lji model to just get the answer instead of searching and searching and going through in fact some I you know it's better it's just better sure so if you can do that you have if if your model is so-called advertising which I don't like I think it it it um if it's possible it denigrates advertising as a business it's something different it's paid influence and channeling and amplification it's something darker than advertising has been and advertising was never entirely on this side of the Angels but this is worse anyway yeah if you have like here's here's my really useful output from the language model from the large language model Ai and now here's 20 links you can follow if you want nobody's gonna follow those those links you know so they're going to be forced to some kind of different model either people are going to pay to use the AI or something or the AI will be corrupted and will start to tell the people well before you fix your water heater you really need to do something about your breath and here's our new whatever okay something H and I I I think at the end of that system of things Google's forced into a different business model that does less damage to society I hope I hope yeah yeah yeah totally so just to Circle back for a second by I'm not speaking as Microsoft I know I totally hear you I just also just to I've sold a company to Google their buddies I'm not I'm saying this with affection and camaraderie not I I don't mean to just dump on them really no that's that's totally how I'm taking it I trust our audience will as well but I just want to Circle back you know we started talking you know about all the instruments behind you and and your deep connection to music have you used any of these AI I'm sure you have to explore composition and I mean do you find there there's utility there or do you find there's creativity there I for me no for me no and the reason why is I'm actually going totally in the opposite direction to me musical instruments are the best user interfaces that have ever been invented in some ways if if what you think a technology is for is to help a person affect the world with ever greater Acuity than musical instruments are the most Advanced Technologies that have ever existed and so what I'm trying to do is make computers more like them not the other direction seems absurd to me I just want to make sure I understand I think I do I mean you know obviously we've all had those Transcendent moments where you're experiencing some spectacular performance I mean is that the kind of influence in the world that that you're making I think more about the act of playing an instrument do you play anything I don't remember you know not well a little bit you know actually I just started playing piano again so in my Advanced age I'm trying to see if I can't do it but yeah not really good good my my warmest wishes to your neurons thank you I appreciate that when you play an instrument you you start to get this connection where you're getting a lot of intent and data out there and it's not just a question of the volume of data but it's it's the focus and the Acuity you know it's it can be really remarkable the control a violinist has over the string and the bow is by some measures close to Quantum limit on occasion you know it's like it's this very intense thing and so what I want is for computers to become more like that I want computers to be more expressive machines that people can connect to with their whole bodies with their whole nervous systems with their whole cognition with ever more subtlety and ever more Acuity so to me the instruments have so much more to teach the computers in the other direction that the idea of using the current software we have which I it's very important for us to not think too much of ourselves like I I always think of that you know in the late 19th century people were say oh physics is done there's nothing more to explore this they're wrong you know like we shouldn't think oh we have it computers are done AI is done you know that's ridiculous like um we should be humble and in particular we shouldn't think that the Technologies of past centuries that aren't computers are somehow just inferior and ready to be disposed of and replaced in some ways they might be superior yeah they might have a lot to teach us so the other direction to me just seems so much more important is so much has so much more potential and interest like it's like it would just seem nutty to me to you know I I I like I'm not for people if somebody I'm not judging anyone else if someone else finds it meaningful go but yeah totally yeah but along those lines in terms of the the the interface I mean again as we said at the outset and I'm sure the audience knows you're kind of the originator of VR right I mean it kind of begins with you and your company well just to clarify the the very idea of of putting up head trck 3D that goes to Ivan southernland who also invented 3D Graphics in the first place what I did is I did the first social ones with avatars I did the first commercial ones I did the first ones that were head supported and where there were sort of goggles that look like what we currently think of so and I made up the term virality and blah blah blah you know I did a lot of stuff so we should clarify but sure in some sense perhaps yeah but profound connection to it in any event and you know just recently I don't know week or two ago I mean Apple comes out with you know its its version yeah have you have I'm sure you've tried it probably long before I yeah so I mean there's a long history to that oh my god um 40 years ago uh when the Mac was shipped some of the key people from the Mac including Andy herzfeld who wrote the first Mac OS came over to help us with with the first VR OS and to ship the first VR system yeah and the first commercial headset was called an iPhone but spilled eye and that's good uh there's I don't know exactly what the story but I've been told by a number of people that that directly led to the iPhone with an why because the original IDE the original idea from jobs and others in apple when it was a very small nice and thing when the the Macintosh hadn't even shipped yet uh was that someday Apple would sell a headset and we all estimated it at uh 2010 oh that that you you thought that that's when back in 1982 or three or something we estimated 2010 well it's actually pretty good terrible not terrible yeah so um uh but yeah so there's always been a certain contingent that continued with an apple that wanted Apple to eventually get into headsets and so um and we've known they've been working on this specific thing for a long time and um I I don't want to get into the details but I I had some access to earlier versions of it and whatnot and uh um I'm glad they're doing it you know I think it has a wayte to go but I'm I'm happy it's there but you don't seem particularly excited about am I reading it wrong or I'm cursed I've seen all you know like this is I I would like I would sometimes I really do wish I had a different past so I could be more startled by some of these things because you know I see people using VR devices these days or I see them using chat GPT and being so surprised and typically delighted occasionally horrified but um I I'm sort of jealous of those who can enjoy that you know surprise um but I'm wondering if about that because every time for instance I teach quantum mechanics or special relativity it's as if I've never seen this subject before I find it so amazing that we were able to figure this stuff out and how it works what it tells us don't do you not have that I have that for the underlying science I totally have that um the the particular product experiences are a little different um this might be oh I don't know let's say one of these Fusion startup starts to work and there's like a fusion energy source and U there's a slight chance that you would be a little less startled by it because you would have read the paper you'd understand it right yeah that's oh yeah okay there it is they they describe the process it makes sense here it is here's what yeah I like that analogy you know what I mean and so that doesn't mean that you're any less interested in that you don't find the underlying physics delightful and I do very much I mean I I'm still totally interested in the uh the physiology the the eyes the cognition um the optical challenge of doing a proper Optical PA through is still not really solved and really really interesting and right on the edge of what we understand about light and and how to manipulate it with materials it's great I still get great joy from it no no I heard you once talk about a virtual reality general relativity yeah thing which I I've never seen I've never experienced what was that oh gosh if been a zillion of them I use this as an example of something about VR which is um uh that it's really harder to maintain apps in VR over the years than it is in other formats so I first encountered somebody doing a relativity experiential teaching thing in VR around 1992 I don't remember it very well and I believe they were from somewhere state New York probably Cornell but I'm not positive and I might have that s it's just so long I just but at any rate you would go in and you'd be in this thing where things are warped and if you tried to move your body it would kind of distort and um there have been some some that demonstrate special some that generate General yeah they've been around for but the thing is every single one of them only lasted a year right right because the technology on yeah yeah so like let's say um I say this with love the first generation Apple headset has a rather narrow field of view um Hollow was worse so I'm not saying there any but they're trying to do something a little harder than what other people have done so they have this thing as soon as you make it broader you can't just use the same apps anymore everything's going to have to be redone because it's so fundamental sure and so when somebody let's say somebody says okay we're gonna do a great relativity teacher thing for this apple headset a year later you're going to have graduated school they're going to be in different things it's not going to get maintained it'll all way no I hear you I hear you because we made I mean it turns out that we World science we you know we work with Verizon and we made these virtual R experiences one of which is what's it like to move near the speed of light we have okay great so you made is it still use by anybody but the but the people who we work with to develop the underlying programming constantly need funding to keep the experience able to run on the Next Generation this that or the other thing it's like yeah I know exactly what you mean um so the solution to that in my opinion is generative AI because one of the um best successes of generative AI has been encoding right so we often see you I've seen different measurements but the lowest General Improvement in productivity I've seen claimed from somebody who's done a serious study is 40% better but then there are lot there are many that's say no no no it's much higher than that I I don't but the point is it's non-trivial like programmers are more efficient and one of the great things and and one of the things I've been I've been working with my research interns on for the last few years and it's been in Crazy successful is prompt based virtual world creation that's spontaneous while you're in it so you go around I'd like a relativity simulator and furthermore I have ADHD and I'm color blind or whatever that's yeah sure you know and um I'd like it I'd like it to be this and that and it should just come up like you shouldn't have to go through this whole process of developers you now I don't know Apple might not like that because they like to be more control so it might not be on their ecosystem I don't know that'll be up to them but the point is that that's the way around this problem and and and I'm kind of you just build it fresh every time you need it as yeah you don't even bother saving it because it's not it becomes a a n it's just this thing you ask for there it is and then you ask for it again and it might come up or maybe you can save it if you like a particular this really does feel like the Hol de on right I mean you know you go in there and you just create a world with a few prompts or something well sure and I mean uh um not but to find a point on it but back in the day my my group and the people who made The Star Trek were talking and that's not that's not entirely a coincidence at all ah there was always kind kind of the vision so yeah wow that that that's amazing but do you see let's just take VR as a case in point do you see VR as going Beyond cool experiences and actually shaping our understanding of key ideas and issues and philosophical conundra that we have tussled with throughout the ages and now by virtue of creating these worlds and experiencing things differently we could shed new light on these old problems well okay so my experience of VR is that it shines a two extremes one is Extreme utility which is to and and there it's been a great industrial technology for decades it's been universally used to improve designs of planes and Automobiles and ships and spacecraft it's almost Universal it's been used for surgical procedure design and training and real-time assist it's been used for City Planning it's all the so all that stuff and that's a whole separate ecosystem that people don't even see it's different and that is that is I mean I'm less aware that like city planners have have made use yeah it's uh I the city the vehicle prototyping I would say is Universal City Planning is very widespread but not Universal yeah um and uh yeah and and so there's an ecosystem that supports those people um it's rapidly in the hardware of it is rapidly merging with the consumer devices as as always happens however it has its own world of software and distribution its own culture its own people it's a whole it's like another thing yeah and it's a set of different niches uh people designing uh chemicals both both uh both uh uh drug uh for drugs and for all sorts of other things for industrial processes organic and not inorganic and then um there's uh anyway so there's that world and that world's great I love that World um but then uh there's the Other Extreme which is the crazy art stuff and that you know it's very subject to interpretation because what happens is you might create a virtual world experience that at least for me seems to teach a very specific lesson about philosophy and thought but for somebody else it might teach a different lesson you know like uh there's no way to perceive without your pre-existing philosophy coloring what you perceive it's it's just not possible so for me when when I'm in a virtual world and I make everything really weird I you can change your body into animal bodies which is an amazing thing about cognition and you can uh you can you can change the perception of the flow of time and many many other crazy things but the more crazy it gets the more you notice that this thing that there's this thing in the middle which I call the little nub sometimes it's like your Consciousness is there so it's a Consciousness noticing machine and I'm an Unapologetic dualist I think Consciousness is a thing apart I think experience is a thing apart really that's a whole other conversation that we should have at some point but yeah and so and I love fighting with people about it I used to make a living fighting with people like Daniel denn oh actually his his books right here proof sure yeah I I I I made I would argue with him and other people about it for money and it's great I love it but anyway so I'm I'm a dieh hard dualist I believe in experience and Consciousness as a separate thing I think we need it in high-tech times in order to identify who the beneficiary of the technology is I think if we blend people into the background we can't identify who we're do this for anymore so there's a new pragmatic reason to care about Consciousness but whatever so that's that's what I get out of it yeah I totally recognize someone else can go to it with a different philosophy and get a different message no totally even as a non-dist myself you know one of the things that has always struck me as interesting you know this framing of Thomas Nagel in terms of you know how do you understand whether something has an inner world is there something that it's like to be that entity and of course used the Bat as his primary example what is it like to be a bat now if you can go into a virtual world and really be a bat and really experience what it is to be a bat perhaps your Consciousness can shift in a way and get closer to the consciousness of a bat you know just as a concrete example and I think philosophically those kinds of experiences could have an unusual impact to really shift discussion this is my very favorite stuff about reality really like this is for me the very best way to use it the very coolest thing I love this yeah yeah so uh one hopes that that's where where this uh will all go at some point well J we're out of time it's been a fascinating conversation one that if you're willing I'd love to have part two at some point because I feel like there's so much more to talk about but thanks so much uh thanks so much for joining us and uh look best of luck in pushing the frontiers of all these fields going forward and it's really really exciting thank you cool cool cool really really nice to talk to you great to see you all right everybody thank you for joining us that's our conversation for today again sign up for World Science Festival newsletter or join our YouTube channel to be alerted when these various conversations and programs are brought forward when they're posted and also you should know that we are going to have a nice live event in New York City toward the end of May here in 2024 so if you find yourself in New York around then you should certain join our live programming thanks so much for joining us I'm Brian Green from New York World Science Festival see you [Music] soon [Music]

Channel: World Science Festival

Views: 57,013

Keywords: Jaron Lanier, Brian Greene, Virtual Reality, Yann LeCun, Alan Turing, Atari, VR goggles, Apple Vision Pro, Microsoft, computer programming, technology, creativity, consciousness, mental models, simulation, man vs machine, supervised learning, artificial general intelligence, research, superintelligence Big Ideas Series, World Science Festival, New York City

Id: caepEUi2IZ4

Length: 66min 14sec (3974 seconds)

Published: Fri Feb 09 2024