Debunking the great AI lie | Noam Chomsky, Gary Marcus, Jeremy Kahn

[Music] foreign oh my word there's so many people in here for a final two acts in a moment I'll introduce Noam Chomsky and two amazing people to speak alongside him and then we're going to close out as is tradition with the pretty incredible president of Portugal it's been an incredible four days I hope everybody has had have you had a good time it's been remarkable the father of modern Linguistics Noam Chomsky joins scientist author and entrepreneur Gary Marcus for a wide-ranging discussion on why the myths surrounding artificial intelligence are so dangerous and why we should not rely on AI to save us all please welcome to the stage Noam Chomsky and Gary Marcus of robust.ai in conversation with Jeremy Khan from Fortune thank you [Music] he's here oh hi gnome thanks for joining us and Gary thanks for joining us a lot of people think we're living in the Golden Age of AI there's a lot of type around generative AI these systems like Dolly and stable diffusion that can create amazing images just from text prompts there's also a lot of discussion about gpt3 a System created by openai that can create long passages of coherent text again just from a small human written prompt a lot of you are very excited about where this technology is going but no my understanding is that you are not that you're actually disappointed in where we are with this technology why is that oh he's I don't know if he can hear us there we go yeah I am indeed yes be efficient one way is it's not strong enough fails to do certain things the other way is it's too strong it does what it shouldn't do well my own interests happen to be language and cognition language specifically so take GPT it there's uh Gary Marcus others have found lots of ways in which the system's deficient this system and others doesn't do certain things that there's a the other that can in principle at least be fixed you add another trillion parameters double the number of terabytes and maybe do better when a system is too strong it's unfixable typically and that's the problem with GPT and the other systems so if you give a database to the GPT system which happens to be from an impossible language one that violates the rules of language they'll do just as well often better because the rules can be simpler for example one of the fundamental properties of the the way language works there's good reasons for it is that the rules the the core rules do not ignore linear order of words they're ignore everything that you hear they uh attend only to abstract structures that the mind creates so it's very easy to construct impossible languages which use very simple procedures involving linear order of words trouble is that's not language but jpt will do just fine with them so it's kind of as if somebody were to propose uh say a revised version of the of the periodic table which included all the elements all the possible elements and all the impossible elements and didn't make any distinction between them that wouldn't tell us anything about elements and if a system works just as well for impossible languages as for possible ones by definition not telling us anything about language and that's the way these systems work it generalizes the other systems too so the problem is the Deep problem that concerns me is too much strength I don't see any conceivable way to remedy that interesting Gary you've you've examined how some of these systems fail can you talk a little bit about what those sort of modes of failure are that you've discovered sure and usually also gnome and I have had this interchange over the last few months where he tells me that I'm too nice when I'm negative about AI in this case there's a certain way in which I thought he was too nice so he said that in the ways in which these systems are too weak just adding more parameters maybe that will help and there I actually take a darker view maybe than gnome does some examples of the kinds of problems these systems have are Dolly if you tell it to draw a blue cube on top of a Red Cube it might just give you a Red Cube on top of a blue cube so one of the most basic things about language is that we put together meanings from the orders of words this is an idea that goes back to Fraga and even even further these systems don't understand the relation between the orders of words and their underlying meanings um another version of this system you say something like draw a picture of a man chasing a woman or draw a picture of a a woman chasing a man and the system is basically a chance it really can't tell the difference and it's not clear that just adding in more of the same what people call scaling today is actually going to help I think that there's something fundamental missing from the systems which is a understanding of the world how objects work what objects it's describing it was another article that just came out on theory of mind and whether these systems can understand what you believe about other people and there was failure on those kinds of things there's failure on a system to understand something like if I want some grapes or sorry I ask you if you have something did you touch this glass I can't remember the exact example and you say well I had gloves on so my finger implying your fingerprints aren't there right and the system doesn't understand it that was another Benchmark that came out this week so when Dolly came out Sam Altman of open AI said AGI is going to be wild artificial general intelligence is going to be here and implied that the system are closed artificial intelligence and scientists like me weren't given access and in the last couple of months we've been given access and These Things Fall Apart left and right so yes they can draw pretty pictures but no they really have a very shallow understanding of language and then to gnome's point if they have a shallow understanding of language they're not really helping with the question that gnome has worked on all his career which is why is human language the way that it is and what is he's saying is these systems could learn computer languages or languages that aren't like humans they could learn anything they don't do any of it perfectly and they don't really tell much about why we're the special creatures that we are interesting uh no I know you you've sort of said that these systems are maybe good engineering but not very good science at all can you explain what you meant by that and why you think you know the science here isn't isn't really valid well let's take engineering achievements so I happen to be hard of hearing so I can't hear most of what you're saying I'm reading life transcription captions it's pretty helpful I'm all in favor of it it's achieved by a Brute Force no scientific interest but it's fine I I'm happy with it I'm happy with the a snow plow that clears the streets so you don't have to do it by hand I don't see anything I have no criticism with that I think it's great uh sometimes the engineering achievements can contribute to science so a telescope for example uh made together they allowed Galileo to discover things he couldn't have discovered without it well that's fine there are apparently some engineering achievements in with the Deep learning approaches that have been like a telescope so they've apparently produced some helpful results with regard to protein folding look massive computation Beyond and computation so that's that's good but science is a different is a different concern you're trying to understand what the world is like not how can I make something useful nothing wrong with making useful things but the project of trying to understand what the world is like is different so turning to Gary's point I'm perfectly willing to concede that I've been too nice and thinking that if you had a thousand parameters maybe you'll get somewhere maybe you won't but my other my concern is different if the system doesn't distinguish what's the actual World from non-actual world it's not telling us any just as in the case of the souped-up periodic table the egpt systems and others like them can find superficial regularities in astronomical amounts of data and produce something that looks more or less like what their data was but it can do exactly the same thing even better often with data that violates all the principles of language and cognition so they're first telling us nothing about them you might ask whether it's making an engineering contribution actually in this case I don't really see any I don't see what point there is to GPT except maybe helping some student fake an exam or something but so it doesn't seem it has no scientific contribution doesn't seem to have any engineering contribution so as far as I can see it's basically a way of wasting a lot of the energy in California right and uh inspiring yeah that's a good point yeah what what do you think the dangers are here I mean if if it's you know these are tools that are in some ways useful is there a danger though in in thinking that they're more capable than they are is there a danger in the hype Gary what do you have to say about that I think there's two kinds of dangers one kind of danger here is that I think we're following the wrong path in Ai and the right path to AI I think is going to involve looking more at humans and doing the science that gnome is talking about and so we're in this moment where we're at a local maximum we have something that looks good but isn't really as deep as we need it and it's sucking the oxygen away from the field of cognitive science looking at how humans understand language how humans understand the world because it's so much fun to play with these toys and it's fine for people to play with the toys but it takes a away a lot of the energy your young graduate students if they have a choice between going into Linguistics or a choice between working on gpt3 and getting paid a lot of money they're probably going to work on gpt3 but I think it's short-sighted that we wind up with a technology that actually has a kind of short road ahead isn't going to last as long as people think and you look at the history of AI there have been fads over and over again expert systems and support Vector machines the people in this crowd probably have never heard of that we're super popular for a while and we need to have something that's more stable and one more Fat's probably not helping us then there's another kind of danger which is the systems we have now what they do is they perpetuate past data they don't really understand the world so for example that means that they're sexist and they're racist they're not built to be that way but because they just copy the data that's there and don't have values about equality for example they perpetuate past bias another example is because they don't have models of the world they kind of lie indiscriminately now they don't do that literally because they have no intention There's No Malice but they produce misinformation and it's going to completely change the world in the next couple of years the amount of misinformation the troll Farms are going to be able to produce and the cost of that is really I think going to be devastating to the Democratic process and so the fact that you have a system that kind of looks good can make sentences that are grammatical but make up stuff like why is it good to eat socks after meditating and then say some experts believe that it's good to eat socks after meditating because such and such it all sounds plausible you read it on a website when you're like falling asleep and you believe it well now you can do this at scale you can get your own custom version of gpt3 for less than half a million dollars so if you're a troll Farm that's a great investment but it's terrible for society so there's some specific dangers around it then there are lesser dangers like a lot of people spent a lot of money on driverless cars and it doesn't really look like it's working out like maybe it'll happen in 20 years but we're not going to have routine level five self driving where you can type in a destination and go anywhere the way musk has been promising it for six or seven years that's just not going to happen there's been a hundred billion dollars put into that industry and the money is wasted I'm more concerned by the threat to the Democracy than investors wasting their money but it is another kind of side product of all of this is there a danger also in sort of not understanding modes of failure I mean some of those things you've discovered recently when you have had access these systems about the way in which they do not understand language is there a danger in in sort of believing the hype and thinking these things are more capable than they actually are well I actually have a prediction I think it comes out tomorrow in wired for the year 2023 I predict that it's going to be the first year in which there's a death attributable to one of these systems either because it tells somebody to commit suicide or somebody falls in love with one of these systems but the system doesn't actually love them back and they've added it or because it tells them to you know mix Clorox with Drano and they drink it all so people trust these things because they appear plausible but it's kind of like a magic trick what they're really doing we haven't said this is basically a version of autocomplete like in your phone you predict the next word GPT is really just autocomplete on steroids but it gives this illusion that it's more than that and if people get sucked in then they can take bad advice from these systems and I think we will see a lot of that next year as these things become widely distributed and cheap interesting no I I know you said you know you don't think there's been much contribution here do you think there's been sort of any contribution from these large language models to the actual understanding of linguistics which is you know what you've devoted your career to I can't think of one single thing I mean I can give you an example of the kind of stuff that comes out just a couple of days ago I don't call the literature on this closely I'm not like Gary but somebody sent me a friend sent me an article from something called I think the Federation of associations of behavioral and computers science or something like that it was a study a massive study in which the guy had was published in their Journal it had he had all of Reddit and a ton of other things in his database all kind of super computers and studying what you can learn about frequently of works and how important this would be for language understanding and they gave one example that was really the chief achievement that discovered that the word occasion is used much more frequently than the word molecule and the reason they explained it was because it's used in many domains whereas molecules only in a few domains who would have guessed it that kind of contribution that comes out if there's anything else I've missed it I mean these things can be useful like I say and basically what lies behind the life transcription that I'm using so again I don't have anything against snowplows happy to have them around but they shouldn't mislead people indicating it if we're Gary's reasons and others that they're making some kind of contribution to science they're not they're not helping us understand the world and they can't in principle because they're too strong irremediately interesting and no I I think you know you've said uh when we were speaking uh before this that you know looking back at the history of AI that the purpose of the field was not simply to make better snowplows that there was this idea of trying to actually understand human intelligence and then to to build on that um can you talk a little bit about you know do you feel like the whole Field's kind of gone awry here in in focusing so much on uh these deep learning systems well I I was around when AI was being developed and knew many of the people the original goal was to use the capacities that had been developed in computer science both Technical and intellectual because it was a major intellectual breakthrough the theory of computation developed by touring and others the original idea was let's use this technical intellectual contribution to see if we can understand something about how humans think you go back to the paper that initiated the field touring Ellen turing's famous paper on a 1950 paper on on cold can machines thing machine of course means program not the object uh well he was interested in seeing if we can understand what thinking is you go to early Pioneers in the field for Simon marvinsky others that was their goal as well can we learn something about cognition thinking well that's why now considered old-fashioned Ai and there's various disparity and it was gradually replaced over the years right what Gary was just describing playing with fancy quarries that is understandable I mean I've spent most of my life at Mi I'm gonna take good teenage kids and smart teenage kids hand them fancy computers it can do all sorts of things massive programming systems it's going to be a lot of fun when I got to MIT in the 50s one of the ways that students used to have fun was by playing with a very elaborate electric Railways which you could do all kind of complicated things one well 10 years later in the 60s when they started having these toys no more electric Railways this was a lot more fun you could do all sorts of things figure out a way to make the elevator run when nobody was running it but it's running a program and you can imagine what it's like so it's exciting to play with the toys you can do a lot of complicated things it does get intellectually interesting when you work into how the programming works so the statistical analysis works and so on it's just not going anywhere it may produce useful things like life transcription but that's not very exciting I mean it's helpful but not a real contradiction understanding the world as far as understanding the nature of language or any other cognitive process I literally can't think of anything that's been contributed and Gary I know you've thought that there's a lot of things we do know about cognitive science that is being ignored sort of by the current uh you know enthusiasm for deep learning and ever larger deep Learning Systems yeah I I look at what people are doing now and I think have they ever taken a class in linguistics so in linguistics you learn about the relationship between syntax and semantics for example in between those in pragmatics and if you build a language production system for example in the classic way you start with a meaning that you want to express and you translate that into words or in language comprehension you start with a sentence that you want to understand and you translate that into a meaning well gpt3 doesn't do that and people don't even notice that it doesn't do that what gpt3 does is it here's a sequence of words and it predicts the next word but let's say you want to talk to a user you know what you want to say well GPT is like a wild bucking bronco it's very powerful people it might produce something grammatically interesting but whether it winds up giving you what you want is an entirely different matter and if you want to extract from it a model of the world it doesn't really do that so there's this paper called Palm seiken I think by Google where they put gpt3 in a robot and it works like three quarters of the time and it's amazing but a quarter of the time it doesn't work now imagine that you want to put your grandpa in bed and you want to tell the robot to do that and three quarters of the time it does that and one quarter time it drops your grandpa like this is not good yeah we don't want any drop grandpas around so you know cognitive science tells us that we want for example it's just a basic thing but you want to map a meaning onto a sentence or the other way around and people are like I've got all this data I don't need to do that well no you really do still need to do that another example is we have lots of knowledge you know you go to college and read books or when you're five you read books you absorb verbal symbolic Knowledge from the world and there's this funny any politics that goes back 50 years about using neural networks versus a knowledge based approach it turns out the knowledge-based approach has some value but it had some trouble you know 30 years ago so now we're doing all this stuff without knowledge which means like you have all the knowledge in Wikipedia and nobody really knows how to put it into these systems and because they're statistical mimics you ask a question like who's the president of the United States and they might say Donald Trump because in their data set there are more examples of Donald Trump than Joe Biden where if you want to be able to reason and say well Biden is now the president Trump was the president before you want to use your understanding of sequences of events and time cognitive scientists think about this stuff they talk about things like discourse models putting together what it is that we're talking about how we're talking about it there are lots of ideas another one that we didn't talk about yet but which I learned from gnome and which is really foundational I know he learned it from Plato is innateness the idea that something is built into the mind there's lots of reason to think that when we learn language we're not starting from a blank slate gnome has made this argument for years and years I think it's still true and the failures that we're seeing for example with Dolly and I'll tell you about one another are showing you that if you start with a blank slate that just accumulates statistics you don't really understand language the paper I have yesterday which I sent to both of you guys shows the same thing in Vision so there's some models that label activities so it can say you're nodding your head those people are sitting out there you can do some visual labeling and the myth is always the right things will emerge which is kind of magic it's like a latinate word for magic it will emerge when you give enough data to these systems that they will understand the world so we built a benchmark based on what 10 month olds do in the lab or four months old for some of the experiments understanding basic physics like things drop or if they hidden you can still you know find them eventually and these systems don't understand at all the Allen AI Institute and some developmental psychologists from Illinois and I published this paper yesterday it's a comp complete failure of the empiricist hypothesis that if we just give a lot of data that cognition will emerge and developmental psychologists have done lots of work Liz felke in particular but many people Renee Byers on one of our collaborators on this paper have done lots of work showing that children seem to have probably an innate sense of how objects exist in the world like we don't start with nothing Kant said we start with time and space and causality he's probably right we probably start with those things and over and over people keep pursuing this hypothesis that ignores the cognitive science around that and says we'll just use all the data and because it works like 75 percent of the time they think they're making progress sometimes you make progress 75 percent of the way it's not enough we saw that in the driverless car industry right you know getting close doesn't really seem to solve that problem and that's what we're seeing here right and Gary I know you have some ideas about how you would go forward um you know if if today's AI despite all the hype around it is not delivering what we want you know what would what would be a path forward so I have an article called next decade in AI which people can read later in archive but it points to four things one is what we call neurosymbolic AI which is putting together this tradition of neural networks that's popular now with the old-fashioned AI trying to find some kind of synthesis between the two because they each have some value the neural networks are good at learning from data but they're not very good at abstract knowledge and the old symbolic stuff is good with abstract knowledge but not learning from data so that's step one step two is we need to have a large database of machine interpretable knowledge so all of these things kind of fake their way through but knowledge is never Rich enough and Abstract enough so like you know that if I have a bottle you can't see it but if you saw me bring it up if I knock it over there's probably going to be water on this stage right even if you haven't seen this particular bottle in this configuration a very general abstract knowledge and that's crucial I think to any real general intelligence then we can reason about things we can make inferences about them and we have cognitive models of the world so you know right now that mentioned it that this bottle is here even if you can't see it and so you're maintaining a internal representation of the stuff that's out there it might be imperfect but you have a representation in your head of the things out there in the world and you reason over these are foundational ideas in cognitive science and they're not represented in current Ai and we're seeing the cost of that interesting gnome do you think that what Gary's saying is probably the way forward some of these ideas about innateness should be incorporated into the way we think about AI systems I think those are good ways to proceed but I also think that if AI wants to proceed it should take into account what has come to be understood about the basis for acquiring our cognitive capacities which of course are do involve D fundamental built-in innate properties we know have some understanding of them and even some understanding of whether the kind that Evolution would provide uh we should recall that back in the early days the days that I happened to favor myself uh AI was basically not distinct from cognitive science it was just one of the ways of approaching the problems of cognitive science using achievements great achievements that had been made both intellectually and technologically with the development of the theory of computability and its the physical devices that were able to implement it this was a great way to advance cognitive science as long as you were part of part of science not when you're playing with toys we're trying to impress uh science reporters and journals but when you're pursuing the message of Science in the cases that Gary was discussing taking into account the kind of discoveries of the spielky grenade buyers on them and others about the innate basis for our ability to maneuver in the world if you're working on language the and thought which are virtually indistinguishable then you look at what has been discovered about the innate basis for the acquisition of language on the basis of virtually no evidence that has been shown by now been experimental work a lot of it by the late by the claytman others that would show that a child just acquires language the way it it walks you know it it's just sitting there waiting to be triggered by some stimulation then it all comes out well you want to study uh the uh introducing such understanding into the methods of AI could lead to carrying forward the earlier programs which I thought were on the right track of using the achievements of uh intellectual achievements technological achievements as a way of carrying cognitive science forward in its effort to understand what the world is like in this case what our minds are like great well we're out of time but I want to thank gnome Chomsky and Gary Marquez for coming here and and debunking a bit of the hype around today's Ai and also pointing forward to some some ways we might be able to take the technology forward thank you very much and thank you for listening

Channel: Web Summit

Views: 199,776

Keywords: web summit, web summit conference lisbon, web summit paddy, web summit youtube, web summit lisboa, Lisbon, Lisboa

Id: PBdZi_JtV4c

Length: 32min 23sec (1943 seconds)

Published: Mon Nov 14 2022