A(G)I: Where we have been, and where we need to go - Gary Marcus - Scottish AI Summit 2023

and I'm really excited for this one um because let's just say Gary isn't exactly shy sharing his views exactly sure um Gary Marquez will be joining us virtually from Vancouver Canada so we'll have him joining us up on the screen shortly his keynote is titled chat GPT godsend or disaster I love it I love how binary is it's excellent and he'll be taking questions afterwards so please please do submit any voting questions you've got via the online platform and I'll ask them and obviously Gary pretty much knows everything so you can ask him anything and see um now Gary's a leading voice in artificial intelligence he's a scientist best-selling author and serial entrepreneur founder of robust Ai and geometric AI which is acquired by Uber he's well known for his challenges to contemporary AI anticipating many of the current limitations decades in advance a bit of a psychic and for his research in human language development and cognitive neuroscience and amateur's professor of Psychology and Neuroscience at NYU he's written five books including the algebraic mind Cluj the birth of the mind and the New York Times bestseller guitar zero he's often contributed to the New Yorker and wired his most recent book with Ernest Davis rebooting AI is one of Forbes seven must read books in AI and without any further ado please let's all welcome Professor Gary Marcus thank you very much can you all hear me we can all hear you Gary can you hear me excellent I can hear you just fine um I don't have a title slide that matches what you promised but you can ask me questions if my views are not clear enough in the godsend uh Department um I'm gonna give today a little bit of an idiosyncratic first person narrative on how I see the history of AI of course I wasn't there at the very beginning but I've been there for a lot of it so here's a very brief history of AI it's given a little bit tongue-in-cheek um but here's how it goes in the beginning there was Hubris uh within a generation the problem of artificial intelligence will be substantially solved said Marvin Minsky in 1967 after the field had been running for about a decade and all that hubris was good because it raised a ton of money but not every promise was kept by 2012 35 years after Minsky's famous prediction the problem of artificial intelligence most certainly had not been substantially solved and then so the legend goes there were gpus Graphics processing units and the graphics processing units were good because they allowed so The Story Goes a Legend um sorry so The Story Goes they allowed for a revolution and I think everybody knows the revolution that I'm talking about the Deep learning Revolution and in November of 2012 John Markov wrote about that revolution in the front page of the New York Times and part of what he said I think is true and part of it I think we should pay careful attention to the wording he said there's been a number of stunning new results with these new learning methods quoting Jan lacun who later went on to be my Nemesis and vice versa I suppose um is it the kind of jump that we're seeing in the accuracy of these systems is very rare indeed and I think that's true and then he went on to say that these advances have led to widespread enthusiasm among researchers who designed software to perform human activities like seeing listening and thinking and I think all of that held up and then I think the key word word in the next sentence is the promise he said Markov said they offer the promise of machines that converse with humans and perform tasks like driving cars and working in factories raising the Specter excuse me of automated robots that could replace human workers well I would say 11 years later we still don't have machines that can drive cars they still aren't replacing human workers all that much and we still actually have a long way to go despite all the enthusiasm in any case Markov wrote this in November 2012 and two days later I replied in the New Yorker and I think that also helped set a lot of what we've been seeing ever since so I wrote a piece called this deep learning a revolution in artificial intelligence and I said realistically deep learning is only part of the larger challenge of building intelligent machines such techniques lack ways of representing causal relationships such as between diseases and their symptoms and are likely to face challenges in acquiring abstract ideas like sibling or identical to I said they have no obvious ways of Performing logical instances and they are still a long way from integrating abstract knowledge such as information about what objects are for and how they're used and then I went on to say to paraphrase an old Parable Hinton has built a better ladder but a better ladder doesn't necessarily get you to the Moon and at that moment in November of 2012 I would say that the basic tension of the next decade continuing today uh was established in a decade passed in March 2022 I wrote an essay that really ruffled a lot of feathers it was called Deep learning is hitting a wall March 10th just over a year ago of 2022 and I said even though people are very excited about deep learning these systems are still particularly problematic when it comes to outliers that differ substantially from things that they're trained on I gave an example of a Tesla maybe running into a jet or something like that um I said for all its fluency gpt3 can either integrate information from basic web searches nor reason about the most basic of everyday phenomena I said the current deep Learning Systems frequently succumb to stupid errors and I said that gpt3 was prone to producing toxic language and promulgating misinformation all of this made people really really mad and then Dali came along and not only were they mad but they gloated they said Dali is here it can generate images from text and Sam Altman the CEO of openai said AGI is going to be wild and Nando de Freitas from uh deep Minds the high level executive there said his opinion AI is over it's all about scale now we'll just make these systems bigger all this stuff that Marcus worried about we don't need to worry about that anymore and so people were writing really high in April of 2022 and in fact skepticism of me became a meme um not not just skepticism but ridiculed so people were skeptical about the skeptic the best one I think was uh Yasha box a picture of a robot leaping over a wall making fun of me it was pretty funny even I laughed um Sam Altman said give me the confidence of a mediocre deep learning skeptic which I think was unnecessarily unkind and it doesn't hold up very well over time Greg Altman Drew with Dali deep learning is hitting a wall but if you read a careful hopefully it said deep lefting which sort of seems like hitting a wall to me and John McCune said not only is AI not hitting a wall the cars with AI powered driving assistance aren't hitting walls or anything else either which wasn't actually true but sure seemed like a nice comeback to me later in that epical year of 2022 even more progress was made so jokes and sentience and compositionality were all said to have been solved at Google but one must always look at the fine print as a scientist and say can I look at the data and so far Google hasn't let us look at the data for any of these things so buyer beware the media meanwhile expressed considerable skepticism the New York County I mean sorry enthusiasm um this is jet lag speaking these speech errors um the media expressed considerable enthusiasm Uh Kevin Roos in the New York Times said we need to talk about how good AI is getting and he talked about the Brilliance and weirdness of chat DBT there was a lot of enthusiasm in the media and then suddenly and unexpectedly even to me there was a plot twist by the end of 2022 The Narrative actually began to change a little bit the first thing that I noticed is that promises around driverless cars started being scaled back there was a great article by Max chafkin in Business Week where he noted that even after a hundred billion dollars have been investing self-driven driving cars still seem to be going nowhere Apple delayed its long room self-driving car and Elon Musk were actually tassels and Tesla's lawyers said that although self-driving technology might be a failure that hadn't materialized it wasn't a fraud well that's a really big walk back from hey we're going to deliver this next year which is what musk had said for the previous ten or previous seven anyway second thing that happened is people began to worry publicly about large language models some of this really began with Meta Meta had a system called Galactica that they released in November of 2022 and it made up really fluent for example you could have it write a story about the benefits of eating crushed glass and it would proceed to say that it had conducted studies and it was examining whether the benefits of fresh glass were coming from the phosphorus content or some other source they look like very good science and it was totally garbage and a lot of people got upset I wrote a essay called a few words about and a couple days later the thing was taken down I also expressed worry in the Scientific American that although these platforms were easy to use they were also potentially dangerous in part because they could create an avalanche of misinformation and then this is the thing that surprised me the most the Skeptics Skeptics started to change their tune so Jan McCune started saying most of today's AI approaches will never lead to True intelligence which was the whole point of the deep learning is hitting a wall essay and then he seemed to borrow a page even more directly from my book he said you have to take a step back and say okay we built this ladder but we want to go to the Moon there's no way the latter is going to get us there foreign Sam Altman who had been so critical of me in December of 2022 said Chachi BT is incredibly limited but good enough at some things to create a misleading impression of greatness it's a mistake he said to be relying on it for anything important now it's a preview of progress we need to have lots of work we have lots of work to do excuse me on robustness and truthfulness he's basically saying everything that I said in the Deep learning uh is hitting a wall essay so he went from ridiculing me to basically imitating me that um I thought I had another slide here I guess I don't um so this all of this raises a question um if deep learning is a better ladder and a letter better ladder doesn't necessarily get you the moon what should we do about it well one suggestion I'd like to make is that it takes a village to raise an AI and that this is not a solution or sorry a problem to be solved by one person alone so what I've been doing for the last few years is trying to put on debates with Vince Boucher where we get a lot of different voices in December of 2022 for example we had everybody from gnome Chomsky to Jurgen Smith Uber Kai Fu Lee Francesca Rossi uh agent Choi Ben gertzel and so forth it was really quite a spectacular cast and we asked people five questions we said can we turn to the cognitive neurosciences for inspiration how can we make progress in common sense reasoning how should we structure and develop our AI systems how can we build AI systems that reflect human values and what should we do morally and legally to ensure a bright future all of this is at agi.agidate.com um it was a three and a half hour debate many people were in Europe and stayed up until the wee hours because they were so interested I highly recommend it and that will allow you to go beyond what I can say today my own suggestion in that debate was that we should look towards human cognition I don't think machines should slavishly replicate human cognition airplanes don't fly exactly like birds but the Wright brothers learn something from Birds um we might learn something from Human cognition and here are four suggestions the first is that humans are really good at abstraction we can learn things like how to sort something or how to count something and we can do it indefinitely and infinitely whereas chat CPT can kind of fake a lot of that stuff can look really impressive but often the details are wrong so there's definitely a dream right now of having computer programs write code directly from Human text the reality is most of those systems only work some of the time and we really need human debuggers um human coders to figure out what's wrong with them here's a great example of what we should get say write 10 sentences about baseball that's an American sport for those of you who know Cricket it's it's kind of the American version of that right 10 sentences about baseball and then print the sentences in sorted order from shortest to longest in terms of number of words in each sentence in parentheses after each sentence State the number of words it contains and the system seems to be off to a great start if you squint your eyes and don't look very carefully it says baseball is a popular sport three words it's played with a baton ball six words team at bat tries to score runs by hitting the ball blah blah blah the only thing is if you actually count the number of words it's wrong baseball is a popular sport is not three words similarly when you say sorted from shortest to longest it seems like it's off to a great store start it says three words six words Seven Ten it's counting up twenty then it's back down to 16 and 16 and 12 and 12 and then up to 25. doesn't really understand what sorting means it can get little bits and pieces of it by matching bits of text but it doesn't really understand the abstraction a second example is reasoning so if I say suppose a container holds eight pennies and if I start with six pennies and then someone gives me five more will they all fit inside of X any ordinary human adult or even you know child who's not four years old um is able to solve this no problem but these systems are very um we sometimes call them stochastic or random or just unpredictable on these things sometimes they get it right sometimes they get it wrong Francois Chalet um had a nice way of putting it he said so far all the evidence that large language models can perform few shot reasonable novel problems seems to boil down to llm store patterns they can reply reapply to new inputs which is to say that it worked for problems that follow a structure that the model has seen before but doesn't work on new problems a third thing that's incredibly Salient to linguists and has been Central to the study of linguistics for the last 100 years is that we understand compositionality we understand that larger holes are made of smaller parts so if I say the sentence of red basketball with flowers on it in front of a blue one with a similar pattern you're not just looking for red things and basketballs and flowers you have structure from the whole that defines what your expectation is this system this is Dolly isn't able to do that reliably it's getting lots of the bits and pieces but not assembling them correctly into a hole so for example in many of the images the blue one doesn't have the same pattern as the red one where the flowers are on the blue one but instead of the red one and so forth and so on in a more systematic study in October in archive Evelina levada um Elliott Murphy and I looked at a whole bunch of things that you would find in the introduction to Linguistics class things like negation and coordination putting different sentences together and these systems were basically a chance they're amazing they can grab something from every word there but they still don't really understand how all those words fit together and the fourth thing the one that I'm most concerned about in terms of its political implications for the world um social implications is factuality large language models are about predicting words and sentences I sometimes call them auto-complete on steroids um they're not fundamentally about facts and this gets them into trouble here's an example Galactica made up an essay that went like this on March 18th of 2018 Tesla Inc CEO Elon Musk was involved in a fader mode of fatal motor vehicle collision in which he was a passenger in a Tesla um the drive the other vehicle was unharmed so it's telling us that he died um but we all know that Elon Musk didn't die there's nothing in the training set that supports this some of the systems sometimes things that are untrue because they borrow them from Reddit or something like that but this is not a case of that it's not a case of the system simply mimicking something that's in the database there's lots of things in the training set that go against it why does it the reason it does that is because really it's learning about relations between clusters of words so the cluster of words on March 2018 could be followed by a lot of things Tesla is is one plausible one the word Tesla could be followed by a lot of things Inc is one of them Tesla Inc could be followed by CEO Tesla Inc CEO could be followed by Elon Musk and then it's abstracting in a certain sense from these categories abstraction is not quite the right where it's clustering it's putting Elon Musk together with something like other white males from California who own Teslas and there was another white male from California who died in a fatal Tesla accident in 2018. but he doesn't know which one's which it's not actually keeping track of facts well this is a serious problem because it means you can use these things to make up misinformation or they can do that all by themselves people have been talking a lot about chatbots for search I don't even think I need this slide anymore when I first wrote it it wasn't widely recognized how bad these systems are how intermittent they are they work some of the time but not all the time they frequently make weird and bizarre mistakes probably everybody in the room has experienced or read about that um at this point so we need to learn a lot from Human cognition I think and here's an overall summary I think the covid should be a wake-up call for AI it's motivation for us to stop building AI for ad Tech and things like that to start making AI that can really make a difference but to build deeper AI the kind of AI that can for example read a medical literature we need to do more than just learn the frequencies of words and sentences um if we want to build AI that can help us with medicine that can help us address climate change power robots to take on some of the risks that human health care workers were taking and so forth we need to do something different we don't need to just fool people into thinking that a program is intelligent by predicting text we need to build programs that actually are intelligent so to get to a deeper AI that can operate in trustworthy ways even in novel environments we need to start working towards Building Systems with deep understanding and not just deep learning and I thank you very much Gary Marcus Professor Gary Marcus everyone hi Gary can you hear me I can hear you just fine how is it over there in Vancouver it's a lovely day here it's no longer raining the way it was from recent months that's different to Scotland then it's always raining here um thanks so much for that extraordinary insightful brilliant thanks so much um got a few questions for you now Gary if you can stick around with us for a few minutes please please I finished early so you could ask million okay so um the first question is um which I'm not sure you necessarily addressed is how fast are we moving right now I mean you've obviously been in this at the Forefront of the eye industry for many years now decades do you think something very interesting is happening right now and by right now I mean the last like two three months how quickly are we moving we're sure something very interesting is happening um there are some good things about it and some bad things about it so um in my opinion we're moving a little bit too fast um it's certainly seductive to play with these machines I think to some people it feels like they have a new friend or something like that um but we don't really understand the risks there was a report yesterday by europol um over on your side of the Atlantic um they're pointing to a lot of problems one was misinformation that I've pointed to a lot another was the ways of which Bad actors could use these things for fishing people's credentials in other ways in which they could take the fishing for credentials and turn that around for example into helping terrorists do various things there are a lot of potentially negative applications the one that I worry about most is producing misinformation at scale and swaying elections which I think could be devastating for democracy so we have all these negative cases on the one hand and there are a lot of positive cases but my sense is that people are moving really fast up until um a few months ago you know Google had Technologies like what we're seeing now and they thought they're too unstable they're not reliable enough let's not release them till we fix them and suddenly lose immense pressure to release them even though they are obviously problematic and we get into this question about are the net benefits better than the net harms we don't actually know the answer the only intellectually honest answer is to say we don't quite know the trade-offs here maybe we should slow down so we have a little bit better handle on it so yes something is absolutely happening whether it's the right thing or the best thing I don't think we have a handle on that yet do we think that this is separate how other Technologies are rolled out across the tech sector I mean traditionally as an industry the tech you know tech companies are fast to release stuff without the proper testing or without knowing the implications right we'll just get it out of the door well I mean this is what the tech industry does and I'm not sure that it's you know interest of society as a whole I think that with social media things we're we're kind of kicked out the door and we had problems we had influence on elections and and depression and and so forth um I hope we won't repeat the mistakes of the past but then even so this is much faster than before and the the potential impact I think is even larger than it was with social media um in part because it interacts with social media um because of the tendency misinformation and in part because the sheer scale is even much faster than it was um for social media so we 100 million people adopted chat GPT in its first few weeks we've never seen anything like that before and it's just clear every day a new kind of negative use cases popping up it's just clear that we don't have a a good handle on it relative to the speed at which it's being pushed out how worried are you and should we be worried uh in general I didn't know if you're gonna finish this out um yeah how worried am I I'm pretty worried well I mean you you personally Gary because obviously you're you're an expert here um and you can kind of the way I look at things when I'm trying to gauge my concerns uh in life is basically to compare it to what's happened over history and over time so where are you at right now I'm pretty worried and the fears come from a kind of Confluence like a perfect storm of factors one is the speed at which this stuff has been released another is the kind of instability of it the unreliability of these systems for example Bard was just released and um there was a paragraph about my book rebooting Ai and three sentences had five errors and that you know that's kind of disconcerting the unreliability of it and then the third thing that's unreal or disconcerting rather is the degree to which the big tech companies seem to have abandoned the principles that they had talked about I noticed words behind you trustworthy ethical inclusive and so forth um those words were were repeated by these companies and suddenly they seem to be of of less interest so you have the scale of it the kind of unknown periphery of how many weird use cases there are potentially dangerous use cases and the fact that the leadership that we didn't think we were getting from the corporates just suddenly it doesn't seem to be there anymore there's going to be a big announcement tomorrow I won't quite give it away but it's already been leaked a bit you know there are a bunch of us that are concerned and are going to speak up thanks guys now there's a few questions that have come in from the audience so I'm just going to ask you those um once comment that says in this era of AI aided writing and research in this era of AI 80 directing and research there's a big concern for some people in academics around the issue of copyright and plagiarism what are the things being considered to address these concerns I think they're they're difficult so um on the copyright issue it remains to be seen whether existing laws can really protect for example artists and so forth um we may need new laws I think Society has to decide what we think is adequate compensation for an artist whose work is is swallowed up by these systems um there's some technical questions there about the degree to which they are copying things that exist degree to which they're really going Beyond what's there and and there are social questions we have to decide what we think is fair on the plagiarism issue I mean it's kind of similar these systems are very good on the written side of kind of substituting with synonyms things that are already there they're not as good at being genuinely original um it's already you know extremely attractive to students to use these things and I don't think that we can actually stop that if I were a professor what I would do is to say go ahead use chat CPT or one of these systems and then write an essay about what it can it's wrong try to learn about the strengths and weaknesses of these systems try to explain how you would write something more interesting because most of what they write is not very interesting what would be your own unique angle on something so I would try to teach all turn all of this into a teachable moment as best I could thanks Connie now you're in Vancouver we're in Scotland here in Glasgow um what do you think about Scotland on the global AI stage do you think we're there do you think we've got the potential what do you think about us as a nation when it comes to AI I mean I I'm not sure I'm completely knowledgeable there there's long been a very strong condition in um Edinburgh and and Glasgow um in the cognitive Sciences um that I think is very relevant to AI research um you know I think I I know Edinburgh a little bit better but for years there's been really good connections there between people working on AI and working on the human mind and I think that's important and I think that's one of the things that that you guys can try to leverage um to become an important player and if I could take a half a step back um most of the field right now thinks that large language models are the answer to AI my own view is that they're kind of a detour in fact John lacun who historically was my Nemesis has been lately saying that um these things are large language models are kind of off-ramp um in in the journey to AI so the question is what do you need and what I'm arguing is you need a firm basis in human cognition to solve a lot of problem that haven't actually been solved yet and I think you know that's a good place for Scotland to get into the mix thanks so much another question from the audience here is um so they loved your breakdown of the aspects of human cognition currently I can't mimic so to what extent do you believe cognition needs to be embodied either as a humanoid or a different type of physical body to gain these capabilities I don't know if it's absolutely required but I think it's certainly helpful if you look at human children there are some that grow up with disabilities so they can't really move around the world and they still learn a lot um from the other hand human children once they can move around start learning things really fast um my guess is that when we can make something of embodied cognition for robots that will be really helpful but that we're starting with systems that are blank slates that don't really know how to assimilate the experience that they're having so they don't make much of it but eventually I think we'll be able to do better with that thanks Gary um another question that's in here can you explain how GPT is lying or is it a misnomer isn't it just reporting uh is it just let let me start again with that question can you explain how is chat GPT lying or is it a misnomer it's just reporting false information right isn't it it's not literally lying it has no intent all it is is predicting next words and sentences um but there are two ways in which it can go astray so one is it can report something in the database that isn't true so for example in the database or probably a bunch of sentences about um covid involving microchips um sorry coveted vaccines having microchips in them so system reads that from Reddit and it repeats it it's nonsense and a really smart system wouldn't repeat that nonsense it would actually consider is this plausible could you stick a microchip in there would people go to that expense is there any evidence for this the current systems are just mimics they're not actually reasoning about it so that's one problem is they can mimic things that are untrue without validating them and then the Elon Musk car crash example that I gave is an instance where the system because it doesn't actually understand the properties of the world and is just predicting words that kind of sound good together can come up with stuff that just isn't true at all it's not deliberately lying it's not within tension but it means that we can't trust these systems to give one more example Bard you know made up a subtitle for my book that might have made sense it said that my book which was in 2019 talked about large language models which weren't really popular then and we didn't talk about you know just put stuff together that like books of this sort are affiliated with words of that sort without thinking about like the timing without actually looking to see what's in the book and so forth so the complete lack of fact checking um winds up with a lot of things that aren't true even though it's not done with malice per se sure you've been a very critical of chat Bots regarding how they don't really understand the context of language I don't really I guess have a full understanding of anything um but given how rapidly things are moving in the AI space right now do you think your opinion may change soon or maybe sooner than you think well it's possible but I'll tell you this in 2001 I wrote a book about in part why these things would hallucinate and I've been right for 22 years since in 2016 I wrote an essay about why driverless cars weren't going to work anytime soon I've been correct about that in December of 2022 I wrote seven predictions about gpt4 and all seven of them were correct I don't know that anybody else did better than that or even close um so you know I have to look at my own track record among other things and then I have to look at how much the you know mechanisms have changed and so I'm gonna you know stand where I am for the moment but I look every day I see you know did something new come out does it change something fundamental and you know I'm a scientist so show me the data Professor guy Marcus thank you so much for joining us here today I Summit thanks my pleasure and sorry I couldn't be there in person bye-bye

Channel: Scottish AI Alliance

Views: 3,932

Keywords:

Id: rmpPtX2-ok0

Length: 32min 5sec (1925 seconds)

Published: Mon Apr 24 2023