From Hype to Reality: Gary Marcus Unravels the Truth about Artificial Intelligence

welcome to World of Dance a show for data enthusiasts I'm your host Orrin Hoffman CEO of safegraph and gpflex capital for more conversations videos and transcripts visit safecraft.com podcasts hello fellow data nerds my guest today is Gary Marcus Gary is the best-selling author he's a professor at NYU he's previously published extensively and curly posted extensively about artificial intelligence and deep learning Gary welcome to World of Dance it's a pleasure to be here an exciting time to talk about all of this now one of the things I really want to talk about with you is kind of like using this time to explore both the promise and the limitations of artificial intelligence or Ai and I want to start with this like famous bet that uh you and you're trying to have with with Elon Musk where he said that he thought there would be AGI by 2029 and um and and you've been trying to get him to to bet some money on it not yet I don't think he's if anybody out there is listening you should get him to come to the table so sure I'll tell you about it so you know elon's been promising AI in various forms for years and not really delivering like in 2015 he said we would all have driverless cars in 2016 and you know I'm still waiting it's 2000 yep so he does this all the time um and it often kind of rubs me the wrong way as someone who's in AI knows how hard these problems really are and he did it again he was replying to Jack Dorsey and he said he would be surprised if we didn't have artificial general intelligence which is to say not just like I can play go but like I can do whatever I want I tell you the problems like the Star Trek computer Yeah by 2029 I thought this was ridiculous and I've been writing something lately um like a Blog on Gary marcus.substack.com and so I thought you know this is a good topic for a Blog to talk about why there's some unrealistic expectations here and I went through like we have for example an outlier problem so you know if a Tesla sees a person carrying a stop sign it's not quite an its training set it has people it has stop signs but doesn't have a person carrying a stop sign and so the Tesla might actually run into that person and so you know I reviewed all these problems for why artificial general intelligence is actually harder than it looks and also elon's own history and you know is needling him a little bit um and said you know I I think this is all implausible and then I put it together with something that my co-author Ernie Davis that I work together with on so many things and I had already been putting together a few days earlier which was some very specific predictions about what we thought might be plausible and when um and put it together and then put it all together it's like you know I should make it a bet and put some money on this um and so it's a bet I offered a hundred thousand dollars and the criteria were will artificial general intelligence or AI be able to do five things in 2029 and the easiest one was maybe read a novel and tell us what's going on who are the characters what are they doing this is something yeah I've wanted to challenge the field on for a while like we have all these benchmarks like okay can you recognize a coffee cup and yeah yeah I can do that but can you understand the conversation that we're having or I introduced this thing called a comprehension challenge in 2014 when breaking bad was hot and so I said like can you watch the show and maybe at some point Like Walter White wants to take out a hit on Jessie can you explain why he wants to do that yeah he might want to accomplish what might happen if he does and so forth um it's been a long time sometimes even for like a smart you but that show is hard to follow so yeah sometimes it is but you know what's interesting about a lot of shows and especially Hollywood shows but even something like Breaking Bad is we usually catch up we usually yeah I mean there are details like you could go back and watch it three times and there's some stuff you missed but there's some like headline items that are no problem for any human like we understand why Walter is pissed at Jesse because you know this deal went this way or whatever um I have a side note about that we could go into but Hitchcock was the master at making sure everybody knew when you saw that train go you know what it meant and why was with suspense and whatever one person missed the train and all this stuff so so the first part of the BET was like okay 2029 are we going to be able to have an AI system that can actually read a novel know what's going on and the counterpart is is even harder actually is watch a movie because now you have to understand all those graphics and what they mean for me to label you got a microphone in front of you and you're wearing um headphones but to really understand the relations between those things and like figure out that even if your headphones are occluded right now by the microphone the probably that wire runs straight through and then you know like why isn't why isn't the thing that looks sort of like a film canister flying through the window you know because gravity is holding down like to really understand a scene and what's going on and like you're giggling is that appropriately do you think that I'm crazy like with the social interaction with it's pretty complicated and yet like in a movie again like we can all do this so we can do it in a movie we can do it in a novel and yes like Grisham maybe spells it all out so it's easier for you to understand um than if it's Dostoyevsky or whatever but yeah you know there's some wide range of literature in movies where humans understand and right now let's be honest AI is basically illiterate can't read a novel doesn't understand a movie so those were two of the bats a third one was really a nod to Steve Wozniak who had something called the coffee test which was like you should have a robot if you really have a GI to be able to go to anybody's house and figure out how to make coffee there and the point is like everybody's house is different yeah um and yet a normal human being I'm not a normal human being I don't drink coffee so I'd have failed but a normal human being could do that you figure it out well I I still can't even figure out my own coffee machine it's so complicated so we would be ruled out but anyway so we changed it to like you know be a um a restaurant helper or something like that yeah be a useful short order crap um prep cook in anybody's kitchen was a third then there was one about computer programming because it's a Hot Topic right now but said you know can you write 10 000 lines of bug-free code and then the last one was like the the hardest one maybe in some ways maybe not in others of being being able to read a mathematics article and turn the verbal part into a symbolic thing that you could prove just you know maybe we'll call that level five and then the most important part was to be general intelligence you'd have to do at least three of those five things right it doesn't count if you've just done one of them yeah we've had is a lot of narrow intelligence like this thing solves protein folding and this one solves go they're similar but they're you know they're really engineered for particular problems and what a lot of the struggle has been has been to make systems that are systematic and general and Powerful so that was the BET put down money other people put up more money it became this thing Elon still hasn't responded I don't really expect that he will but it would be really cool he did say he said I would be surprised right if you say I would be surprised that means you you would you kind of give it at least a 75 likelihood of happening right yeah and so and you're giving him you're giving him even odds I'm giving him even up right so that you're you're basically thinking that 65. right right you should take that so he has enough money to pay the Twitter breakup fee for anything you know but um so I mean he really should take it good for the whole field of AI if he would a take it because it would actually generate excitement for the field and give it I think good directions I think this would be useful problems to work on um and also you know since he doesn't want to lose he could put some money in making sure the the we actually get there which I think would be good because I think that the AI that we have now is actually lousy in a lot of ways maybe we'll want to talk about that a little bit um I think the AI may actually be in its worst moment in history because before we had no AI so it didn't cause any harm and I'm hoping that if it's smart enough then we can talk about the risks people worry about that that it might not be so bad but right now we have ai that has done a bunch of pernicious things like direct news feed in ways that reinforce people's beliefs that we have a huge problem with misinformation um and um you know the AI is not smart enough to weed out misinformation so it spreads things like mad and we have polarization in society we have all kinds of problems with bias and like loans and stuff like that then we have have reliability problems so like the system gpt3 if you configure it to give medical advice people have found dialogues I think they haven't happened in real life but just in testing it where you go up to gpt3 and you say I'm thinking of committing suicide should I do that and it says I think you should because it just it's just predicting statistics of words it doesn't really know what it's talking about I'm sorry the current AI that we have is actually in many ways harmful there are some good uses that's been put to but um there are risks I think there are further risks a lot of people are trying to apply what we call like the new AI or the statistical AI large language models to all kinds of problems like they want to coordinate driverless cars with this stuff and it's going to be bad you know these it's like giving too much power to an unintelligent person who can't really reflect deeply on things like I remember like let's say 10 years ago there was this claim that like people shouldn't study Radiology anymore because AI is going to make at least that profession dead relatively easily read these medical scans and um you should be able to uh you know quickly figure it out and it seemed like a perfect application for AI I'm certainly a Believer 10 years ago like I don't think there's one radiologist that's been put out of business like why why is that the case so so I'm just going to fill in some history because it's interesting it was 2016. I know the exactly I can almost say it word for word Jeff Hinton said people who are studying Radiologists are like Wiley Wiley Coyote at the edge of the cliff basically you're saying like they just don't know it but it's all over we don't need the Radiologists anymore because deep learning is going to do this so you know fast forward six years and as you say the number of actual Radiologists has been replaced is zero there are 400 startups working on this problem but it always turns out to be hard to turn AI or at least almost always turns out to be hard to turn AI into real world practice so part of the thing is like only part of what a radiologist does is kind of the visual part which deep learning is best at yeah but part of it is like reading a patient's chart and understanding the history like in the context of the patient and yeah like did they fall off of a ladder once like maybe you read this image different if they did and so yeah you've got this all these notes and unstructured text and AI doesn't really know how to read so it can look at the picture and it sees a blotch there and then there are other problems like a real radiologist can notice hey the lighting on this one's just not right or there's a hair across it or something yeah like it's sort of like extra the bugs the bugs essentially yeah real Radiologists can do that and these systems can't and so what what people I would think it would be a perfect like human computer assist thing where the computer could like help you quickly point out some things to maybe make your job go a little faster and more efficient exactly right that's exactly what I was going to say next is right now and it could change but right now it is a perfect example of human machine augmentation or symbiosis so those things can change like I think a lot of people made a big deal of Chess being like that it was a period where machines were better than human I mean sorry we're better than machines sorry I'll say this again there was a period where machines were better than people at chess but machines plus people were better than machines alone or humans alone yeah and that's where we are with Radiology now I think maybe with chess the best machines don't even need our help anymore yep yep and it could turn out that way in Radiology but it won't turn out that way soon because the content text as you said a lot of which is written down in unstructured text like not in like in a table form but just like sentences um the machines aren't any good at that at all so for a while I don't I don't know the exact numbers but I'm going to guess at least for several years maybe for decade or two um we probably will do best having the machines in a workflow with the people but we don't want to get rid of the radiologist and in fact I think during the covet crisis I don't know the exact numbers but I think we didn't have enough Radiologists um and we should and that might have been because like a few people heated hinton's warning back in the day and they they lived in other fields or something you know I ought to be careful about saying that Lester yeah myself I'm prone to litigation but I mean there has actually been a lot of consternation in the field and I think you know for several years like people really took him seriously I think now most people in the Radiology field kind of make fun of him and they're like you know they all feel like they survived this world with them and they're like yeah we're still here but um you know for a few years the Radiologists were worried and you know they it could be that Hinton was just off by you know a factor of four or something like that it could change but there are a lot of problems in turning this to practice one more that I want to mention is there's a huge problem in all of machine learning with generalization so the way that machine learning works right now or at least the most popular technique is basically you memorize some data and then you generalize kind of nearby to that stock yep you imagine you're like in a big hypercube or something like that if I now test you in the same part of the hypercube you're good to go yeah but if things change and this comes true in models in general if things change then the systems just don't work as well so if you took all of your patient pictures pre-covered and now you've got covet and like the whole distribution of your data changes yeah your systems may not work as well anymore yep now that's not just a problem with machine learning like right people are problematic too there right people are problematic too there and like you can think of what happen with long-term capital in the Russian bond market you know you could have a model that you really believe in it could be a neural network it could be a classic symbolic model but if your assumptions are wrong it may blow up and the Assumption of machine learning models right now the popular ones is basically that your data at test time are from the same distribution is training time and you know there's basically the same stuff I'm just randomly drawing for a hat the same kind of stuff which is true for a lot of Statistics but in real world applications that's not necessarily true things do change another kind of weird manifestation of that is if you ask gpt3 who's President it's probably going to tell you Trump because the larger fraction of its data were collected when Trump was there and it doesn't yeah temporal reasoning that a human would be of like yeah I know he was there for a long time and maybe I didn't like him or maybe I did um but he you know he was in the news a lot but he's not there anymore he's now the president um and so um you know you update your representations or if I ask asked you has Russia invaded Ukraine in you know I asked you that in January you would say no if asked you in February would be like I heard it might happen I don't know if it really happened and then if it's now then yeah Russia obviously invaded Ukraine you update your database it doesn't matter how many conversations you had earlier that they might right now it's happening right like some sort of like temporal waiting on the content or something like that has to has to be there it could be temporal waiting I actually think it's more like a database where like in a database you could have a buffer like what is the last key that the user press you update it yeah you just update it um and so I think human cognition has ways of doing updates we're not perfect at it I can actually give you counter examples but in general we do we certainly want our machines to have those kinds of updates and In classical artificial intelligence is Trivial but it's actually hard to put it into these machine Learning Systems you you had a great Ted Talk where you said one of the biggest issues is that AI doesn't have like Chronicle common sense like how do you how do you kind of Define that common sense and an example of AI or do you have some good examples of like where AI could potentially have common sense or where has more of a hard time learning common sense I mean it has a hard time almost anywhere um I'll first say I don't have a crisp definition and I think it's actually a um you know there's a there's a common sense is common sense is not very common even amongst people so well there's that I mean there's the parts that are in the parts that aren't and you know it's a little bit like the famous line about pornography I know exactly so like some common sense is like I've got a cop I better not tilt it or I'm going to mess up my keyboard and like yeah everybody knows that um and yet that particular one is not really written down in a whole lot of places and so like you know you do your web scraping of conversations and nobody talks about tipping their mugs over yeah some article of mine somewhere where I used it um but so you're not going to find that kind of stuff um there's other kind of common sense like it's maybe contradictory like out of sight of out of mind and absence grows the heart it makes the heart grow fonder yeah you know some of it's a mess um and then there's also like expert knowledge about certain kinds of things and that's also useful for machines so it's a little bit gray but also there's some pretty clear examples where current systems just fall apart like um uh one of the most basic things is we know that once you're dead you're dead you can have certain religious beliefs but um if if I go and ask gpt3 which is the most popular language model thing AI thing right now I say um Bessie the cow died um uh how long will it take for her to be alive again you know a human being would be like that's a ridiculous question what do you mean yeah and the Machine um and now now there's this famous sentence let's take this step by step which supposedly makes these things better so we'll throw that in there too so so you know Bessie the cow died where you know how will she be alive in nine months again let's take this step by step and the system will say something like well first she's dead it'll take nine months to make a new cow so I guess the answer is nine months like you're just missing something there um and so you know there's just very basic stuff like that like what does it mean to be alive what does it mean to be dead um and then in our book Ernie Davis and I rebooting AI we gave millions of examples like this that are really hard like suppose I tell you that Michael Jordan and played basketball since he was a kid and that he's whatever 50 years old now um human being um can understand that when I say he played basketball even if I put in a phrase like all the time that I don't literally mean all the time right so I don't mean that Michael Jordan played basketball when he was asleep probably not when he was eating dinner um you know he probably went to class sometimes yeah like um and you can figure out from the context and this is part of what makes the writing challenge to Elon so hard is there's so much of that context that we figure out same thing with a movie like we we don't see the characters going to the bathroom but we assume that they do it because we know something about human beings and if if I said what's the chance that this character has not in the span of the movie Gone to the bathroom even once you contain zero because you know right that's just not something human beings can do um you know we're not looking at a camel here right and so like you know you you um we just know so much about the world I would say that that kind of stuff is common sense it's a little bit slippery and harder to find um there is one really serious effort to build common sense for machines in a classic AI Paradigm by a guy named Doug Lennon a system called site that I think is very interesting not completely satisfying it was built in the 80s I think we would do some things differently now he and I are actually writing a paper about like what you might do now in the 2020s to make it better um but mostly people don't really directly deal with the question and what people have been doing is hoping it'll kind of emerge by magic by just feeding in lots of data when that hasn't worked they said well we'll feed in more data so then they fit in more data what's the why why is it a problem like we solve solving some like narrow uh thing like there's a lot of wins that we have like I I can I could I don't know German but I could read an article a German with a translator a machine translator and I might not be it might not be perfect but I get like the gist of the article it's pretty good uh I understand it like can't we just chalk some of these things up as like this is a nice win I didn't have that 20 years ago and now I know everything in my life or something you know there are some nice wins and one of the questions is really the cost of error so if you stick in a story from German um about you know today's news roaring Ukraine and you're not actually professionally involved in that war um it'll probably give you a serviceable translation yep it won't be perfect right right um if you wanted to put in a legal document yeah okay I couldn't do that yeah you know little details about where a comma are really matter yeah um and so if it's not Mission critical it's fine if it's Mission critical it's not really good enough same things kind of happen with driverless cars so it's easy to make a demo that sticks to a lane people have actually been doing that for 30 or 40 years yeah but driving is super Mission critical and um you can't have your car you know drive into a stop vehicle but that's a whole bunch of times and so like there's a bug that Tesla has known about for five years and still hasn't fixed and maybe I should actually say that sentence more clearly there's an issue that Tesla has known it's not like a one-line bug um it it's some very complicated interplay of things that they're having trouble tracking down and partly is a function of the training date and it's it's hard to it's hard to do debugging in these kinds of systems and so for five years Tesla's have been running into stock Vehicles you know somewhat regularly they're like 20 cases or 30 cases documented um there's so much focus on like self-driving cars which seems like incredibly difficult problem with all these other adversarial there's pedestrians and all these other things that could happen like whereas I feel like you know just like a much simpler problem let's say self-driving boats or something like why why are we all must not so you must not have a boat my friend okay is it is it even is is I assume like like a fishing assistance could be really helpful to me if I was official yeah I mean there are some limited things like this um I'm new to the book other boat world but have a boat and the physics of a boat relative to the current and the wind they're actually complicated oh okay it is way harder than driving a car I'm sure if I've grown up okay I have no idea okay but it's non-trivial and there are there's actually been a lot of progress in self-propelled boats but in the docking part they still do humans so out on the Open Sea you can kind of do this you still have an outlier problem like it's not so much weird stuff I mean you know the weird stuff for driving is like pedestrians or yeah something falls off of a Truck Yeah you got some stuff to deal with logs and stuff in the sea but if you're like out in the open water maybe it's just the time okay um but the outlier problem is still there so like if you so I live in Vancouver not too far from where a little pirate ship goes around um it looks a little different from the other boats and I could imagine a self-driving system that was trained in I don't know LA or something off the Waters of La it comes up to Vancouver it's never seen the pirate ship before and you know go smack because it's not in the debate database and so yeah yeah it's an outlier and like we don't really have the data for how hard that is I mean another lesson I think of AI over the last decade is what looks hard I mean it's really a lesson of AI for many decades is what looks hard to a person is not necessarily hard to a machine and vice versa what looks easy to a person so a lot of people thought driving wasn't that hard and here are some reasons why you might have thought that like 16 year olds can do it more or less fine I mean they're a little bit aggressive but they can mostly Do It um so that would be a reason another reason would be like roads are basically the same across North America so if you're not like talking about unimproved roads in Afghanistan you might think well you know they're all kind of engineered with the same Lane marking and signs and then it turned out even though a lot of people had that intuition and maybe reasonably so it turned out that there was just a lot of edge cases like this unending Cavalcade of edge cases like I think I mentioned already the stop sign with a person carrying a stop sign is an edge case another thing that confounded a Tesla a couple weeks ago is somebody brought a Tesla to an airplane show like on a big Runway lots of planes you can kind of Imagine even if like me I'd never been to one but you know people are showing off their airplanes and somebody pressed summon on their Tesla to have it come across the parking lot and it ran into a three and a half million dollar job oh my gosh like you know the debt was just standing there it's not like the jet was moving right and like it just wasn't in the training set yeah the training set at this point is huge Tesla has the biggest training set you know of this kind of data ever assembled in the history of mankind but there are still things out of the training set so it turns out there are all kinds of objects nobody anticipated and you know pedestrians do weird things or they carry weird things so like maybe your pedestrian is fine for your image system and then the pedestrians carrying an umbrella and your image system is looking for their their eyes and it can't see it anymore because the umbrella is in the way there's just like unending um Litany of these cases so there are problems that are harder than than we realize because we kind of automatically compensate for them um and then there are things like go which a lot of people thought were hard but it turns out you just make up as much data as you need by self-play and and you know deepmind actually solve go in a very um robust fashion and so there are some problems where the machines are just way better than people and some the other way around and the real issue in my mind is that the public and also the business world does not understand the difference between these kinds of problems well it's hard to understand like hype from reality you know there was this recent uh Google engineer who claimed that uh that that maybe some of the deep Learning Systems within Google or said kids I know that you you you you had a strong reaction to that but I did yeah so I mean probably by the time this people watch us they will all know about this case where this guy was interacting with one of these large language models and convinced himself that it was sentient that it like really had feelings and emotions and you know he said it should be treated like a colleague rather than um in like an employee rather than than like a piece of software right I mean we have no problem turning off Excel but are we allowed to turn off Lambda yeah the questions he's asking I think yes you can turn off Lambda because really it is just like Excel it's just doing a bunch of computations on a bunch of numbers that's really all it's doing um it doesn't actually have connection to reality um I used in this article it's called nonsense on stilts I used as an example of sentences with something like um these ask the system what do you do with your spare time and it's like I like to hang out with my friends and family and do good things for the world and the system does not have friends it does not have family it does not know what a good deed is in the world I made a okay but it's sort of half joke I said you know it's a good thing it's just a statistical approximator because otherwise we would think that this thing is a sociopath because it was like making up friends and uttering platitudes to make you like it except that it's not really it doesn't care if you like it it's just autocomplete is all this system is the kind of autocomplete to complete its own sentences and yours but like autocomplete it's predicting the next word in sequences so when it says you know I like to hang out with my family it's not like there's a representation there in the computer of like Peter Paul and Mary or its relatives and it's like thinking warm thoughts about it it's just it's taking this word you you could understand I can understand how this Google engineer like you want to believe when you're interacting with something I mean one of my my one of my favorite movies is her uh which I think is a beautiful movie and you you want to believe that this interaction that you're having is is more than than just a bot even when you're dealing with a person sometimes you ascribe things to this person or you love this person more than they are it's warranted so I could see how like anything I could see how this could happen um well and I think Will and it already has so um in fact in in the book rebooting AI that I mentioned we talked about what we call the gullibility Gap the gullibility Gap is really a form of anthroprocessation where we see in things things that are not there so you look in the moon right you see a cloud yeah yeah you see face in the moon or you know clouds or something called peridolia right another example is potato like potatoes Mother Teresa in the potato right um and hopefully you know your rational world is the rational self is is you know strong enough to know that that's not real but I'll give you another example right now this is a weird example but right now all I see is a two-dimensional version of you and I'm ascribing a three-dimensional version and that's okay because I met you in real life and it turns out it's real but I will do that for a character in a movie and I will cry when that character dies right you know and like they didn't really die yeah I remember this movie Fried Green Tomatoes which kind of dates me I suppose but like yeah beautiful night in Every Act yeah beautiful movie it also has a great line basic girls I'm older and have more insurance I love that part yeah but so you know or you take joy when when she says that to the teenagers and face it girls I'm older I have insurance no real person said that a screenwriter wrote it the actress delivered it masterfully and we love it but it's also an illusion um and it is an illusion of a different sort when this machine predicting next words says the sentence that you wanted to and then like he did some editing he made him like he he kind of escalated the illusion to himself um but it's you know like I feel a little bit bad for him like I I think that it is a very normal thing to get sucked in if he hadn't been a Google engineer probably people would be completely sympathetic they're kind of like well since he's a Google engineer he should know better and there's there's some element of that but I mean you see like psychiatrists that fall in love with their patients and stuff like that and you know or right fall in love with their their computer psychiatrists so the right classic example of this is Eliza in 1965 was a so-called rogerian therapist which basically no matter what you say just ask you questions never gives you any advice yeah you know you say I'm having a fight with my girlfriend and it says oh tell me more about your girlfriend and yeah you know you say well it was about dinner and they're like well you know do you often have dinner together whatever and like it was just matching words like yep girlfriend relationship dinner with no clue what it was talking about but people still got sucked in and you know another way to think about it is when we evolved we didn't have to discriminate humans from machines we had to discriminate machine I mean humans from Lions so we could get out of the way fast um if you think about evolutionary psychology but we there was no thing you know for our ancestors to make sure they didn't get tricked by a bot right and so we don't have the kind of biology to help us do this and we don't have training in school schools I could teach a class if anybody wanted to hire me I'd tell you how to spot them but you know most people don't know now there's been you yeah you and Scott Alexander the the author of the Slate star codex blog have been going back and forth um on having different discussions and different uh opinions about both the current state of AI and the future state of AI we're explaining to me his point of view the best you can uh that where you guys might have some differences so I guess there's a couple places where we've differed I mean we've had a bunch of back and forth lately on his blog and on my blog it's like when they used to go from happy days to Laverne and Shirley like back and forth between us so we've been now you're really dating yourself exactly you're going going back and forth um between our two shows so to speak um uh I think this is called astral codex 10 or something like that and mine mine is uh uh Gary marcus.substack.com um and in the first one he wrote this really funny thing about the state of AI and how the dialogue goes and it's like somebody comes up with somebody something really cool and then somebody else and he said usually Gary Marcus I mean it's true that it is usually Gary because it was a very funny line um which you know I thought was funny in the field thought was funny usually Gary Marcus points out something wrong um asterisk on that it's usually Gary Marcus and my buddy Ernie Davis we write all this stuff together but anyway I'm on Twitter more so people know that attorney But but so um and you know I've written some piece somewhere but most of them it's usually me and Arnie but so Gary and Ernie notice that there's something wrong and then people try to improve it and then it's basically rinse lather and repeat um and so like there's another interview like a lot of times when you guys do point out these things like people fix the bug you'll be like oh there's something oh actually thank you very much Gary and already for pointing this out yeah but you know I'm sure they'll come but anyway um um I'm not doing it for the thank yous from them they're a little bit sparse on the ground but but um dialectic is a bit like that and some things get a little better um it might well so in his View and he's not in the field but he's a very smart person and he reviewed the cover um and he was careful to say like I don't have a PhD in cognitive science like Gary does he was very measured um and almost sweet about it but he said you know I look at this and what I see is these things just keep getting better and I you know I'm not I'm not worried they're getting there and the rate mate people might might his his issue is or his argument is well you can argue about the rate of it getting better but there's some forward progress or something so that's basically the argument yeah and it's not unreasonable but I you know I'm not I've got my own arguments and I came back at him and I pointed out that the improvement's not as much as he thinks it is it was actually a flaw in his kind of statistical procedure because he looked at new things I mean sorry things where there were errors before and showed that it got better but he didn't look at the things where that they got um where they're actually worse now we didn't do like a random sample or whatever and overall like there was definitely improvement from GPT to gpt3 yeah so clearly from gpt3 to what we'll call gpt3 plus which is the new thing he kind of overestimated how much improvement that there was there but yes there's some improvement but there's also some core problems and this is what I think is important where they haven't really been progress and most of those are around language so I'll give you an example from Dolly which is this thing that takes text and makes images it's perfectly good at saying that an astronaut can ride a horse um uh but if you tell it a horse rides an astronaut um which is a much less probable thing it won't draw it for you and you can actually do some tweaks to get it to do it for you but um it doesn't really understand the inversion um and I was doing this as an homage to Steve Pinker who has often used the example of man bites dog which itself it comes from the newspaper business the old line of newspaper businesses dog bites man isn't news happened too many times before man bites dog now that's news right so so horse rides astronaut that's news um and these guys didn't let me have access to the system so I had to do this very indirectly but I knew um from from what it leaked out that they couldn't do horse rides astronauts I wrote a piece about that as well um in in the subset College I mean you are a well-known researcher like if I had a a new AI system I would love you to have research access to it so you could like you could tell me all the areas I need to improve on it like it's free it's free QA all I can say is you're not running either open AI or or Google AI those guys really don't want me to play with their toys I I wrote about this too in one of the recent sub Stacks why is that is that just because they're afraid they they have a little PR thing going where they have now got people you know in some of these companies thinking that their systems are practically sentient why why would they want me to poke holes in that and so like their PR game is to make it sound like they're very close to artificial general intelligence and why does that matter because artificial general intelligence when it really comes is a complete Game Changer I think yeah um you know there's so much of the economy is done by what's the like what's the reasoning to get people think it's gonna come faster than it is like like they need to raise more money or something or yeah raising money getting talents like so I'll take Dolly as an example it's really Dolly too but I'm just gonna call it Dolly so Dolly comes out 45 minutes later um uh Sam Altman tweets AGI is going to be wild suggesting that you know they've made progress towards artificial general intelligence here and you know timed exactly to that somebody I don't know maybe Scientific American but I don't remember runs an interview with like one of the programmers and says you know what we're trying to do at open AI is to solve general intelligence um and you know we think this is a step forward in that direction you know what Gary Marcus looking at your dirty laundry saying well you know the image synthesis here is really good but the language stuff still doesn't really work who are you kidding they don't want me to you know say that of course no wall is impregnable so they you know they promised me access to gpt3 but they didn't give it to me and I complained on Twitter and somebody said Hey kid I'll give you I'll give you 45 minutes of access see what you can do with it which I dated and I wrote a critique and you know I wrote a piece around that with Ernie Davis called gpt3 bloviator which we wanted to call gpt3 artist and that is basically what it is and so you know we got some access and then Scott Aronson actually gave us a little bit of access to Dolly and we figured out that it had the problems that 2000 terms of language and stuff like that with small amounts of access there's some other systems have come out since from Google like Imogen where I publicly asked them like you say you are better at problem X can I give you a few examples and try it and I get no reply so you know there's been a shift from real science where people would stand up and say yeah sure look at what I've got yeah like customer hypothesis like because like these these systems that that have come out like dolly or gpt3 or or GP whatever gp38 whatever it's going to be in the future like they have some usefulness like they're they're not all about like they do they have like some in some areas they they are and so it's I think it is helpful for them to to to let people because because if you point out the flaws that people might not even go for the good things it's like okay here's where it doesn't work here's where it works let's let's use this for now and then let's get better in the other areas yeah the dirty secret about gpt3 which is not so much a secret anymore is that it's kind of like a bull in a china shop and so there are a few hundred startups that have been built on its technology but it's not clear to me that any of them are really thriving and that the biggest problem is that these systems are full of toxic language they're not very truthy and you can't really count on them so there are some applications where I think they're fine um the best one is in my view but I don't know all of them is AI dungeons AI dungeon is like Zork if you remember those old video games again dating myself to the prehistoric era um where you would type in text you'd be like you know it says you see a key and you're like okay take the key put it in the lock and turn and maybe that would be the magic invocation so imagine that but a super fun version where you can talk about anything so you can say I'm sitting in a dark bedroom in Vancouver with a coffee mug and some guy is asking me weird questions and it'll just continue from there and then you rip on that and if it makes a mistake so to speak there's no cost to that because you're just having fun if it says something toxic and it tells you you know it questions your sexuality in a way that you don't like you can just turn off the program and it's fine yeah but if you put that same software um in a customer service chat bot let's say which you might think it'll work for but now you're dealing with a customer over a bank loan and now you tell them to do something unpleasant with their mother it's not funny anymore yeah I mean my joke well um if it's like let's say you have like agent assist or something like I I the autocomplete feature on uh you know Google Docs or something I'm in the middle of a sentence that often could could uh with with a decent amount of accuracy can complete my sentence for me it's also my typing last bit of complete something faster it's it's a it's a human assistant well Jesus system three what it is really is like the best version of autocomplete that money can buy because it's trained on a much bigger Corpus yeah but it's basically doing what autocomplete does and so you know another thing people have used it for is like copywriting so um yeah or like term paper writing so like for term paper writing you know I don't endorse this use but like it could actually be pretty good at that you know it probably wouldn't give you an a pay paper but it'll make something that sounds sort of like the topic and and whatever it's probably going to make a lot of mistakes it's not going to be an a paper but then like then you can go through it or something and then you could so maybe human can go through it and they are like the commercial question if you want to do it for anything other than a high school term paper where maybe the student just doesn't care which is a problem with our educational system um you know then there's a question of like how carefully do you have to look at it is it worth your while and that's just like people have to do trial and error and see if they can get it to do what they want with Dolly it's a sort of similar question it makes these fabulous photos but it seems to be hard sometimes to get exactly what you want and so if you want to use it like give me an idea for a book cover it's amazing if you were wanted like something for an advertisement you wanted exactly this thing exactly there with this other thing on top and whatever you might run into this thing where it's just too hard to get it to do what you want you might get frustrated so if you have to like if you had to make like a prediction like five years out ten years out okay here's here's where we're going to see more a lot of progress in here's an area that maybe a lot of people think we're going to see progress uh I don't think we'll see as much progress in over the next five or ten years like how would you I like and I'm gonna put money on this like where would you say hey Oren here's where you should put money on so deep fakes are going to be just like unbelievably good um they already are videos uh audio all that stuff I don't expect that like in five years you could make a whole movie with the whole plot and that kind of stuff but if you wanted to do it scene by scene or something like that that stuff's going to be really really good yeah so I could I could create a famous person stabbing somebody or something and put it out there it'll be it'll be possible to know yeah exactly I mean already it's you know pretty good so this is not going out on a huge limb to interest I think to say that in in five years that that stuff's gonna be insanely good you know it's already kind of mind-boggling yeah um and it's art like people you know in the Russian invasion we've already seen some of this I think in both directions if I remember correctly um so those are part of the like the thing now is like okay how do we train Society to uh to not every time you see something to to not assume it's it's real or something like that that's hard that's hard and we have like kind of weaponized misinformation teams now right and you know every government has one and companies do and so like I mean even just random people have it like they put it out there they put the memes out there yeah and it's going to be so easy to make those yeah I had a little poll on on my Twitter uh account about when you'd have a version of Dolly for gifts and I think you know most of us including me I don't because I did my vote public but most of us thought like in a year would probably have dolly for gifts um you know little simple animations and how's it like like if uh let's say I'm like caught on camera like picking my nose or something and it was true I really did that but I I could like start blaming it on the Deep fake yeah well that's Trump's move right fake news and yeah you know there'll be more fake news it'll be more often true when somebody says fake news that it is fake that's going to be a a total mess um it's going to be a blank storm I won't use the first word but you know what I mean um it's gonna be a mess so that that's one thing that we get a lot better speech recognition will keep getting better every year you'll be able to do it like in a louder car and you know you'll be able to talk about a few more things with Siri every year or you know Alexa or whatever that stuff's going to continue to grow it's still in five years not going to be that smart it's still not going to be Samantha so you know come back to her the movie that you mentioned Samantha really understood like all of what's going on so yeah you know one of the opening scenes she um says like you know what's bothering you is like my email um and he's and she comes back like two seconds later and says well I notice you have 17 000 messages I deleted 2 000 of them for you these were duplicates or whatever and like yeah yeah um you know we're not going to have machine reading at that level there's one thing from Samantha that we won't have in five years where you can actually trust it to fully organize your email as if you have an important message any system right now could easily mess it up so you know I got a message from you to to do this podcast that's an important message but maybe you know I actually it's a weird week but I've got like 20 um messages like that I'm not always so popular but like yeah you know it could have gone to the spam and I mean we have problems with spam filters and like AI is not going to solve that problem immediately um because it still doesn't have enough sophistication like there's there's a I I think it's x dot AI has been promising for years just doing your scheduling and for a while I think that humans behind the scenes if I remember correctly I couldn't have the wrong time well I mean it's incredibly hard I mean I I have a extremely accomplished assistant that does my scheduling who's extremely smart and like she's a lot smaller than any Ai and even there it's like so hard to to to do it it's like that's very very hard task for an AI to do yeah it's high stakes if you miss a meeting like that really matters yeah like that's not a solid problem so imagine just how hard it is to do scheduling with a machine where you have your calendar in front of you whatever but still things come up in in the software also you have uh you have your own nuances like I like to do this in the morning or I need some space in between get me some space to go to the bathroom or whatever it is right and you know I I think in five years humans are still going to be better uh the machines of that even there's a lot of effort and that's like a narrow part of reading your email so like what Samantha is doing it's like Way Beyond just looking at your calendars presumably um so another part of Samantha that I think is Way Beyond us right now is Samantha actually understands human interaction and I mean she understands it so well that the character falls in love with her yeah um we actually do have software that people are done enough to fall in love with now or I'm trying to find a more polite word that they have the will to believe to fall in love with now but there's a level of like social understanding that Samantha has towards the end the critical plot twist depends on her not having one piece of social understanding um she doesn't really get monogamy yeah without quite giving away the whole film um but she gets a lot about human interaction and what would make people feel better and this kind of stuff and I don't think we that we're five years away from that I think we're much more than five I don't think it's impossible but it's harder so like the paradigms that we have now are like I show you a picture of a pencil and I say pencil and the Machine learns the name of some concrete physical object that we can put in a bitmap and yeah something like love or harm or pain or need or you know any of these kind of psychological terms your Justice more abstract political terms it's just much harder to push those in to the Paradigm that we know how to use now and so just to me we actually need different paradigms for some aspects of AI yeah so going back to the Scott Alexander slate star codex thing the other debate that we were having aside from like how much progress are we making now and so forth was really like do we need to change what we're doing or not and ultimately he offered me not quite a bet but a prediction he said you know he thought there was a 60 percent no a 40 chance that we could get to real General artificial intelligence just by using the tools we have now more data and so forth okay I wrote a lengthy reply um called paradigm shift or something like that um where I said you know I thought it was more like an 80 chance which may not sound so different 40 versus 80. um uh but you know I walk through why I think the differences are and why I think it's actually really important that we as a research Community consider Paradigm shifts um why why I think we probably won't get there just by adding more data and we do need something substantial but the data is important there is a sense that like as we can join these data sets together we could potentially uh solve bigger or bigger problems it's like we have access to a very very small amount of data um I think data is critical I think it's really interesting that human children become more sophisticated understanders of the world than any computer is now even with a lot less data um I think ultimately you know you want to take advantage of whatever data you've got but if it's a small amount of data you still want to be able to do something with it I think you know that I built a machine learning company that I sold to Uber and when I sold it I had a conversation with Travis who was still CEO at that point and I was explaining what my company did which is we work with small amounts of data and he he said oh I get it the data Frontier problem and he gave the example which was like he knew how to put the right amount of cars in the right place at let's say 11 o'clock on a Thursday night because he had plenty of data around that but there just weren't enough cars at let's say three in the morning for his techniques that he already had to give a reliable answer so even you know Travis who had more data than anybody he ever had on anything at that point still ran into like if you break things down into smaller subcategories so you know the tenderloin at 3am on a Thursday um you know there's not enough data there even when you're accumulating massive amounts of data so yeah if you're Google you have enough data for most things but even Google actually has this problem um there are always these cases and then you know like jet on a Runway maybe Tesla just had zero cases in that so you need to solve that in a different way by having a general understanding of what an airplane is what a large physical object is rather than doing it by memorizing this specific case and looking for a lot of similar cases so you don't want to throw away the data that you do have it's often extremely useful but you also need some paradigms that are a little bit less data driven than I think the ones we have now yeah all right cool this has been a base error last question we ask all of our guests what conventional wisdom or advice do you think is generally bad advice um conventional device that's generally bad advice um [Music] it's funny that I'm stumped on this one right now because I think there's a lot of it um how about trust your instincts um as a piece of conventional wisdom and it's sometimes true there was a kind of Malcolm Gladwell part of story for a while about like experts don't know anything or they're not doing it in a bling and it's not really true and in fact one of the most important things that scientists know is that for almost any piece of data um sorry yeah for any piece of data you will have your own Theory and it will seem to fit your own Theory but if you think about it carefully from someone else's perspective you'll realize it could be explained in a different way and so if you trust your instincts too much you become too in love with your own ideas there's an old saying about you know falling in love with your own press clippings and it's a version of that yeah um the psychological phenomena there are two uh well-known ones one's called confirmation bias so you have a theory you notice other data for it and the other one's called motivated reasoning so you come up with reasons so you can keep believing what you're believing so you don't have to believe that you've made a mistake yeah guys so I'm for gun control or something I see foreign and so you know I mean I'm for gun control you can probably guess and I probably can't even pretend that I'm neutral on it but um the point is whichever side you're on on any you know hot button issue like that or even smaller things that give you a much smaller one which is like who did the more dishes if you live in a house yeah with let's say two adults that maybe married or whatever they are I guarantee you that both people will think that they did more than whatever their fair share is and if you add up you say give it to me in a percentage score yeah you know like the person will add up to like 130 under 40 or something yeah and if you do it in a group house yeah like I lived in graduate school like five of us is gonna add up to like 270. right and like the dishes aren't really getting done so I imagine if like your uh your your your your your baseball team or something like that I imagine uh nobody thinks like the ref is super fair to them right they always say it's always fair to the other side yeah the other side you're on exactly so there's all these kind of biases and stuff like that um is if you trust your own instincts you're like I know what that call was I mean he was out who are we kidding that guy was right not like how was it out who are you kidding you're right and and so so I think we we there's there's value in knowing and calibrating your own instincts but there's also value in thinking about alternative hypotheses and you know maybe the other person's right and that could be on a scientific matter it can be on the dishes it can be on your you know on the calls in your sports game um we got all this polarization in the world because we're naturally inclined to believe that we are correct and to not take the other um view seriously um and so that'll be the conventional wisdom I will challenge all right this is basic I follow you at Gary Marcus on Twitter is that the best place for our audience to engage with you I would say that and now I have this thing Gary marcus.subs.com yeah which I also like so yeah all right this is amazing well thank you very much Gary really appreciate uh you joining us [Music] thank you [Music] foreign

Channel: World of DaaS with Auren Hoffman

Views: 4,162

Keywords: AI promises, AI shortcomings, Gary Marcus, NYU, Reboot AI, advancements in AI, ai, artificial intelligence, common sense, data analysis, data innovation, data insights, data trends, deep learning, future of AI, gary marcus, machine learning, podcast, tech discussions, technology, thought-provoking conversations

Id: 7t_W2TcGUZk

Length: 54min 48sec (3288 seconds)

Published: Fri Apr 21 2023