Are we on the verge of artificial generalintelligence, or are we still far behind? Can wetrust these systems and what the heck is ai anywayand? Today i have the perfect guest to discussthese things. So hello and welcome to my showmy name, is roya zevich and in this channeli discussed science philosophy, religion, andcreativity, artificial intelligence with themost, interesting people and scholars from allaround the world. If this is your first time, hereplease consider subscribing and hit the bellbutton and donate a few bitcoins to my walletand today, i’m honored and privileged to have a myguest on the show uh. This is the first time i havemixed feelings for my guest and before i explainlet me introduce my guest, so my guest today isprofessor gary marcos, professor marcus, is, is ascientist author entrepreneur and the professorin. The department of psychology at new yorkuniversity is the founder of two ai startupsand, probably will be more soon. His books, includeguitar zero cluj and most recently rebooting ai, which he wrote with ernest davis, so hi gary andthank, you very much for coming to the show todaythanks for having me and sorry ittook so many months to set it up. Butokay and many thankful to irish parent. For uhmaking this possible now let me explain why i havemixed feelings for you. A year ago i decidedto write a book on the subject of artificialintelligence to the general public and i trulybelieve that we are almost there deep learningwinning goal deep mind inception the progress ofnlp, but then i spoke with iris and she referredme to your book, rebooting ai, and This book justblew my mind it i didn’t read it. I just swallow itand, but basically your book just winseverything for me, people ask me: whatabout your artificial intelligence book. Youyou say that you are going to publish it saylisten in recent months i read a book. Thatcompletely changed. My mind and if i can summarizeyour book in a sentence, it’s not, we are notthere. Yet you are saying that we are on the onetrack now. Would you consider yourself like the badboy of ai or like a sort of prophet of apocalypseoh? I don’t know i. I do think that my predictionshave been a lot better than the ones that you’dread in the media, so you can go back to my 2012article in the new yorker on deep learning and ithink.

I projected all the problems that we’reseeing now problems with causal, reasoningand knowledge and generalizability and there’s a2016 interview. I gave john brockman on the edgewhich i just re-read recently, and i thinkevery word that i said there was truei’m not widely known as a prophet, but i think ifyou, look, you know at the data. I think you know ami a prophet of doom. I don’t know, but i’m a prophetof. We need to get our act together and we haven’tyet, and so i i like what you said a minute agowhich is i’m not saying we’ll, never get there i’msaying we’re on the wrong track. It’S people liketo um, to caricature me and say well, marcus thinksai is never going to happen and that’s stupidbut. Of course, that’s not what i think i think aiwill happen and i wrote a book with ernie davis totry to get us on a better track. I do think it canhappen, but i think we’re on a lousy track, rightnow, where we have ai, that we can’t really trustand, isn’t really very smart, and so it sort ofseems to me like the worst moment in historyfor aios right now, because the stuff that we haveis So mediocre it’s dangerous. You know when i wasa kid. There wasn’t much ai in practice, so itwasn’t, very dangerous and the ai, when my kidsgrow up will probably be pretty good but rightnow. It’S not a great moment for ai. Doesn’T meani hate ai. It means i hate. You know the way we’redoing it now and i want to see us do it betterokay. Now, when i read your book, there is one namethat came to my mind over and over again andthis. This name was marvin minsky and i think thatcom comparison is inevitable. Now in 1969, minskyi’m giving you a trivia fact by the way, marvin’slast appearance in in semi public at a privatefunction um, but last stage appearance was withme kind of a interesting historical fact. Yes, uhtti vanguard, i don’t know if it’s recordedonline, but that was, i think, his last appearanceoh. So now uh we and he he resembles uh myneighbor. I have an old neighbor which looksexactly like marvin minsky and it’s it’s scary. Nowmarvin means in 1969, one of the founding fathersof ai, the godfather of ai, along with samuel puppetwrote, a book called perceptrons and in which youargue that a simple perceptron is unable to learneven a very simple logical function like so andthose function can be learned only withmulti-layer, perceptrons Or neural network buthe proved that the training of the weight willtake forever and many people today regard thisbook as what started the winter of ai. Now now weare. Can i jump in for one second yeah yeah there’sthe claim that they really made was thatthere was no guarantee of convergencein learning, non-linear functions with multi-layernetworks and a lot of people blame them for sortof the end of neural networks, which ithink is false. We can talk about, whythey, certainly slowed them down, but i don’t thinkthey ended them. I think that they slowed downbecause. They weren’t, really working um andyou know in 2005 was a long time. After 1969and, people were playing around with themand, they just weren’t very good thenand, so they were ridiculed not from the kind ofminsky side of the spectrum, but the svm peopleridiculed the neural networks, because they weren’tworking then – and you know – i think we all know – thatwhat changed. That was having a lot more data, andgpus and so forth, but people like to pin thaton minsky and i think what minsky and papert saidwas actually true, which was there’s no guaranteeof convergence and we what they got wrong, was youcould do a lot of stuff, but not Everything withoutconvergence without a sort of guarantee you knowminsky was a mathematician originally and and hewas looking for proof, and there was no proof. Thatthose systems were going to work and turns outeven. In the absence of proof, they can work, insome things and not others, and it’s still veryempirical in a way that a mathematician might findmessy like yeah. These things work, sometimes theydon’t, always work, and when and and so forthso i think minsky gets a raw deal in historyon. That one, i think, if you go back and readcarefully what he said, that it was actually stilltrue and it was you know, left to other peopleto figure out what you could do in the absenceof conversions. So my question regarding minskyi totally agree with you, because right now we westill don’t know why neural network work theway they work and we have like expandable aiand when or and when. But my point is that minskybasically said uh just like i think 15 yearsbefore uh the invention of the backpropagationalgorithm, that we don’t have. I a i don’t know. Ifthis algorithm will converge and b it will takemany many. It will take for for days to train oreven. For months and but my question, is it seemsthat, you say something completely different, yousay, something fundamental about the essenceof, artificial intelligence. You say it’s notthe, cpu powers that we are missing, it’s theunderstanding of intelligence. So i think i thinkin a way. Your claim is much stronger claim. Thatwhat minsky said in is perceptrons. Would you agreewell? I think it’s a lot of different claims. Andhe’S. You know he. He made a number of claims. Onhis career that particular book was narrowlyconstructed. It was constructed in a mathematicalway saying we have these great proofs aboutwhat. You can do for linearly separable functionssimple functions and we can’t find that for thenonlinear separable ones. They were actually peopleplaying around with back propagation, around thesame time as wrote that book jurgen schmidt, huberhas written a really detailed history of backpropagation, which has been invented. Many timesincluding in the 1960s when that that bookwas written um, just as a historical asideum, but so that book was narrowly construed, it wascan. We do this mathematical thing and the answerwas no – and they were right about that and thenthere were implications for like how we shouldbuild models and so forth, um he at other timesthought about the nature of intelligence, but thatbook wasn’t really about it. My book sorry notmy book, my my work has in large part been aboutcomparing the human mind to machines you knowa lot of my work for many years was just aboutthe human mind how kids learn language, which is atruly, amazing feat. Questions like that and i wouldsay that the arguments that i have been makingwith respect to ai for the last few decades, havereally been about generalization, um they’ve reallybeen about if you have a certain amount of datais that enough to get you to the rest of whatyou want And the the simplest demonstrationthat i made that i think is really the mostprecious thing that i’ve done is i showed in1998 that you could take a neural networktrain it on a function. The identity functionf of x equals x times one and if you trainedit on even numbers, you wouldn’t generalizeit to the odd numbers and i was accused ofmaking. A terrorist attack on neural networksunpublished review. People hated me forthis, and they thought it was irrelevantbut in the subsequent. Almost 30 years there or 25years um, i think that’s actually become receivedwisdom um and the measure of that is thatyahshua. Bengio has now oriented his wholeresearch program around it, so he talks aboutum. What is his vocabulary? Something like out: ofout of uh training, extrapolation. I forget hisexact formulation, but it’s basically the samething, as i was saying all along um bangio hasrealized, that the achilles heel of these systemsis, they don’t generalize very far. They generalizesome and that’s, been a source of enormousconfusion. So these models do generalize. Some butthey typically have problems with outliers becauseif, you think of like a cloud of points. They’Rereally good inside the cloud of data, pointsthat they’ve got you go outside that cloud of datapoints and they’re just clueless and that’s reallythe achilles heel and that’s why for exampleyou know with all the data that tesla has theystill can’t make a safe driverless car becauseeventually you get To these outlier cases andthat’s a problem with these language worlds, one ofthe problems with these so-called large languagemodels. I always teach my students that thosesystems are very good in interpolation and notextrapolation, because extrapolation means deducesomething beyond what you have seen, and this isvery very hard to do. Even a a for a machines, buthuman beings are notoriously good at this now letme, please go when we start.When we discuss thesubject now we first need to define the termsand the time of what is artificial. Intelligencelet me just tell you that when i startedwriting my artificial intelligence book, i thoughthow you know artificial intelligence aim, toemulate human intelligence, so i might write achapter devoted to human intelligence and thisand. I knew nothing about human intelligenceand iq and all those great research and ii was so fascinated that i ended up. Writing afull a 400 pages book about human intelligenceand. Human intelligence is a very interestingvery sophisticated. Very it’s like the biggestadventure of psychology. In the last one 120 yearsand, let me please show you what patrick winstonis he was a the head of the ai lab in mitwhat. He says about artificial intelligence. Nowthis is the last minute from his ai talk and inthis minute. He said that one student approachedapproached him about a an ai program that cansolve integrals. So please just it’s 40 secondslong time ago i was talking with a student whosaid, computers cannot be intelligent and i saidokay maybe you’re right, but let me show you thisprogram, so i showed at the integration programworking on problems like this and after uh ishowed him. A couple of those examples he says: wellall right, i guess maybe they can be intelligenti’m learning how to do that. It’S not always easythen. I made a fatal mistake. Isaid, let me show you how it worksand we spent the an hour going through it likethis and at the end of that time he turned tome and said i take it back, it’s not intelligentafter all it does integration the same way. I do. I take it back. They’Re notintelligent they do theythey. Does they do exact integrationlike? I do so. If, if the computeracts like you, it is not intelligent, butwhen you’re doing it, you are intelligent, soit’s, not fair. What is the definition of ai inyour perspective? Well, i mean first, i i mean it’s afunny anecdote, but i don’t think per se that it’sright. So you know if you could get machines to doessentially the same things as people do iwould grant them uh the notion of intelligencei think in fact that the notion of intelligenceis um often treated as if it’s a one-dimensionalvariable. You know you have an iq, it’s a number110 or 92 or whatever um, but really there aremany dimensions of intelligence. There are manythings that go into intelligence, so there’sverbal intelligence and symbolic, intelligencemathematical intelligence, physical intelligenceand. Even those are too coarse a categoryso. There are lots of different things, thatdifferent aspects to intelligence and it is hardto define, i think, what’s missing from currentsystems is not kind of brute force. Computationlike alpha fold and alpha go are amazing, examplesof things that you could call intelligent if youwant, but what’s missing, is the ability to look ata problem they’ve, never seen before and related toother problems that you have seen before and comeup with a solution to that. The kind of flexibilityand adaptability is certainly core to intelligenceand. That’S the piece. That’S missing and relatedto that as a kind of verbal intelligence, ofbeing able to relate a sentence to a situationin, the world or to somebody’s beliefs or desiresgoals and so forth, and you know currentsystems just don’t have that at allnow. One thing before we move on to a understandinglanguage, which i think is a fundamental thing. Inyour argument it is, i i was surprised to knowthat, even alphago zero if you change the domainof the of the board, if you just like take onepiece out of the board, it won’t operate wellso, so it even alphago. Zero cannot generallygeneralize from the board of gold to a verytiny modification of goal, which i was mesmerizedby this, but how? How how could it not generalizeit’s just essentially memorizing andinterpolating, there’s no fundamentalabstraction there i mean they build in acouple of things, but there really isn’tthe comprehension. Even of a notion liketerritory, it’s really just a massivenumber of examples and saying you knowin the closest example. What would i doand? That’S just all it is. If you took one stoneaway, i think they’d probably still do fine. I don’tknow a result like that, but if you change theshape of the board, it would have a lot of troublenow again. I want to quote something from your booknow. You say those systems tend to land superficialstatistical correlations and do not somethingprofound. Now there is like a a a very nicememe on the internet, and it goes like this now iguess that you are familiar with it so basicallyso. Basically, you will get like a scrap in thewall and the scrap is like a statistic and noone like it and when you phrase it and when youframe it, so you can call it machine learningand when you call it artificial intelligencebut. Basically, machine learning and artificialintelligence are merely statistical. Functionsstatistical correlations. The best book on aior machine learning is called islrintroduction to statistical learningso. Haven’T we made any progress from, i don’t knowfrom svm and nearest neighbor, or i don’t knowfrom. The work of ronald fisher in the 20s to 2021it is still just a good statistics. Most of it isi. Don’T think that artificial intelligence, andmachine learning are at all the same thing. I thinkmachine learning is a set of tools for artificialintelligence. We do have another set of toolsfor symbolic knowledge, representation for exampletraversing, a taxonomy. It’S not a statistical toolthough. You can add some statistics into it. Umso, we have other tools and we we do things likeplotting the best route from point a topoint b without using statistical, learningagain there’s you know lots of room, forquantitative data and so forth, but that aren’tusing the same kinds of techniques so nobody’srunning their traffic navigation systems, byusing Pure deep learning, for example, you knowthere are places where certain tools are goodand places where those tools are not very good. Itis. True that there’s been a lot of progress, usingthe, deep learning, kind of stuff for particularproblems, please, let’s pause for a second all right, so there’s been a lot of progress, inclassification problems and not that muchprogress using the same techniques for problemslike reasoning and problems like understandinglanguage. We can talk a bunch about gpt3 and whyi think that it’s fundamentally misleading buti, don’t think it’s real progress there, but whenwhen you had to talk with lex friedman about whatai can do, and you said, like: okay, we have like thevision thing. You know from alexnet to imagenet toinception and we know that those vision, systemsthose convolutional neural network can performmuch better than humans, but in your latest inyour latest article, his a i found, a new foundationyou show this thing and you show like iever, get in an apple. Yes, a granny smith, andthe network know it knows: it’s not justan apple, it’s a granite mix and you canclassify this specific apple. But if i take uhi, don’t know a a post-it note and i just sticka post like an ipod, a i just write ipod on theapple and many people say no. This is cyber whatwe are doing, is like cyber attack, but okay, i agreethat this is, can can be considered a cyber attackbut. It’S not a natural attack. It’S it’s! It’S anevaluation of a technology. It doesn’t have to bean attack to figure out hey. What does this systemactually? Do what is it good for and what we see isit’s, not very subtle, it doesn’t have. The abilityto represent a complex hypothesis like an applebehind, a piece of paper called a piece of paperwith the words ipad on it. The system is, is verybrittle now. Are you not an attack, it’s an analysisokay. I totally agree, but i i wanted to ask youthere was like i think, two years ago, a very bigfiasco in google, when google image classifiedone of uh one of google’s friends as a gorillaso, it was a very famous. I i think that you’refamiliar with this it’s more like eight or tenyears ago, and it yeah yeah, so we had the sameproblem earlier this year and then you say: [ __ ], you, google, my friend, is not a gorilla nowwhat. I wanted to say in regarding to this scenariothis fiasco that if you take out all the emotionalthings that we have of classifying people as apesokay and classify – and you know, white people blackpeople the metrics. The objective metric was notfar away. Okay, it it. If you don’t take, if you don’tinsert the weight of of what does it mean tomisclassify, a human being as an ape, so themetric function didn’t fall far behind? But if youshow me an apple and you write on the apple ipodand this and the network cannot deduce. Itthis is an apple, so basically the networkdoesn’t understand what it sees. Andthis is basically your argument: yesthere’s very little comprehension, there’s noability to understand nuance. So when i lookin your room, you know, even though some of it’sout of focus i recognize. There’S some furniturei recognize a person. A microphone i understandthat, you are between the microphone and the let’ssay. It’S the desk back there um, i understand thatyou’re, probably supported by a chair, even though ionly see part of the chair, so making inferencesabout the relations between entities when i seethat picture of an apple with a piece of paperwith, the word ipod in front of it. I don’t justpick the most popular thing in the room. I picki understand as a relation between entitiesand these systems. Just can’t do that now, one theydo is say which pattern in my database is thismost like okay. Now one thing that you said in yourbook, which i tremendously love, is how much priorknowledge do we need when we interpret the worldnow, you gave an example of a a person with athat lost his wallet, so he tapped his pocket soyou to find out whether his wallet Is still thereand, you said, and i was wow you are absolutelyright. There are so much prior knowledge that weput into even a very simple sentence that is notencoded in the sentence itself. So could you pleaseelaborate on this, because this is some somethingthat – is completely like flipped? The way i seethose things, those things need many many pieces ofinformation that they are not being given and weknow them just because we are human being who livein the outer world. I think if you take any scenein any movie that you like, you will see that a lotof the connections are not spelled out, that thedirector is leaving it for you to figure thingsout and you feel smart if you figure them outand. You know the director has some senseof what you will be able to figure. Outbut almost. Never is every little thing. Spelledout. The director is not going to sit thereand, explain to you what a glass is the director isgoing to assume that you understand what a glassis from your prior experience. If the glass fallsthe director is not going to say, hey, there’s sharppieces there, the director is going to assume thatyou can figure that out or same thing. If you reada story, whether it’s fiction or non-fiction, ifthe writer articulated absolutely everything. Thatthey expect that you would know it would beincredibly tedious, but they’re assuming thattheir viewer. Their audience is a human being whoknows how the world works, and you are able to makean interpretation of what’s going on by puttingtogether. Excuse me all of these little piecesthat. You have kind of come across through yourlife, with what’s happening right now in order tobuild. A cognitive model is what i like to callit of the scene. You know what’s happening, nowand form and interpretation, and that’s just notwhat. Current systems are equipped to do. Whenpeople are dead. They know they never going back tolife or when you’re thirsty and you drink you’llprobably be, will be less thirsty. Now these thingsare, so simple so like transparent to us thatwe forget that we need to encode them into thisthose systems, and this is why you said we have nounderstanding whatsoever of language and you toldme that if you ask siri, give me a restaurant whichis, not mcdonald’s, it Will give you mcdonald’sand wow? This is a very simple sentence to graspto. Understand. Give me a restaurant which isnot mcdonald’s, okay, not mcdonald’s, restaurantand. Even this simple sentence is very hardfor. Those system to understand how can it bei mean? I i think, the systems that we have arefundamentally wrong-headed they’re trying toshort-circuit the problem. So it’s it’s not thatthis problem of common sense is unknown mccarthywrote about it. In 1959, minsky talked about itoccasionally doug lennon spent his entire careertrying to distill common sense into machineinterpretable form, but it’s hard work. People don’twant to do the hard work. They want to be able tosubstitute massive data for doing that. Hard workand they’ve confused themselves into thinking thatthat’s possible using systems like gpt3, which getthings right. You know 75 percent of the time, andgive the illusion that they have more going onthan. They actually do so all the clues are therethat. This is not going to work, but people keepgoing after it, because there’s low hangingfruit, they’re very close to running out oflow-hanging fruit in the field has really kind ofchanged. In the last few months, people have startedto realize hey this isn’t working, because i yousaid that you haven’t been giving a permissionto access, gpt suite to check it. It’S not aftera year i mean there’s great irony there rightthere, there’s a company called openai madegpt3 in the presence. George, i think yesno, no he’s not part of it. Um open ai was supposedto kind of like save the world from bad ai bydoing everything open and then they figuredsome stuff out. They’Re like let’s make moneywe’ll still be a non-profit in part and wewon’t be open anymore. But we’ll keep the nameand, we won’t share with scientists. We don’t wantreputable scientists to look at what we’re doingbecause. We know it’s not that great and they theywill call us out on it. So to be better marketingif. We don’t share it with them. We will share whatwe’re. Doing with the media, it’s gullible and notwith people like marcus, who actually understandwhat’s going on, and so it’s been over a year. It’S like it’s like this. You know but you’rea a very unique voice in the world of aii. It’S you and melanie mitchell, which speaks onthose subjects over and over again, but i don’t seeother people in this industry in this fascinatingworld say: listen. I know that benign startslike. Listen. We need to re rethink what what we aredoing, but you are still you and melanie mitchellare unique voices, note well and ernie davis myco-author, who yes, there’s many many pieces, withum me but yeah there aren’t that many ofus. It’S always, you know easier to makea living being an optimist. I think so youknow the books that sell the most on the aion ai are the ones you’re like. Oh we’re on theverge of of you know a miracle a singularityand the public eats that up, but they don’t reallywant to know that hey. This is really hard and it’snot actually going to happen in your lifetimenobody wants to hear that message. You must readthis book guys. I tell you it’s like a a it’s nota life-changing, but it’s like it was basicallylife-changing for me because i need to. I need torethink my entire book. All over again now wheni read your book. I came across immanuel kantthe. Very famous 18th century philosopher whobasically said that when we approach to understandthe world, we are provided like with inherentuproar categories of the mind and the top mostthe top or the most the three most important thethree most important are time, space and causalitynow causality. So, let’s start with time in spaceand move on to causality because causality becausecorrelation is not causation and we need to uh digit through. But what do you mean that we need toequip, those machine learning or iai systems withtime? With the notion of time and space? You have tobe able to ask yourself when you’re thinking aboutsomething, where is it in? You know four dimensionsin time and space and you have to ask yourselfwhen, you see two things that are related to oneanother. Are they actually causing it? You knowis a causing, b or c, causing a and b and theseare really basic to human beings and there’sgood evidence with infants that even human infantsyou know have some recognition of these things. Butthe machines that are popular right now, don’treally have those things i mean. There’S tensiongoing back to minsky in the 1950s for decadesthere’s been two kind of schools of thought in aiand people disrespect each other and don’t wantto borrow from one another, and that’s reallypart of the problem here is that the people doingneural networks, don’t want to take anything fromclassical Computation, the classical computation iswhat makes most of the software in the world workand, so they’re kind of like for political reasonstying,

their hands behind their backs, and then whenthey get a little bit of success. Then they getreally full of themselves and that’s kind of wherewe are right. Now, without kind of looking downthe football field to see where this is all goingand, your argument is that all the big ai systemsfor, just for example, a deep blue and alphago zerothey use they use a fusion of many techniquesdeep neural network and convolutional network andsearch trees. So you cannot solve everything it’slook like it’s, not like a thorough hammer when youcan smash every problem with artificial generalintel, with neural networks or deep neural networksyou need to fuse some of the problemsthat i face with my students or with mycompany, is just like strictly computer visionand, Not every computer vision problem is aisome of the computer vision. Problem can be solvedeven better with traditional approaches. Yes, yeahthere’s a little edit there i mean i would saythe. Traditional approaches are part of ai thatyou know again, ai. Is this much broader yeah yeahyes? Yes, yes, i’m sorry, not everything is out of isdeep neural nets, not everything’s deep neuralnets, and in fact, if you look at some of the mostsuccessful systems, they have deep neural networksas part of what they do, but they have other stuffit’s less sexy people talk about less. So alphafold sorry alphago has tree search, which is aboutas old. You know as classical and ai techniqueas there is. It works with symbols like we’dfind in a classical computer, alpha foldhas, a bunch of stuff, that’s built in aboutthe, very nature of the specific problemof protein folding that it’s working onand. I think we find, over and over again weactually need some of those classical techniquesbut they’re sort of out of favor, so people kindof sweep them under the rug. What we really needis a better science about how to integrate thoseold techniques with the new techniques not justone-off. For protein folding, but in general nowif, i think about it. There is a very nice bookwritten by peter novig called artificialintelligence general approach and it sumsmany of the techniques like bfs dfs, many othertechniques and research and many many techniquesand. I think it doesn’t cover deep neural netsbut. We have many techniques up in our bagbag of tricks that people usually dismiss whenwe speak about ai because deep neural, networkbecame, so sexy and when you want you know to toraise money or to find with money. You need tosay yeah. This is a deep neural network and etc. Sowe just know the old quote from santiana. He who’she you freak. Well, i can’t quite get the wordsbut. He who forgets history is doomed to repeat itand, that’s sort of what’s happening now: right, thewhole, norvig and russell book of techniques, allof those have their place right and – and you havepeople that, like don’t even read the old stuffthey’re, just like i’ve got this new tool and I’Mgoing to use it for everything, so it’s you know toa man who has a hammer. Everything is a nail, a lotof that going on okay now and the other thing thataccount wrote about was causality. Now we alwaysteach our students that correlation isn’tcausation and all these things, but causality isa very slippery thing, because even david yum saidwe, don’t find causality in nature and can’t saywow causality is what within us. We perceive theworld in terms of causality. So but you say inyour book that we have methods we have systemsthat can interpret or can like understandcausality like a. I don’t know like thisthis flow chart, diagram that you have thereso. We have ways to grasp causality. Yes, i think it’s an unsolved problem. I mean you knowone way to think about ai in general. Is it’s sortof like physics before newton there’s a lot ofobservations there’s some techniques, you can useyou know it’s not as if we couldn’t do anyengineering before newton. We can do a bunch ofengineering now, but we’re missing some fundamentalinsights and some of those are around causality sothere’s, some mathematics that people have doesn’tfit that well with deep learning some people aretrying to see how to integrate it um, but there’ssomething, even deeper. That, i think, is missing youknow you and i we look at the world and we thinkof it in terms of causes and effects. We thinkif, i press this pedal. The car is goingto, go forward, it’s not a correlationyou know. I mean it’s a causalcorrelation. It’S not just a correlationif. We try to interpret a medical study, we tryto say you know, did a cause b or is it just acoincidence? Was there maybe a confound andnow? The subjects were chosen. We reason aboutthings. We just don’t really know how to do thatyet to systematically reason about the cause offorces in order to build interpretations of theworld is to build interpretations of the worldyou need to come back to all these things, likesymbols and representations that are out offashion, but i think, still Useful and if you don’thave that it’s hard to see even how to get goingnow in your most recent article youlament on the gap between ai researchersand, cognitive psychologists, linguistic expertand, let me quote so so from your latest uhfrom, your lastest article, you quote someone elsewhile many approaches That have dominated overthe last decade tend to focus mostlyon one end of the spectrum we believethat exploring manners to reach better balancebetween. The two are promising avenue for futureresearch, for example. It seems and then use you sayseems like an acquired recasting of what stephenpinker said in his 1999 bestseller warden wolvesokay. So there is a rich literature in cognitivepsychology, linguistic pinkie rock, but the peopleof ai. I are basically unaware of what other peopledid in other disciplines like people like irisbaron and lisa feldman and other people whostudy. What intelligence is what representationalsystems means? Yes, yeah. I think that there’s anarrogance right now in machine learning, becausethey’ve made a bunch of progress. Um there’san arrogance to think that they don’t need tolisten to other people. So here’s an example: isevery linguist knows that part of what goes onwhen. You interpret a sentence. Is you understandthe syntax of a sentence and then you understandthe semantics, the meaning and then the pragmaticshow? It fits into the world. This is like axiomaticif you’re a linguist, but we have this system gpt-3that’s, very popular that doesn’t do any of thatyou know it actually is able to replicate thestatistics using sorry replicate the syntaxusing, the statistics, but the semantics aren’treally there. So you know, if you ask linguistswhat, the semantics: are they’ll start tellingyou about things like well who’s, the agentin, this sentence: what is the object and youknow what’s happening to that object? You can’task gpt that and it’s not there anywhere andin a rational world. People would say: heywe have these fancy systems but they’renot making contact with the things thatpeople, who study language for a livingtake to be fundamental. Why is that? What canwe do about that? Let me just even get thatconversation going. Let me just give you an anda. Great example. Is that gary gives about howwhat? Does it? What does he mean that by saying gpt3doesn’t understand you give a a fabulous examplewith like an orange juice, he wanted to drinkorange juice, but then you you added a a rewardthat was like a. It was grape juice and cranberryjuice yeah, yeah yeah yeah talking about the sameexample yeah. Yes, could you please repeat, thisexample, because i think that this example nailswhat. You just said those systems don’t have an onand, profound understanding of what is being toldso. So the example is in an article calledgpt3, bloviator um. It was in technologyreview and i can tell you that the wordbloviator was supposed to be [ __ ] artistbecause, that’s what gpt3 is as a [ __ ] artistlike. Somebody comes into a room and pretends toknow about things, but doesn’t really and so tomake that point we had some examples like thisso. What gpt 3 does is it finishes sentences yougive it a bunch of sentences and finish themso. It’S like you’re in a room. You havesome cranberry juice, but there’s notbut you’re really thirsty it’s not enoughto drink, so you add some grape juiceand. Then what happens next basically and thesystem says you drink it, because you know you’vegot this mixture that sort of makes sense butit’s just doing it, because the statistics of drinkhappen to correlate with the rest of the sentenceand, then it says you die well, nobody has ever diedfrom Drinking cranberry juice, mixed together, withgrape juice, in fact, there’s a commercial productin, the us called cran grape juice. That is evenextruded too, and you know you have input of 500billion words of text or whatever it is probablycrayon grape was actually in the input. Corpusbut. The system doesn’t reason it doesn’t likesay i’m going to cross-check in my databaseto, see if anybody’s ever had cranberry juice, andgrape juice before it’s just correlating the wordsdrink and thirsty, and you can’t smell or somethinglike that with death. And so it’s like picking outsome random little bits to come up with itsanswer, but it’s not reasoning about it. It’Svery very different from how a person would wouldlook at the same sentences. It’S like nlp when whenyou teach people about nlp and the vector and andthe vector. So in many cases it’s it actually worksso dog, the vector of the wall dog will be similarto to the vector of of i i don’t know the word catand. The vector of happy will be similarto, the rapture of joyful but again it’sjust mere statistics. It’S not yes. So when iwas in grad school, a guy named tom landauerpredicted a lot of or foreshadowed in a waya lot of, what’s happened and i think made amistake that i saw then and i’ve seen repeatedmany times since, which is he showed truthfullythat. You could correlate words with geography. Andso, if you do co-occurrence statistics, um new yorkand boston come up together, pretty often and sodoes, los angeles and san francisco, because likethey’re both on the east coast or they’reboth in california, you could actually do likemulti-dimensional scaling or something like thaton vectors that were made from the Co-Occurrencestatistics and come up with something that looksa little bit like a map. But the question is isa little bit like a map, what you want or do youwant an actual map. So you get something a littlebit like a map, because the way that people usewords is actually correlated loosely with thegeography. But it’s not perfect. So, like new yorkand la come up together a lot because, like youknow actors, go back and forth and musicians goback and forth and they’re both big cities andso forth. So the geography isn’t perfect. So ifyou’re, given the choice between a map and thishalf-assed correlate, you really want to use themap and not the half-assed correlate, and you don’twant to think that, because i have a half-assedcorrelate that i have come up with the solution. Togeography that allows me to not use maps but whatwe are doing now is exactly that, but writ largeyou get correlates. You know it is correlated youdrink things when you’re thirsty, but that doesn’tmean, just because you’ve noticed that correlationyou actually have a representation of beveragesand drinking and hydration and physiology and soforth and people get misled over and over and overand. Then they scratch their heads and they’relike. Why don’t these systems work? Betterwhat? Can i do about it? You can’t do anythingabout it because it’s just statisticsand there is no deeper representation. It’Slike! It’S sorry. The big thing now is that thesesystems all say nasty things right, they’retrained on the nasty ravings of some of theyou know worst people on reddit plus, you know lotsof reasonable stuff too, not everybody’s nasty butif. You train them on nasty stuff. They will repeatthe nasty stuff. So now there’s a whole industryand. How do we make these systems less rude, andimpolite and evil and so forth and the way theywork all these systems? Is they take the nastysystem that is just replicating the garbage thatit sees and they try to filter it, but they can’treally. Do that well and they filter some of thegood stuff, it’s just a mess. It’S like band-aidson top of band-aids on top of the band-aids it’sit’s like what you said at the beginning of ourtalk. It’S like it works 80 percent of the timethanks to correlation, but we don’t know when andwhy will be the missing 20 and those 20 mightbe crucial. And this is why the name of your bookis building, artificial intelligence, we can trustthis, is fun for the 80. But if i want like fullproof, i cannot get the 20 just off balance or justyeah i mean so that was the subtitle of thebook is build. Building machines we can trustand. The title is rebooting. Ai and thepoint is without a reboot, we’re justgoing to keep incrementally changing from 80 to81 percent, but not get really where we need to besometimes things. Look like they’re on theright path and aren’t now reading your bookmade me think, or we think about john johnsonargument about the chinese room. So what doyou expect, because, basically, what you’re sayingright now guys? There is no understanding, likehuman level understanding or i don’t know how howyou want to call it transcendental understandingor. But what do you expect those system to haveor? How do you imagine an artificial generalintelligence system with understanding would yousay that such a system can be operated on like amacbook m1 or we need a quantum computing poweror, something that is profoundly different inhardware? Maybe we need something? Maybe they couldthe current computer hardware, doesn’t supportthe concept of human un understanding? Well, i meani don’t know, maybe you need you know four m1shooked up or something like that, but i i don’tthink that the hardware is the problem. I thinkit’s the software. We don’t know how to leveragethe hardware that we’ve got to make systems thatare, efficient learners and good reasonersand so forth. I don’t think it’s impossibleyou know. We right now have a regime where thequality of the intelligence is very proportionalto. The amount of data and the amount of computei mean it’s not very good intelligence anyway. Butthe extent that it does stuffyou need a lot of compute butnew humans are a different paradigm. You can takehumans that don’t have a lot of experience withthe world, but still have pretty deep comprehensionlike. My seven and eight year old, kids, don’t havean enormous amount of experience because they’reonly seven eight years old, but they completelyunderstand the physical world around themselvesthat, doesn’t mean they understand. Quantummechanics, but if they run through a playgroundthey understand, you know the relation between allthe entities and themselves and how their bodiesfit, in with these kinds of things, much better thanthe robot of boston dynamics. This is basicallywhat’s much much better. All the boston dynamicsthings that you see are you know carefullyrehearsed like when they do parkour they’vetried it 21 times. You see one successful iterationand. They have like special marks on the floor. Sothe robot needs to know where it starts. You knowwhen, my kids go to the playground. They can dotheir parkour kind of thing on a new playgroundthat they’ve, never been to when it’s a little bitwet and the light is different from it was beforethey have much richer. More general representationsof physical objects causality their bodiesthat allow themselves to be vastly more generalthe same thing. Is true for language? True, like youknow, i can talk to siri and say: hey siri turn onthe lights, but i can say essentially anythingto my kids and if it doesn’t require likeyou, know, college level, math or whatever they’llunderstand right and understand at the level likethey could paraphrase it. They can act on it. Theycan ask questions about why and fill in thingsthat. They don’t know, given those answers. Andyou know put all this stuff together, there’sno machine right now. This is very important, thisis very important. Now current uh, nlp or currentlanguage understanding systems. Don’T ask you heyi, didn’t understand this, i didn’t understand, thisplease, repeat, they don’t know what they didn’tunderstand. I think that this is a crucial argumentyeah. They they don’t and so like the the levelwhich these systems can clarify, is pretty crude. Sosiri can now clarify a little bit like if i sayset a timer. They might ask me, like i’m, trying tothink of a good example of what siri actually doesbut. I know if there’s two sets of lights, it mightask me, you know: do you mean the living room? Lightsof the dining room light so there’s a little bitof clarification, but there’s nothing likethe clarification you get with a human. I meanour whole conversation is clarification. Did youreally mean to say this thing in the book can youelaborate like it’s. You know much more abstractlevel of clarification. A lot of conversations arelike that not all of them, but many of them are youknow. You tell me you’re feeling bad and i’m likewell. Could you tell me a little bit more andyou say well, it’s about my relationship and thenyou know what what happened, and so it’s like alot of clarification back and forth. I’M buildinga richer cognitive model of your situationso. I can give you advice, you i mean this ishypothetical. I just saw your wife uh before andshe’s. You seem to be getting along, but let’ssay that you just had a fight, and you know and imight ask you, you know what was the fight abouthave. You thought about her perspective where wasshe coming from she is great. She is great. My wifeis great, my wife is great. My wife is great andnow. You understand what i just said, because i iit’s like i. I am scary that my wife will listento our interview and then say: oh wait, wait there’slike a wrong idea, so i’ve chosen the wrongyes. No, no, let’s say i had a fight with my wifethis is great because in this uh simple chatthat we had. We had many prior knowledge that weknew about relationship husband-wife relationshipthat was not specifically mentioned in thissentence, and you know what i’m talking. Abouti know what what you’re talking about so forexample. I can understand that, since your wife, iswatching, us recording or maybe watching recordingyou would be uncomfortable in speaking fullycandidly right. So i know enough about human naturenot about you in particular yeah, but i know thathuman beings, for example, sometimes share thingswith one of people and not another, and so like ican, make all of the these inferences around it andunderstand why you might be in an uncomfortablesituation while I’M even talking about thisthere is no computer that could like sitthere and make that analysis of these layersof complexity, layers of knowledge, it’s likethis old television, show called get smart theone of the things the character would say. Isi know that you know that. I know that you knowwell, we can reason about that. I know that youknow that i know that you know you know that i’mkidding here and you know that makes it all okayand again. Let me just put it because some of theviewers will uh are like a student of computerscience and again even the most. Like the bestconvolutional neural network inception, four okaythat can classify one 000 different uh categoriesif a child, an in