"the following is a conversation with francois chalet his second time in the podcast he's both a world-class engineer and a philosopher in the realm of deep learning and artificial intelligence this ti..."

"let me say that the serious rigorous scientific study of artificial general intelligence is a rare thing the mainstream machine learning community works on very narrow ai with very narrow benchmarks t..."

"one day if you enjoy this thing subscribe on youtube review it with five stars on apple podcast follow on spotify support on patreon or connect with me on twitter at lex friedman as usual i'll do a fe..."

"and website that gets you speaking in a new language within weeks go to babble.com and use colex to get three months free they offer 14 languages including spanish french italian german and yes russia..."

"and ends with the vodka now the latter part is definitely not endorsed or provided by babel it will probably lose me this sponsorship although it hasn't yet but once you graduate with babel you can ro..."

"hope to have him in this podcast one day neil degrasse tyson on scientific thinking communication neil two will wright creator of simcity and sims on game design carlos santana on guitar carrie caspar..."

"let me mention a surprising fact related to physical money of all the currency in the world roughly eight percent of it is actually physical money the other 92 percent of the money only exists digital..."

"with francois chalet what philosophers thinkers or ideas had a big impact on you growing up and today so one author that had a big impact on me when i read these books as a teenager with jean pierre w..."

"time that had a big impact on me uh and and there was actually a little bit of overlap with john pierre as well and i read it around the same time is jeff hawkins on intelligence which is a classic an..."

"in here i've been around for a long time even before on intelligence i mean they've been around since the 1980s um and yeah that was before deep learning but of course i think these ideas really found..."

"our brains yes the brain is more for sparse access memory so that you can actually retrieve um very precisely like bits of your experience the retrieval aspect you can like introspect you can ask your..."

"memories without language well the interesting thing you mentioned is you can also program the memory you can change it probably with language yeah using language yes well let me ask you a chomsky que..."

"uh inspect it introspect it how do you think about language like we use actually sort of human interpretable language but is there something like a deeper that's closer to like like logical type of st..."

"that or that they can perform and in terms of the in in terms of motions of objects in their environment before they start thinking in terms of words it's amazing to think about that as the building b..."

"i certainly met i certainly visualize mathematical concepts but you mean like in concept space visually you're embedding ideas into some into a three-dimensional space you can explore with your mind e..."

"it's almost like um what's that called when um you do a path planning search from both directions from the start and from the end but and then you find you do like shortest path but like uh you know i..."

"i guess mind map is a way to make connected mess inside your mind to just put it on paper so that you gain more control over it it's a way to organize things on paper and as as kind of like a conseque..."

"connected in a more effective way see but i'm so ocd that you just mentioned intelligence and language emotion i would start becoming paranoid that the categorization isn't perfect like that i'll beco..."

"being wrong if you want to if you want to organize your ideas by writing down what you think which i think is is very effective like how do you know what you think about something if you don't write i..."

"it's it's not going to be purely like a tree for instance yeah it's fascinating to think that if there's something to that about our about the way our mind thinks by the way i just kind of remembered ..."

"because the categories are very you kind of mention intelligence language and motion they're very strong semantic like it feels like the mind map forces you to be semantically clear and specific the b..."

"the brain that the brain is very associative in nature and i also believe that a topological space is a better medium to encode thoughts than a geometric space then so i think what's the difference in..."

"well let me uh let me zoom out for a second uh let's get into your paper on the measure of intelligence that uh did you put on 2019 yes okay yeah november november yeah remember 2018 that was a differ..."

"yes so do you think the the superintendent ais of the future will want to remember us do we remember humans from the past and do you think they would be you know they won't be ashamed of having a biol..."

"within that context there'll be a chapter about these biological systems seems to have a very detailed vision of that feature you should write a sci-fi novel about it i said i'm working i'm working on..."

"yes so you would see intelligence on display for instance whenever you see a human being or you know an ai creature adapt to a new environment that it has not seen before that its creators did not ant..."

"know how efficiently you're able to adapt how efficiently you're able to basically master your environment how efficiently you can acquire new skills and i think there's a there's a big distinction to..."

"from point a to point b but a road building company can take you from you can make a path from anywhere to anywhere else yeah that's beautifully put but it's also to play devil's advocate a little bit..."

"recently in machine learning and people are you know extrapolating from that progress that we're about to solve general intelligence and if you want to be able to evaluate these statements you need to..."

"so at the first level you draw a distinction between two divergent views of intelligence of um as we just talked about intelligence is a collection of tax task specific skills and a general learning a..."

"very long time and um early ai researchers people like marvin minsky for instance they clearly subscribed to this view and they saw they saw the mind as a kind of you know collection of static program..."

"textbooks until the 1980s with the rise of machine learning it's kind of fun to think about that learning was the outcast like the the weird people were learning like the mainstream ai world was um i ..."

"mind so the other view of the mind is the brain as a sort of blank slate right this is a very old idea you find it in john locke's writings this is the tabulata and this is this idea that the mind is ..."

"ai researchers are conceptualizing the mind via a deep learning metaphor like they see the mind as a kind of randomly initialized neural network that starts blank when you're born and then that gets t..."

"whether like like scaling size eventually might lead to incredible results to us mere humans will appear as if it's general i mean if you if you ask people who are seriously thinking about intelligenc..."

"that all you need to do is to scale it up to all the available training data and that's if you look at the the waves that open ai's gpt stream model has made you see echoes of this idea so on that top..."

"language model that's trained in unsupervised way without any fine tuning is able to achieve so what do you make of that what are your thoughts about gbt3 yeah so i think what's interesting about gpg3..."

"right but there's a side to interrupt there's there's a parallels to what you said before which is it's possible to see gpt3 as like the prompts that's given as a kind of sql query into this thing tha..."

"ceiling do you think you'll be able to reason um no that's a bad question uh like what is the ceiling is the better question how well is it going to scale how good is gptn going to be yeah so i believ..."

"the flaws lgbt3 which is that it can generate plausible text but that text is not constrained by anything else other than plausibility so in particular it's not constrained by factualness uh or even c..."

"question right and if you if you ask the same question in two different ways that are basically adversarially engineered to produce certain answers you will get two different answers to contractor ans..."

"who are not familiar and is uh is the idea of being able to convert the internet or be able to attach like semantic meaning to the words on the internet this the sentences the paragraphs to be able to..."

"ever work simply because it would be a lot of work right to make to provide that information in structured form and there is not really any incentive for anyone to provide that work uh so i think the ..."

"if it's presented with a novel uh situation that's the opening question between the illusion of reasoning and actual reasoning yes the power to adapt to something that is genuinely new because the thi..."

"coronavirus yes which is why you know uh a system that's uh you you tell that the system is intelligent when it's capable to adapt so intelligence is gonna require uh some amount of continuous learnin..."

"a driving model it wasn't even l2 right and that's a lot of data that's a lot more data than the the 20 or 30 hours of driving that a human needs to learn to drive given the knowledge they've already ..."

"it's the vice versa so i'll tell you you know the thing is the the amounts of training data you would need to anticipate for pretty much every possible situation you'll encounter in the real world uh ..."

"i mean sure you've seen a lot of cars in your life before you learn to drive but let's say you've learned to drive in silicon valley and now you rent a car in tokyo well now everyone is driving on the..."

"of a test institutions because again there is no task for which skid at that task demonstrates intelligence unless it's a kind of meta task that involves acquiring new skills so i don't think i think ..."

"just that um it strikes me as a fairly inefficient thing to do because you run into this uh this uh scanning issue with diminishing returns whereas if instead you took a more manual engineering approa..."

"of any generality at all again the only possible test of generality in ai would be a test that looks at skill acquisition over unknown tasks but for instance you could take your l5 driver and ask it t..."

"had a limited amount of time their limited amount of energy and of course this started from a different set of priors to solids from uh um you know innate uh human priors um so i think if you want to ..."

"start from the the same set of knowledge priors about the task and you have to control for for experience and that is to say for training data so prior what's priors so prior is whatever information y..."

"the organization of 2d space and the rules of how the the dynamics of the physics of this game in this 2d space yes and the idea that you have what winning is yes exactly so like and all other board g..."

"yes uh so in particular i think if your if your goal is to measure a human-like form of intelligence then you should clearly establish that you want the ai your testing to start from the same set of p..."

"humans what um what do you think is a good test of human intelligence well that's the question that psychometrics is is interested in what is there's an entire subfield of psychology that deals with t..."

"so the ideal goals of the field is you know tests should be be reliable which is a an ocean type reproducibility it should be valid uh meaning that it should actually measure what you say but you say ..."

"creating psychometric tests are very much nighty old i don't think every psychometric test is is really either reliable valid or offer from bias but at least the field is aware of these weaknesses and..."

"in things that sound completely unrelated like writing in english essay for instance and so when you see clusters of correlations uh in in statical statistical terms you would explain them with a late..."

"and they're not wrong and the thing is so personally i'm not interested in psychometrics as a way to characterize one individual person like if if i get your psychometric personality assessment or you..."

"it's like wishful thinking among psychologists but uh the more i learned i realized that there's some i mean i'm not sure what to make about human beings the fact that the jig factor is a thing that t..."

"skill and there are actually different theories of the structure of clinical abilities that just emerge from different statistical analysis of iq test results but they all describe a hierarchy with a ..."

"bottom of the ocean and so on and if you were a scientist say you want you wanted to precisely define and measure physical fitness in humans then you would come up with a battery uh of tests uh like y..."

"and as you just said that doesn't mean that people with a with high physical fitness can fly doesn't mean uh human morphology and human physiology is universal it's actually super specialized we can o..."

"similar to that our cognitive abilities have a degree of generality that goes far beyond what the mind was initially supposed to do which is why we can you know play music and write novels and and go ..."

"as you were being demonstrated at training time a difficult task starts to require some amount of uh test time adaptation some amount of improvisation right so uh consider i don't know you're teaching..."

"with respect to what the test taker already knows and with respect to the information that the test taker has about the task so that's what i mean by controlling for priors what you the information yo..."

"not obvious it's not because the ambiguity like it's uh and you have to look to them but like some number sequences and so on it's not completely clear so like you can get a sense but there's like som..."

"think an ideal uh test of human and machine intelligence is a test that is uh actionable uh that highlights uh the need for progress and that highlights the direction in which you should be making pro..."

"otherwise what you're doing is not exhibiting intelligence what you're doing is just retrieving uh what you've been exposed before it's it's the same thing as deep learning model if you train a deep l..."

"brute force the space of possible questions right to pre-generate every possible question and the answer so it should be tasks that cannot be anticipated not just by the system itself but by the creat..."

"part of the work you did with uh and i guess continue to do uh probably with the with the paper and the arc challenge can you talk about some of the priors that we're talking about here yes so a resea..."

"because really hard wire to acquire them so a bunch of points pixels that move together on objects are partly the same object yes i mean i mean that like i don't i don't smoke weed but if i did that's..."

"understand was this uh idea of abstraction yes so clarifying uh my own ideas about abstraction by forcing myself to produce tasks that would require uh the ability to produce that form of abstraction ..."

"so how did you come up um if you can reveal uh with the questions like what's the process of the questions was it mostly you yeah that came up with the questions what uh how difficult is it to come up..."

"original new ideas um did psychedelics help you at all i'm just gonna but i mean that's fascinating to think about like so you would be like walking or something like that are you constantly thinking ..."

"is hidden yes absolutely it is impressive that uh the test that you're using to actually benchmark algorithms is not accessible to the people developing these algorithms because otherwise what's going..."

"encouraging yeah i'm fascinated by there's a world of people who create iq questions i think i think that's a cool uh it's a cool activity for machines that for humans and people humans are themselves..."

"i i think it's due to the fact that the set of core knowledge priors that arc is built upon is so small it's all a recombination of a very very small set of assumptions okay so what's the future of ar..."

"so one thing i think we did well uh with arc is that it's proving to be a very uh actionable test in the sense that uh machine performance and arcs started at very much zero initially while you know h..."

"and so and i i do believe that you know it will it will take a while uh until we reach a human parity on ark and that by the time we have human party we will have ai systems that are probably pretty c..."

"so i i don't actually think humans are anywhere near the uh optimal intelligence bound if there is such a thing so i think for humans or in general in general i think it's quite likely you know that t..."

"and that system includes non-human non-humans first of all includes all the other biological systems which are probably contributing to the overall intelligence of the organism and then yeah computers..."

"itself or any of its variants for two reasons so first of all it's um it's really copping out of trying to define and measure intelligence because it's entirely outsourcing that to a panel of human ju..."

"for something very much like the human mind indistinguishable from the human mind to be encoded in ensuring machine and at the time that was that was you know um a very daring idea it was stretching c..."

"whole in practice when it turns out that it's very easy to deceive people in the same way that you know you can you can do magic in vegas you can actually very easily convince people that they're talk..."

"but i still think it's really difficult to convince uh if you do like the alexa prize formulation where you know you talk for an hour like there's formulations of the test you can create where it's ve..."

"know i think being very human-like being indistinguishable from a human is actually the very last step in the creation of machine intelligence that the first ais that will show strong generalization u..."

"you rely on human judges it's not something reliable because the images may not may not so you don't like human judges basically yes and i think so i i love the idea of interactivity um i initially wa..."

"difficult but yes but uh i mean what you're doing what you've done with the arc challenge is is uh brilliant i'm also not i'm surprised that it's not more popular but i think it's picking up what does..."

"as a tool to promote simplicity and efficiency in our models but they are not perfectly compressed because they need to include things that are seemingly useless today that have seemingly been useless..."

"well actually probably not which is why if you're a smart investor you're not you're not just going to be following the strategy that corresponds to compression of the past uh you're going to be follo..."

"especially hedging so uh cognition leverages compression as a tool to promote uh uh to promote efficiency right and so in that sense in our models it's like einstein said make it simpler but not howev..."

"but you need diversity and the reason you need diversity because fundamentally you don't know what you're doing and the same is true of the human mind is that it needs to to behave appropriately in th..."

"so you know the one thing that's uh very important uh in understanding who we are is that everything that makes up uh that makes up ourselves it makes up we are even even your most personal thoughts i..."

"and i i and that that's in a way this is this is our path to immortality is that as we contribute things to culture culture in turn in turn becomes future humans and we keep influencing people you kno..."