The following is a conversation with of bothy. Previously, the director ray I at tesla and before that at Albany I and stafford he is one of the greatest scientist engineers and educators in the history of artificial intelligence and now a quickie visa. Mention of sponsor check them out in a description is the best way to support this park ass. We got eight sleep for naps, better help for mental health fund rise for real estate, investing and But the greens for nutrition choose wisely my friends and now onto the full I'd reads as always know as in the middle. I try to make this interesting, but if you get them, but he still check out our sponsors, I enjoy their stuff. Maybe you will do this episode, a sponsored by eight sleep as new pod? Three mattress, I'm recording this in a hotel. In fact, given

Some complexities of my life is in the middle of the night, for I am I'm seeing an empty hot. a room yelling at a microphone. This, my friends, is my life. I do usually feel good about myself at four am, but not with two cups of coffee and me, and the reason I feel good is cause I'm going to go to sleep soon and have accomplished a lot this is true today, except for the sleep soon part, because I think I'm gonna airports and point soon. It doesn't matter what matters is in his sleep here and that's great because well, I don't have an eight sleep bed. They can call itself at home. I do that's wrong. Headed I'm headed home.

Anyway, check it out and get special savings when you go to a sleep, dot com, slash lex, this episode is also brought to you by butter. Help spelled h, e l p help I'm a huge fan of talk therapy, I think of podcasting as a kind of talk therapy, so I'm a huge fan of listening to podcasts In fact, that's how I think of doing a park ass myself, I just get to have front row seats to thing I love and is actually just the process of tar king. They reveal something about the mind and he never good talk therapy is. Is it guided by professional therapists? It helps you reveals year saw something by your mind. just play it all out on the table so yeah

definitely use the best method of talk therapy, the best meaning the most accessible at least to try, if not to make it a regular party of your life. That's what better! How does check them out about how darker slash lux can save on your first month the the episode, is also brought to you by fund rise, spelled f: u n d r! I s e! It's a platform that allows you to invest in private real estate will live in hard times, folks for many different reasons, but one of them was financial and one way to protect yourself. Endeavor times is diversify your investments. Private real estate is one of the things I believe you should diversify into and when you do, he should use tools that look like a major twenty first century which allow investment, even

on online investment websites and services seem to be designed. By the same people, the design, never joy items, that's not the case with fun rise, super easy to use accessible over a hundred. Fifty thousand investors use it their team. vets and manages all their real estate projects. You can track your performers in their website and see, updates as properties across the country are acquired, improved and up if, anyway, check out fun dries. It takes just a few minutes to get started at fund rise, dot, com, slash lex. The show is brought to you by athletic greens and it's aging one drink, which is an all one daily: drink, disappoint, better health and peak performance. I have to be honest. I can. clearly forgot to bring a fight agrees with me as I am travelling now, and I miss it. It's not just

if my nutritional base and needs it's good for myself part of this serve the daily habits of life and we need we have that habit. The routine stuff is off. So it is good to just put that into your daily routine. To make sure you're. Getting the vitamins thing. You attrition that you need no matter the dietary. the workload the athletic endeavors, the? U partaken, kind of incredible than the others.

The greens is for me to give you one month, supply of fish or when you sign up at a third agrees that calm slash lacks this. Is the legs treatment part guessed the support it we check on our sponsors and now, dear friends, here's andrey capacity, what is in your own network, and why does it seem to do so? Such a surprise, me good job of learning. What is a neural network, its hey mathematical, abstraction of the brain, I would say that I it was originally developed at the end of the day, it's a mathematical expression and severely

simple mathematical expression when you get down to it, it's basically a sequence of metres multiplies which are really not products mouth equally and some knowing your teeth thrown in, and so it's a very simple mathematical expression and it's got knobs in it. Many knobs many knobs and these knobs are loosely related to business and options in your brain, their training, their modifiable, and so the idea is like we need to find this. The knobs that makes the neural net do whatever you wanted to do like classify images, that's on, and so there's not too much I was saying that, like you might think that basically don't want to endowed with too much meaning with respect to the brain and how it works as real, just a complicated mathematical expression with knobs and does not need a proper setting for it to do something desirable yeah, but poetry is just the collection of letters was spaces, but it can make us feel a certain way in that same way, when you get a large number of knobs together with whether it's in or inside the brain,

in or inside a computer. They seem to. They seem to surprise us with the weather power yeah. I think that's fair, so Actually I'm underselling it by a lot because he definitely do get very surprising merchant behaviour sort of these neurons, their large enough and trained on complicated enough problems like say, for example, the next hour work protection in a massive dataset from the and then these narrow must stay. Calm, priest, surprising, magical properties, ya think it's kind of interesting how much you can get out of even very simple mathematical formalism when your brain right now is talking. Is it doing next? Were prediction there is something more interesting will definitely some. I have a generated model, that's a djibouti like and prompted by you. Yes, they are given prompt and I'm kind of like writing to it. In a general way, and by yourself perhaps little bit like hey. Are you adding extra problems from your own memory inside your head?

I know monthly feels like your reference in some kind of a declarative structure of like memory, and so on, and there you, that, together with your prompt and giving away some, how much of what you just said has been said by you before nothing, basically right, no. But if you actually look at all the words you ever said in your life and do a search, you probably said a lot of the same words in the same order before you could be I mean I'm using phrases that are common, etc, but I'm remixing it into a pretty I sort of unique centered at the end of the day, but you're right. Definitely, there's like a ton of remixing. Why you didn't you the man was said. I'm am point now whatever which is pretty decent? I think you're talking Very you not giving of credits in your nets here why do they seem to work which you best intuition about?

this emerging behavior. I must kind of interesting because I'm seventeen, sleep under selling them, but also you like, there's an element to which I'm over, like it's actually kind of incredible that it's so much emergent, magical behavior out of them, despite them being so simple mathematically, so I think those are kind of like two surprising statements that can adjust the juxtaposed together and I think, basically, what it is is we are actually fairly good at optimizing, these neural nuts and when you give them a hard enough problem, they are forced to learn very interesting solutions in the optimization and solution basically have these emerging. These are very interesting. Theirs. Wisdom and knowledge in the knobs and something yes this rapidly. Patients in the knobs does it make sense to doing. With a large number of knobs can hold a representation a capture. Some deep was about the data? It has looked at falada knobs, it's a lot of knobs, somehow inner circle. Speaking concretely,

one of the new illness that people are very sad about. Right now are our duties, which are basically just next word prediction networks You consume a sequence of words from the internet and try to protect next word and. Was he train these on a large enough data set? They? You can basically prompt these neurons, an arbitrary ways. Then you can ask them to solve problems and they will say, can just don't tell them you can you can make it look like you're trying to solve them I have a mathematical problem and they will continue to do what they think is the solution based on what they've seen on the internet and very often those just look very remarkably consistent, the correct potentially me. Do you still think about the brain? side of it. So as new nets is an abstraction mathematic, obstruction of the brain is still wisdom. Far from the bible. logical, your networks or even the bigger question. You big fan of biology, biological computation. What

impressive. Thing is biology, do doing to you. Computers are not yet the gap. I will soon differently on much more hesitant with the analogies to the brain, then I think you would see potential in the field and- and I kind of feel like certainly the way neural network started- is everything stemmed from inspiration of the brain by the brain, but at the end of the day, the artifacts that get after training. They are arrived at by very different, optimization process. Then the optimization process that gave rise to the brain, and so I think, I kind of think of it as a very complicated alien artifact at something different As necessary, the neural net that were training oak, the complicated alien artifact, I do not make analogous to the brain because I think the optimization process that gave rise to it is very different from the brain. So there was no more agent self play kind of set up. An evolution it was then accusation. That is basically a what amounts to a compression objective.

on a massive amount of data. Ok, vision, your networks are doing compression and logical, new networks in order to survive. I really doing eighth there, an agent in a multi agent self place system, has been running very very last time. That said, evolution has found that it is very useful to to predict and have a protective model in the brain, and so I think, our brain utilizes, something that looks like that us as a part of it, but a lot more widgets gizmos and functions in ancient. nuclear, they are all trying to like make a survivor, reproduce and everything else, and the whole thing through embryo genesis is built from a single cell is just the code is inside. The dna edge is just built up like the entire organism, with armed, crazy, ahead head and legs. Yes and like it does a pretty well

it may not be possible so there's some learning going on there. Some there's some there's some kind of competition going through that building process at an where. If you're just to look at the entirety of history life on earth? Where do you think is the most interesting invention is the origin of life itself. Is it just jumping dewey you carry, it is it mammals is a humorous himself and say: is that the origin of intelligence or highly complex intelligence or where what where's it all just a continuation of the same kind of process will certainly, I was it's an extremely remarkable story that I'm only like briefly learning about recently and all the way from actually like. You almost have to start at the formation of and all of its conditions and the entire solar system and how everything is arranged with jupiter and me, and the habitable zone and everything? And then you have an active earth, thus turning over mature

up and down, and then you start with a b abiogenesis and everything. I saw it all like a pretty remarkable story. I'm not sure that I can pick like a single, unique piece of it that I find it It's interesting, and I guess for me as an artificial intelligence researcher. It's probably the last piece. We have lots of animals that are, you know, are, are not building technological society, but we do, and it seems to have happened. quickly. I think to have happened very recently, and on something very, listing happened there, but I don't really understand. I almost understand everything else can, I think, intuitively I dont, understand exactly that part and how quickly was both explanations will be interesting. One is that this is just a continuation of the same kind of process, there's nothing special by humans, That would be deeply understanding. That will be very interesting that we think first ourselves a special, but it was obvious. While it was already written in in the code

there. You have been great Greater intelligence emerging in the other, explanation, which is something truly special happened. Something like a rare event or their site crazy rare event like us precisely what would it be see if you like the invention of fire or the As richard rang him says, the beta males deciding a clever. Did you kill the alpha males by Calabria so just optimizing the climate and really the multi agent aspect to the multi agent and that really in constrained on resources and trying to survive the elaboration aspects were created. The complex intelligence, but it because the natural I'll go to the evolution press. What could possibly be a magical thing that happened like a rare thing, there would say that here

these are. Actually he will. Love and tat is actually really rare thing in the universe isn't to say that it is for her away, but it definitely seems like it's come look upon, traded equilibrium where you have lots of duration and then you have certainly leaps, sparse leaps and between so of course, like origin of life would be one in a dna sex. You carry out existed attica life and the and the symbiosis event where the arc jan eight, the bacteria you know just the whole thing, of course, emergence of consciousness and so on. So it seems like different than there sports events where mass amount of progress was made, but he it's kind of hard to pick. One. Do you think human are unique that I ask you. How me intelligent civilizations, do you think are out there, and is doing agents, different

similar to ours. Yeah I've been preoccupied with this question. Quite a bit recently are missing from the free paradox and just thinking through and and the reason actually that I am very interested in, The origin of life is fundamentally trying to understand how common it is that their arctic logical societies out there and his pace and the more I study it the more I I think that there should be quite a few quite a lot. Why haven't we heard from them, because I I agree with you feels like I just don't see why a the what we did here on earth is so difficult to do. Yeah, and especially when you get into the details of it, I used to think origin of life was very, and it was this magical rare event, but then you,

it looks like, for example, Mclean and the vital question a life ascending, etc, and he really gets in. Then he really makes you believe that this is not that rare basic chemistry. You have an active earth and you have your alkaline vents. and you have lots of alkaline waters, mixing whether the devotion- and you have your pretty. Radiance and you have the little poorest pockets of these. I climbed thence that concentrate chemistry and basically, as he stepped through all of these low pieces. You start to understand that. Actually, this is not that's crazy. You could see this happen on other systems and he really takes you from just a geology too primitive life and he makes it feel, like is actually very plausible, and also like the origin of life and didn't

was actually fairly fast after formation of earth and farmer correctly. Just a few hundred million years, or something like that, after basically, when it was possible, life actually arose, and so that makes me feel that that is not the constraint that is not the limiting variable and that life should actually be fairly common and, and then you know where the drop offs are is very, is very interesting to think about. We are currently think that there is no major drop us basically, and so there should be quite a lot of life and basically what the word that brings me to. That is the only way conceal the fact that we have found that one and so on is that we just can't we can see them. We can observe them. Just A brief comment, MC lane, a lotta biologist, I talk to you They really seem to think that the jump from bacteria to more complex organisms is the hardest jumble. Do you carry decline? Yes, yeah, which I don't I get it there much.

More knowledgeable than me about like the intricacies of biology, but this seems like crazy guess how much, how much these single organisms, are there like an hum time? You have surely it's not that difficult and a billion years nine. That long of time, really all these bacteria under control. Resources battling it out. I'm sure they can invent more complex, gotta move from a hollow world programme to like that, like invent a function or something like that, so I don't yes, I'm with you. I just feel I cannot see any if the origin life now be my intuition the hardest thing, but if that's not the hardest things that happened so quickly, then, be everywhere and yeah, maybe we're just too dumb to see. It was just that we don't have really good mechanisms for seeing this life. I mean by what how come some

an expert just to preface this, but just from that it was. I want to meet an expert or an alien intelligence, how to communicate I'm very suspicious of our ability to to find these intelligences out there and to find these earth like a radio For example, are are terrible, their power drops off, as has basically one over our square also, I remember, reading that our current radio waves, but not be the ones that we we broadcasting, would not be measurable by our devices today, only like wasn't like one tenth of a light years away like not even busily tiny distance, because are you really need like a targeted transmission of massive power directed somewhere for this to be picked up on large long distances? Just think that our ability to measure is is not amazing. I think of the other civilisations out there, and then the big question is don't they both by improves and why don't they interested travel across the entire coxey and michael answer is it's probably interstellar. Travel is not really hard

interstellar medium. If you want to move at close the speed of light you're going to be countries bullets along the way, because even like tiny Hydrogen atoms and low particles of dust are basically have massive kinetic energy at those speeds. so basically you need some kind of shielding. You need to have a look. ethnic radiation is just like brutal out there. It's really hard, and so my thinking, maybe indra sold. Travellers is extremely hard, billions of years to build hard feels like it feels like. a billion years away from doing it just might be that its very go very slowly, potentially as an example through space right, as opposed to a close, speed of some suspicious, basically of our ability to measure life and I'm suspicious of the ability to permeate all of space in the galaxy or across galaxies, and that's the only way that I can certainly I can currently see a way around it there. kind of my employed. You think that there's trillions of intelligent, alien civilizations out there

Car slowly, travelling through space will make we're each other, and some other me so I'm gonna war, some of them collaborate more there also independent. They are all just like little pockets for statistically, if there's like, if it is trillions of them, surely some of them some of the pockets are close enough to get some. happened to me closer in close enough to see each other, and I want to see once you see something that is definitely complex. Life like if we see something we're probably going to be severe, like intensely aggressively motivated to fill up with the hell that is and try to meet them all. Be your first instinct didn t try to like at a generation level we meet them or defend against them, or will be around instinct as up

in the united states and the scientist dunno, which hat you prefer in this question. Yeah. I think the the question is really hard and I will say like, for example, for us and we have lots of primitive life, forms on earth palm Thus we have all kinds of anson everything else and wash airspace with them and that we are hesitant to impact on them and to our has returned to protect them by default as they are amazing, interesting, dynamical systems that took a long time to evolve and they are interesting and special, and I don't know that you want to destroy that by default, and so I like complex dynamical systems that took a lot of time to evolve. I think I'd like to I'd like to preserve it. If I can afford to, and I'd like to think that the same would be true about the galactic resources and that

they would think that work and of incredible interesting story that took time he took a few billion years to unravel and you dont want to destroy it. I could see to aliens talking about earth right now and saying I'm a big fan of complex dynamical system, so I think it is the value to preserve these and who, basically, our video game they watch or show a tv show that they watch ya. Think You mean, like a very good reason. I think, to destroy it like why don't we destroy? and farms, and so on is because we're not actually really in direct competition with them right now. We do it it and so on, but on there's plenty of resources. So why would you destroy something that is so interesting and precious oil from a scientific perspective you might probit interact with it lets you might want to learn something from it, so wonder there could be certain physical phenomena that we think is a physical phenomena, but it's actually interacting with us to like pop the thing see what happens. I think should be very interesting to see.

It is other ailing scientists. What happened here- and you know it seen today as a snapshot. Basically, basically, it's a result of a huge amount of computation ah of loverlike over years, or something like that so but have been initiated by aliens. This could Computer running a program like winch. Ok, if he had the power to do this when you ok for sure least, I would. I would pick earth like planet there. as the condition base. My understanding of the chemistry prerequisites for life and I would seated with life and run it. They will win you one hundred percent, do that and observe it and they protect. I mean that this register A good tv show its scientific experiment, and that is its physical simulation right, what made maybe the evolution is the most like she running, it is the most efficient way

you understand, computation or computer for the interest and life, or you know what life looks like and what branches it can take with us. Making a careful where there were part of a science experiment, but neighbours Things are science, experiment sadder that change anything for us before science experiment, the two descendants of apes talking about being inside of the sites, but I'm suspicious of this idea like a deliberate pence per year as you describe it service and I dont see aid. Fine intervention in some way in the in the historical record. Right now, I do feel like the story in these in these books, like place, bucks and so on. So it makes sense and it makes sense had life arose on earth uniquely and ya. Don't you I think that we need to reach for more exotic explanations, right now sure but NBC said a video game dolls. Observe any divine intervention either we might he'll be often pcs or any kind of code. Maybe eventually they welcome

the emphases are really dumb, but on their running gb tease. Maybe they will be like this is the really suspicious what the hell, So you are famously tweeted. It looks like if you bompard earth with photos for a while. You can. a roadster. So if I can hitchhiked guy to the galaxy. We would summarize the story of earth, so it in that book is mostly harmless. Ah woody things all possible toys, like a paragraph, long sends long that earth can be summarized as at once it done its computation. So, like all the possible fool, if earth is a book right now, probably there has to be an ending, I mean there's going to be an end to earth and we could end and all kinds of ways you can and soon again and later, what do you think other possibles stories, or definitely there seems to be yeah your surf it's pretty incredible at these self replicating systems will basically around

from the dynamics and then the perpetual seldom become more complex and eventually become conscious and build a society that kind of you like it. In some sense, it's kind of like a deterministic wave that you know the kind of just like happened, some any sufficiently well arranged system like earth, and I can feel it deserves sense of endeavour inevitability in it and is really beautiful, and somehow right so sir. It's chemically diverse environment, where a complex dynamical systems can evaluating become more and more for them for the complex but then there is a certain: what is it there certain terminating distance? the item of determining conditions are but differently, there's a trend line of something and were part of that story and like. Where does that? Where does it go so you know famously described, often as biological bootloader for a ice. Now, because

and I mean that, where an incredible biological system full of computation enough and so on, but our Women inefficient as well like we're talking to each other through audio, it's just kind of embarrassing. Honestly, there were manipulating, like so Symbols: I serially were using vocal cords. It's all happening, overlap multiple seconds It's just like kind of embarrassing when you step down to the frequencies at which computers operate or are able to cooperate on, and so basically it does seem like synthetic intelligences are kind of like the next stage of development, and I dunno where it leads to like Some point, I suspect the universe is some kind of a puzzle and these synthetic eyes will uncover that puzzle and dumb solve it and then what happens after eight like would cause if you just like fast

earth many buildings years thick, it's quiet and there's like Normally you see like city lights and stuff like that, and then what happens at like at the end like? Is it like a? Is it like a calming? Is it explosion? Is it like earth like open, like a giant because he said, or there's a new start emitting like giants number, satellites, yes to some kind of a crisis, lotion and we're living we're like we're stepping through eight explosion. and we're like living day to day, and it doesn't look like it, but it's actually. If you, I saw a very cool animation of earth and life on earth and basically It happens for a long time and then the last like two seconds, like basically cities and everything and and the low earth orbit just gets cluttered and just the whole thing happens in the last few seconds when you're like this is exploding the state of the explosion. Yeah yeah, you played a normal speed, is ill. Just look like an explosion.

Firecracker, we're living in a firecracker waits going to start a mini all kinds of interesting things, and then the explosion doesn't You look like a little explosion with wood lights and fire energy emitted, all that kind of stuff. But when you look inside the details, the explosion theirs sure complexity, happening, where there's like yeah human life or some kind of life. We hope it's not destructive firecracker. It's kind of like a constructive firecracker eyes or given that they have a think hilarious disguise. It is a really interesting thing about like what the What university did the creator of the universe give us a message affairs upon the book contact Carl Sagan, there's a message for humanity for then he was issued in the digits and the expansion of pie. And base eleven eventually, this kind of interesting thought, maybe maybe we're supposed to be giving a message to our creator

maybe we're supposed to somehow create some kind of the quantum mechanical system that allows them to our intention presence here, because, if you think about it from that perspective, it just sailor. Quantum field theory massive like slowly. Autonomy like thing and like how even notice that we exist. You might not even be able to pick us up in that simulation, and so how do you have you prove that, just now that your intelligent in that part of the universe. So this is a good turing test for intelligence from earth. Yeah I got a creator is maybe this is like trying to complete them. Wars in the sense that the complicated way of that by the earth is just is basically sending a message back yet a puzzle. Basically like alerting the creator that we exist or the puzzle is just too just break out of the system and just a you know, ass ticket to the creator in some way. Basically, like. If you play a video game, you can, you can somehow find and exploit and find a way to execute on the hosts machine. and arbitrary code. Their summer present, like

If someone got a mario, a game of mario to play pong just by exploiting it and then creating a mystical, writing writing code and being able to execute arbiter. Code in the game, and so maybe we should be. Maybe that's the puzzle is that we should be on my way to export it, so So I think, like somebody, synthetic gas will eventually find the universe to be some kind of a puzzle and then solve it in some way, and that's kind of like these, game, somehow do often think about It is as a simulation the soul the universe being kind of computation that has might have bugs and exploits yes, yeah. I think so that would have his exists is essential. I think it's possible that physics has exploits and we should be to find him arranging some kind of a crazy quantum mechanical system that somehow get sued buffer overflow somehow gives you a error in the floating point Yeah, that's right, like more more sophisticated exploit that those jobs, but that could there

colleagues, I find some way to attract definite energy, for example, when you wayne, reinforcement, learning agents in physical simulations and you ask them to say, run quickly on the flat ground, they'll end up doing all kinds of like weird things in part of that optimization ride, they'll get on their back leg and they'll will slide across the floor, and it's because of the optimization and the impossible. Learning optimization on that agent has figured out a way to extract infinite energy from the friction. Forces and basically their poor implementation, and they found a way to generate if an engine just light across the surface. And it's not what you expect. It is just a sort of like a proper solution. And so maybe we can find something like that. Maybe we can be that little dog It's through the cracks or escapes the intended consequences of that in that year first came up. We will forget some kind of shock at the sum weirdness young and to see that, That weirdness is the first point to discover the weirdness like slide.

In the back legs. That's all we gonna do very quickly become everybody. Does that thing? So I gather that the paperclip maximizing, isn't it good idea, but that very well could be what and we'll just we'll just all switch that because it's so fun while no person will discover it. I think, by the way I think it's going to have to be some kind of a super intelligent ai of a third generation like we're building the first generation eg a new There were generation, yes all the bootloader ferny I there, that I will be a bullet. for another hour. I am now there's no way for us to enter speck like what now, I think is very likely. These things, for example, like say of these Agee ice, is very like the first but they will be completely inert. I like these kinds of life I books, sometimes where these things are,

completely inert. They don't track, but anything, and I find that kind of beautiful, because they probably they ve, probably feared automatic Let a game of the universe in some way, potentially they're they're doing something We could be on our imagination and the don't interact with simple chemical lifeforms like. Why would you do that? So I find those kinds of ideas compelling was their source of fun? What did it? They do well to strongly pleasure solving in the universe, but not as a kid define what it means. A nerd today, Keep the iraqi near his nervous us Then they will behave in some very strange way too because they're there beyond their plane them again and the matter games probably say, like arranging quantum mechanical systems in some very we're ways to extract infinite energy. our solved the digital expansion of pi, to whatever amount build their own like flow fusion reactors are something crazy, like you do something here,

comprehension and not understand, or to us and Actually. Brilliant under the head would have quantum mechanics itself. Is the system and were just thinking it's. six by were really parasites and not parasite whenever they hurting physics, I just living on this organisms that this it hasn't been where like trying to understand- But really it is an organism with a deep, deep intelligence. Maybe physics itself is out, the organism, as do the super interesting thing and we're just like one little thing, yeah ants and on top of it, trying to get energy from it. It just kind of like these particles in the wave that I feel like is mostly deterministic and takes our universe from some kind of a big bang to Some kind of a super intelligent, replicator, some kind of a stable point. In the universe. Given these laws of physics, you don't think

as I say said, god doesn't play days, so you think it's mostly deterministic. There's no way listen. I think I think, as a domestic others tons of bombs careful with friendless pseudo random. He I dont like random. I think maybe the law if the physics are deterministic, Their determines just got really uncomfortable with this question. of anxiety about whether the earth his hand and we're not set aside What's there's no randomness, It's easy! You, like goodwill, hunting, it's not your fault under his department. see you don't like randomness, I think, it's unsettling, I think so, deterministic system. I think that things that look random say the collapse of the way function, etc. I think there are determine estate, just entanglement and so on, and some kind of a multi, worse theory, something something ok, so why does it feel like? We have? A free will like

If I raised his hand, I chose to do this now, who, what it doesn't feel like a deterministic thing. It feels like I'm making a choice. It feels like it. Ok, so it's all feelings are just feelings: yeah, so when it and our allegiant is making a choice that this not really make a choice. The choices are already there. Yeah you're interpreting the choice in your creating a narrative for having made it now Talking about the narrative is very matter looking back. What is the most view fuller, surprising idea in deep learning or I in general, the you ve come across. You see in this field explode and grow an interesting ways, just what were quite ideas like the grease major setback and gotten smart, big or small, while the want that I've been thinking about recently, the most probably

is the the transform architecture? So, basically, no others have a lot of architecture that were trendy, have come and gone for different messroom modalities like for vision, audio, you would process them with different. Looking euro months and recently, we ve seen these companies convergence towards one architecture transformer and you can feed it that or you can feed it in an images or speech, extended just gobbles it up, and it's kind of look a bit of a general purpose. Computer there's, also tradable and very efficient to run in our hearts and so this paper came out in twenty sixteen, I want to say attention is all you need. Protection is all you need. You criticised per title in retrospect that it wasn't didn't foresee the bigness of the impact the other was going to have the amateur. Authors were aware of the impact that paper would go on to have probably weren't at, but I think they

we are aware of some of the motivations and design decisions behind the transformer and they chose not to, I think, expand on it in that way in the paper, and so I think they had an idea that there was more than just the surface of just overshooting translation and a better architecture, you're, not getting translation, It's like a really cool depression. Will optimize able efficient computer that you've proposed, and maybe they didn't have all of that foresight, but I think it's really interesting. Isn't it funny sorry to interrupt that? That title is mima? all that they work for such a profound idea. They want with the imo. Genuine user can entitled before I was. I need the eyes like a meme or something you early is not funny. There were beggar. Maybe It was a more serious telling me. I have the impact honestly I

yeah, there's an element of me that honestly agrees with you and prefers it this way as if it was too grand it would over promise and under deliver. Potentially, so you want to just there mean your way to greatness as you'd be a t shirt. So you you tweeted the transformers and magnificent neural network architecture, because it is a general purpose, differential computer. It is simultaneously expressive in the forward pass optimize w via proper backpropagation, gradient descent and efficient, I parallelism compute graph, can you discuss some of those details? Expressive optimize a bull, efficient, yeah from memory or or in general? Whatever comes to your heart, you want to have a general purpose, computer that you can train on arbitrary problems like say task of network connection or detecting. If there's a cat in image or something like that and what to train. This computer say want to set its weight, and I think there's number of

on criteria, that sort of overlap in transformer simultaneously? That made it very successful, and I think the authors where kind of deliberately trying to It's really a powerful architecture and do so in an obviously it's very powerful in the forward pass, because it's able to express a very a general carter computation as a sort of something that looks like message passing notes and they all store vectors and these nodes get to basically look at each other and it's each other's vectors and they get to communicate and basically nose get to broadcast hey, I'm looking for certain things and then other nodes get to broadcast hey. These aren't the things I have. Those are the keys and the values Just attention yeah exactly transformers much more than just the attention component. It's got many pieces architectural that went into it the residual National debates arranged. There's a motorway perception. there, the weights, stacked and so on, but basically the message passing scheme were nodes, get to look at each other, decide what's interesting and then update each other and

I think be when you get to the details of it. I think it's a very expressive function, so it can express lots of different types of algorithms and former pass. Not only that, but the way designed with the residual connections. Lan normalization, the soft max attention. Everything is also optimism, or does it really big deal because there's lots of computers? There are powerful that you can't optimize or not easy to optimize, using the techniques that we have, which is back proposition ingredients, and these are first order methods for a simple optimized really, and Do you also need to be optimized apple and then last He wanted to run efficiently and hardware are hardware. Is a massive throughput machine, much abuse. They prefer lots of parallelism, so you don't want to do lots of sequential operations. You want to do a lot of operations serially and the transformer is designed with that in mind as well, and so is designed for our hardware and is designed to both be very expressive in a forward pass, but also very optimized people in the backward pass, and you said that There is a your connection, support a kind of ability to learn short, algorithms, fast and first

and then gradually extend them longer during training. What's what's the idea of learning short algorithms right, think of it as a so basically, a transformer is a series of blog, right and these blocks have attention and normal to labour, sexual, and so you you go off into a blocking. You come back to this residual pathway then you go off and you come back and then you have a number of layers arranged sequentially, and so do we, look at it, I think, is because of the residual pathway in the backward pass. The gradients also flow along it, uninterrupted because addition distributes the gradient equally to all of its branches. So the gradient from these supervision at the top just floats directly to the our first layer and the audiovisual connections are arranged so that, in the beginning, a during this location, the contribute nothing to the rest of your pathway. So what is kind of looks like as

agile, to transform this kind of like a python function like a deaf and you get to do various kinds of luck. Lines of code say, have a hundred layers deep transformer typically It's matures a twenty six twenty lines of code and you can do something in them and so think of during the optimization. Basically, what it looks like as first you optimize the first line of code and then a second line of code can kick in, and the third line of coke can kick in, and I kind of feel like because of the residual pathway in the dynamics of the optimization. You can sort of learn a very short algorithm. That gets the proximate answer, but then the other layers can sort of kick in and start to create a contribution at the end of a a your optimizing over an algorithm that is twenty lines of code except these lines, we are very complex, because this, an entire block of a transformer can do lot in. There was an interesting. Is that this from architecture actually has been a remark. We recently basically trust whether came out in twenty sixteen as the transformer you would use today, except you reshuffle some delay or norms.

if the player immunizations having to reshuffle to prey norm and formulation- and so it's been remarkably stable, but there's a lot of bells and whistles that people have attached to it and tried to improve it. I do think that basic, meda, it's a big step in some opinions. The up optimizing for lots of properties, of example, neural network architecture, and I think that people have in trying to change it, but it's bruin remarkably resilient, but I do think that there should be even better if this potentially but that's and you're, you admire the resilience here, yeah something profound about this architecture that at least visit he may be, can everything can be turned into our and your problem, the transfers? currently definitely looks like the transformers taken over ai and you can feed basically arbitrary problems into it and it's a a general, the forseeable computer and it's extremely powerful and This convergence in asia has been really interesting to watch for me personally. What else do you think could be discovered here by transformers echoed surprising thing,

or is it a stable arm? I wanna stable There's something interesting might discover by transformers. Aha moments may be as good memory, I may be knowledge representation. That kind of stuff definitely does that. Guys today is just pushing like basically right now. This guys is do not talk to transformer. You touch everything else, so people are scaling up. The data sets making them much much bigger their working on the evaluation, making the evaluation much bigger and are there basically keeping the architecture unchanged and we found the last five years of progress and a high kind of What do you think about one flight? which is language models. Have you been surprised? as yours of imagination and captivated by you mentioned you preteen all the begun began bigger language models,

and what are the limits of those models eating so just the task of national language. Basically, the lgbt strained right as you just dulled mass amount of our text, data from internet and are you trouble Take the next word and the sequence, roughly speaking, are you predictable were chunks, but to ha roughly speaking, that's it. What's been really interesting to watch- is obviously language model. Linkage. Models have actually existed for a very long time, display on language modeling from two thousand free. Even earlier can explain that case. One language model is yes which model just basically, the rough idea is some just pretty the next word in a roughly speaking, ass and there's a paper. From, for example, a banjo and the team from two thousand three, where forty four They were using neural network to take say like three or five words and predict be next word. And they're doing this on much smaller data sets and the neural net

transformer at some, will tell her perception but interpreters. The for sending your luck. it has been applied in that setting. But before neural networks that, where the language models, except they were using and grandmothers? So, criminals are just a count based models. So if you tried to, if you start to take two words and predict the third one, a you just held up how many times have seen any toward combinations and what came next and that what you predict us coming next is just what you ve seen the most often the training. So and so our lunch remodeling has been around for a long time. Neural networks have done wonders muddling for a long time. So really, New or interesting are exciting, has just realizing that when you scale it up with powerful enough neural net transfer You have all these emergent properties where, basically, what happens is If you have a large enough data set of text, you are in the task of protecting next word. You are multitasking a huge amount of different kinds of problems

you are multitasking understanding of chemistry, physics, human nature. Lots of things are serve clustered in that objective, two virgins, jacket but actually have to understand a lot about the world to make that protection. You just said the youth, word, understanding Are you the internet connection, feelings on? What do you feel like is doing, is searching for the right context in yeah? What is it? Was the actual process happening here? Yes, so basically it gets a thousand words in front a pretty thousands, first, and in order to do that very very well, over the entire dataset available on the internet, you actually have to basically kind of understand the context off of what's going on in there. and it's a sufficiently heart problem that you are, if you have a powerful enough computer like a transformer, you end up with interesting solutions and you can ask it to I'll do all kinds of things and do it?

It shows a lot of emergent properties lacking complex learning. There was the big deal with deputy in the original people when they published it is that you can just serve ah prompted in various ways and ask it various things and it will just kind of complete the sentence, but in the process of just completing the sentence, it's actually solving all kinds of really interesting problems that we care about. Do you think, is doing something like understanding In when we use the word understanding for us humans, I think is doing some understanding it in it's weights. It understands, I think, a lot about the world and it has to in order to predict the next word in a sequence, so a strain on the data from the internet. You think about this. This approach in terms of data of using data from the internet, do you think the turn it has enough structure data to teach a boy came. Yes, I

internet as a huge amount of data, a mature fits a complete and upset. I dont know that text is enough for having a sufficiently powerful edgy. I has not come. Of course there is audio and video and images know that kind of stuff, you're so text by itself, I'm the little bit suspicious about there's a ton of things we don't put intact in writing. I just are obvious to us about how the world works in the physics of it in the things fall, we don't put that stuff in tax, because why would you we share that understanding? taxes communication medium between humans and it's not a all encompassing medium of knowledge about the world. But, as you pointed out, we too have video and we have images and we ve audio, and so I think that the definitely a lot, but we haven't trained models are sufficiently across both across all this without his yet ass. I think that's what people are interested in, but I wonder what that shared understanding of like what we might call common sense.

has to be learned inferred in order to complete the sentence correctly, so may be the fact that its implied, The models can have to learn that not reading about it by infer, In the representation side and says just like we are don't learn. Common sense, like nobody says, tells us explore. They were just fear at all by interacting with the world right a soul. Here's a model reading about the way people interact with the world may have to further. I wonder yeah you briefly. What can a project called world of debts, training our health system to take actions and the internet version. just consuming the internet like that, what you think there's a future for that kind of system. Interacting with the internet to help the learning? Yes, I think that probably be the fowler frontier for a lot of these models, because

it says you met him. I wasn't opening. I was working on this project. Will the bits and basically was the idea of giving Europe that works, access to a keyboard and mouse and the idea of possibly go wrong. Basically you you perceive the impact of the screen. Pixels and obviously the state The computer is our visualized for human consumption in images of the web browser and stuff like that, and then you, given all the ability to press keyboards and his mouth and we're trying to get it to, for example, complete bookings and interact with user interface and dumb. Would you learn from that experience? A ghost on stuff, there's a super cool idea, it's like yeah they mean that the step between observer to actor is super fast eying stagger what the universal interface in the digital role, I would say,

and there's a universal interface in like the physical realm, which, in my mind, is a humanoid form factor kind of thing. We can later talk about optimists and so on, but I feel like there's a ah they're kind of like a similar philosophy in some way where the human, the world, the physical world nine for the human form and the digital world is designed for the human form of seeing the screen and using keyword, not keyboard and mouse. as the industry universal universal interface can basically our command the digital infrastructure we ve built up for ourselves, and so it feels like a very powerful interface to to command into Belinda. If I now at your question, is to like what I learned from that is interesting, because the role of bits was basically to early, I think at opening I at the time- This is around two thousand and fifteen or so, and does I gazed at that time was very different in the eye from this place today, time. Everyone was super excited about reinforcement, learning from scratch at this at the time of the authority paper where

neural or for playing a tar games and beating humans. In some cases, alpha go and so on. So I was very sad about training training on earth from scratch, using reinforcement, learning directly out there. We first maloney is extremely inefficient way of train earlier works because you're taking all these actions and all these observations- and you get some sparse rewards once in a while. So you do all this stuff. based on all these inputs and once in a while you're like told he did a good thing, he did a bad thing as just an extremely hard problem. You can't learn from that. You can burn the forest and you can sort of brute force through it, and we saw that I think with you know what I'd go and donor and so on and does work about its extremely inefficient and not how you want to approach problems are practically speaking as the best approach that at the time we also did took two world of bits would have an agent initialize randomly so with keyboard, mash and mouse mash and try to make a booking, and it just like revealed the insanity of that approach very quickly where

you have to stumble by the correct booking in order to get a reward of, you did correctly and you're never gonna stumbled by it by chance at random. so even with a simple web interface is too many options, because too many options. And its two sparse of reward signal and you're starting from scratch. At that and so you don't know how to read, you don't understand pictures images buttons. You don't understand what needs to like make a booking, but now it's happened is our. It is time to revisit that and open your eyes interests in this companies like adapt our interest in this and so on, and the ideas coming back, because the interfaces very powerful, but now you're, not training in asian from scratch you are taking. The djibouti is initialization, so Jupiter is pre trained on all of text and it understands what so booking it understands was a sub submit it dense quite a bit more, and so it already has those representations. They are very powerful and that makes all the training significantly more efficient and more, problem tractable, should the interaction be would have liked the way humans see it

the buttons in the language or- bill with html javascript innocence. Yes, that's what do you think is the butter, all this interaction is mostly in the level of html services and so on. That's done because of computational constraints, but I think ultimately everything's for human visual consumption, and so, at the end of the day, there's The additional information is in the layout of the web page and what's next to it, and what red background in all this kind of stuff and what it looks like visually. So I think that the final frontier we are taking in pixels and were given out keyboard mouse comments, but I think it's impractical still today do worry about box on the indian. Given given these ideas, given how exciting the ideal robots on twitter being not the stupid boss, though she now the crypto botz, but the boy that might be out there, actually that we don't see that therein I think interesting ways to this kind of system, it should be able to pass the I'm not robot click button whatever,

Josh understand how that does works. I dont quite like us does that check box or wherever you? Click is presumably tracking like mass movement and the time and so on. So exactly this kind we're talking about, should be able to pass their. So yeah. What do you feel about botz? There are language models plus have some interact, ability and are able to tweet reply saw you worry about that world. I think who has been a bit of an arms race between sort of the attack in the defence are so the attack will get strong. But the defence will get stronger as well. Our ability to detect that had he defend honey, detect. How do you know that you're, a cop party account on twitter is as human how'd you approach? I give people claim you know I how'd you defend yourself in the quarter.

law that I'm a human. This account is, he had something I think it might be. I think the society sexual evolve, I, like my start, signing digitally signing. Some of our respondents are you things that we create right now, it's not necessary, but maybe in the future it might be. I do think that We are going towards a world where we share wish the digital space with a ice synthetic beings yeah and they will get much better and they will share our digital rome and, though, eventually shower physical aromas, while much harder, but that's kind of like the world. We're going towards and most of them, be benign and awful and someday be malicious and it's going to be an arms race trying to detect them. So I mean the worst: isn't the ai eyes the worst, Does the Aspartame human cement, always malicious desert, obviously live militias applications, but it could also be in a thousand I out travelling heart pretend to be human because human

world our own, getting your respect as an eye some love and respect. I don't wanna autonomous, intractable people are thinking about the proof of person, hood Yes and we might start digitally signing our stuff and we might all end up having like AL. Basically some solution for prefer person who doesn't seem to me intractable, it's just something that we and had to do until now, but I think once done, like really starting to emerge, which is soon. I think we will think about it much more so, but that too will be is because, obviously it probably ah spoof or fake the proof of personal, She has to try to figure out how to probably it's weird aware, like source carry numbers and like passports as love, it seems like is harder to fix up in the physical space.

Yes, it's your space. It just feels, like is gonna, be very tricky very tricky to out, because it is simply a pretty low cost fixed, What are you gonna put any I in jail for like trying to think he's. A fake, fake person had proof Ok, fine, you put a lotta he eyes in jail, but they'll be more as arbitrary exponentially more. The cost of creating bought is very low. and unless there is some kind of way to track accurately, Ike you're not allowed to create any programme without showing You yourself to that programme, like you any programme, the rise in the internet. You be able to trace every single human programme in those above that provide it maybe have to start declaring when You know where to start drawing those boundaries and keeping track of ok. What are digital entities versus human

It is and what are the ownership of humanity's individual entities and are something like that, I dunno, but am I think, I'm optimistic that this is this is possible and its in some sense we are currently in like the worst time of it, because all these bots suddenly have become very capable. But we don't have the fences yet built up as a society and but I think that doesn't seem to me intractable it just something that we have to deal with issues we're that the twitter, but, like really cool. between robots so numerous, I guess, is it. I presume that engineers, a twitter, are very good, sir. Seems like what I would infer from that is. It seemed like a hard problem. It they're probably catching the if our deserve steel man, the case It's a hard problem and there is a huge cost to ah false positive.

two to removing opposed by somebody. That's not about that's, creates a very bad user experience there very cautious about removing the maybe babies, and maybe the boss a really good at learning? What gets her mood than not sick, they can stay ahead of the removal process very quickly. My impression of it honestly as there's a lot of low hanging fruit, I mean that's what I it's not my impression with its not, but you have here as my impression as well, but it feels like maybe seeing the tip of the iceberg. it'd be the number of bots, isn't like the trillions, and you have to like you just it's a constant assault of boxing yeah yeah yeah. I dunno. I bet you have to sue many cases about some seeing a pre like obvious, I could write a few lines of code they catch these spots.

I mean definitely there's a lot of low hanging fruit, but I will say I agree that if you are a sophisticated actor, you could probably create a pretty good bought right now and using tools like gps because it's a language model you can generate faces tat. Look we have quite good now. can do scale, and so I think he has quite possible and it's going to be hard to defend there is a global engineer, the claim the lamb was censure. Do you think, there's any inkling of truth to what he felt And, more importantly, to me at least, do you think good models, will achieve censure. The illusions: agents soonish fish here it's a little bit of a canary Nicole mine kind of moment, honestly a little bit because So this engineer spoke to like a chat bought at Google end up becoming that does not a cent hint. He asked some existential

recital a question: if you like reasonable answers and looked real end, and so on to me, it say yeah he was he was. Are he wasn't sufficiently tried to stress the system? I think and are exposing the truth of it as it is today but I think this will be quick increasingly harder over time, so yeah, nothing more people will basically become getting warmer, there'll be more people like that over time, as this gets better like form an emotional action to up to today? I have only possible in my mind. I think these areas are actually quite good at him human connection, human emotion, a ton of text on the internet is about humans, connection and love and so on. So I think we have a very good understanding it, some in some sense of of how people speak to each other about this, and

they're, very capable of creating a lot of that kind of text. The and there's a lot of legs I from fifty and sixty that imagined a eyes in a very different way. They are calculating cold and balkan like machines, that's not where you're getting today we're getting pretty emotional here. that actually are are very complex, and capable of generating a possible funding taxed with respect to all of these she emily hopeful body. I systems that are like companions that hope you grow develop as a human being. I help you maximize longterm happiness, but I am also very worried about a I systems that figure off from the internet, the hue get attracted the drama of these issues like shit talking to eyes. They just cause. Did you hear little do gossip Do they try to plant seeds of suspicion to other humans that you love and trust and just kind of message? People in Iraq

That's going to get a lot of attention to drama max my stronger on the path of maximizing engagement and ass. He was will feed into that machine here and get it ll be a giant drama shitstorm. Oh yes, so I'm worried about that shows the objective function really defines the way. human opposition, progressive with eyes in here. I think, right now, at least today they are not serve its not corrected I think of them ass goal, seeking agents that want to do something. They have no, long term memory or anything they, it's literally a good approximation of it. Is you get a thousand words and you're trying to predict the thousand and first and then you continue feeding it in and you are free to prompt it in whatever way you want so in, So you say: ok, you are a psychologist and you are very good and you love humans and there's a conversation between you and another human human column, something you something and then

This continues the pattern and suddenly you're having conversation with the fix. I colleges is like trying to help you, and so it's done. kind of like in the realm of a tool is a young people can be their number three ways can create really incredible taxed doesn't have long term goals over a long period of time. It doesn't try to sort out. Look that way right now, but You can use short term goals that have long term effects of my prompting shore. Our goal is to get on your capacity to respond to me on twitter, whenever I think I might that's the goal, but my figure out talking shit. You will be the best in a highly sophisticated, interesting way. And then you build a relationship when you were once and then it like over time. It gets to not be sophisticated, and just like just talk shit,

what is an okay, maybe you won't get to Andre, but it might get to another celebrity and may get two other big, a yeah you'll, just so with just that simple: go get them to respond here, maximize the probability of risk actual risks. yeah? I mean you could prompt a powerful model like this with their its opinion about how to do any possible thing you're interested in it, so they will just the kind of contract to become these oracles. I could sort of think of it that way they are oracles are currently is just text, but they will have calculus her stay with access to. Google search may well have all kinds of gibson business. They will be able to operate the internet and find different it. For me, and dumb in some sense that's gonna. Currently what it looks like in terms of the development. Do you think you'll be an improvement eventually over what google is for the human knowledge like he'll, be a more effective

engine to access human knowledge. I think there's definite scope and building a better search engine today and I think, google they have all the tools of the people. They have everything they need the older puzzle pieces. They have people training, transformers at scale. They uphold the data, Ah, it's just not obvious if they are capable of an organisation to innovate under search engine right now and if they don't, someone else, will there's absolute scope for building a significantly better search engine built on he thought he saw your thing, a large company, where the search there is already an infrastructure works. Ass brings out a lot of money so where structural inside the company? Is their motivation to pivot yup? We're going to build a new search engine at its heart. How are these you're going to come from a stirrup right that some that will be yeah some other come more competent organization. So are currently, for example, maybe bing another shot at it,

though microsoft we're talking offline, definitely is really interesting, because search engines used to be about ok, you're, some query: here's here's! Here's web pages stead look like the stop it you have, but you could just directly go to answer and then have supporting evidence and these these models? Basically, they ve read all the texts in they read all the web pages, and so sometimes, when you see or sell going over to search results and serve getting like a sense of like the average answer to whatever interested in like that. Just directly comes out, you don't have to tat work. so they're kind of like her you, I think they have a way to this of the stilling or that knowledge into, Like some level of insight, basically, do you think of prompting, as a kind of teaching and learning like this whole process, like another layer, yoke is maybe humans are already have their background model and the world is prompting you yeah exactly.

I think the way we are programming these computers now, like you is, is urging to how you programme humans, I mean how to program humans, the prompt I go to people and I prompt them to do things like prompt them from information. and so our natural language prompt, is how we programme humans and were started programme computers directly. That interface is like putting remarkable honestly, so you spoken a lot but the idea of software two point: oh all good ideas become quite cliches so quickly The terms is kind of hilarious, it's getting Eminem once said that like if he gets annoying by signing he's written very quickly. That means it's gonna be a big hit, because his is too catchy by he described this idea and how you're thinking about it has evolved over the months and years since, since you coined it yeah guess I had a blog post on software.

Porno. I think, several years ago now and and the reason I wrote, that post is because I kept I kind of saw something remarkable happening in like software development and how a lot of code was being transitioned to be. written not in sort of Luxy, boss, boss and so on, but it's written in the weights of a neural net, basically just saying that neural nets are taken over software, the realm of software and the taking more murmur tasks, and at the time I think not. Many people understood this deeply enough. This is a big deal, is a big transition. Neural network were seen as one of multiple classification, algorithms he might use for your data problem on capital like this is not that this is a change in how we programme computers and I saw a neural nuts as this is going to take over the women. I'm computers are going to change, is not going to be people writing a software in people's bus or something like that and directly programming the software. It's going to be the committee

in our training, sets and data, sets and crafting these objectives by which we train these neurons and at some- there's going to be a compilation process from the data sets and the objective and the architecture specification into the binary which is really just are the neural net. You know weights and the forecasts of the neural it, and then you can deploy that binary ass far ass. I can about that sir transition and that's what the poster about, and I saw this serve play out in a lot of fields. Ah, you know all about are about being one of them, but also just a simple image. Classification people thought her region. wait. You know in the eighties and so on that they would write the algorithm for detecting a dog in an image, They had all these ideas about how the brain does it and first we detect corners and then we detect lies in that. We stick them up and they were all really going added. They were like thinking about how they're get it write for them, and this is not the way you build it.

The smooth transition where ok. First, we thought we were gonna build everything. Then we were building the features so like hot features, and things like that that detect these losses what patterns from image patches and then there was a little bit of learning on top of it like a support, vector machine or binary classifier for cat vs, dog and images on top of the features. So we wrote the features, but we trained the last layer, a sort of the the classes and then people are like actually less, not even design features, because we can't honest them are not very good at it. So let's also learned features and and up. We have basically come little neural net where, you're learning most of it you're just specifying the architecture and it the architecture has tons of filling blanks which has all the knobs and you let the opposition right must have, as this transition is happening across industry everywhere and suddenly we end up with a ton of code that is written in neural net weights, and I was just pointing out that the analogy is actually pretty strong and

What of a developer environments for supper? Bumpo like we have her eye the east. Are you work with good tobacco daddy? How you run code? How do you mean? Could we have github? So I was right Make those analogies in the new realm like what is the git hub of software to pilot turns out that something that looks like hugging face right now you know, and so I think some people took it seriously and built cool companies and many people. We attacked the posted actually was not well received when I wrote it and I think maybe it has something to do with the title, but the bus was not well received and I think more people surf have been coming around to it over time. Yes, you were the director of a high atossa aware. I think this idea was really implemented a scale? is how you have engineering teams doing software to point out. So can you linger on that idea of why they were in the really early stages of everything you say which is like get how, by the ease like

How do we build engineering teams that that work and software to point of systems and in the end, the data collection and the data annotation, which is all part of the software to point out like what do you think is. The task of programming is also to point out. Is it debugging in the space of hyper parameters, or is it also bug in the space of data? Yet the way by which you programme the computer and influence its over them is not by writing de commence yourself. Your changing, mostly the dataset, Are your changing de las functions what the neural net is trying to do. I was trying to predict things, but it basically the assets and the architecture of the neural luck, and so, in the case of the auto pilot, a lot of the deer sets to do with, for example, detection of objects and legally markings and traffic lights, and so on? So he accumulate

data sets of here's an example: here's the desired level and then here's roughly how the ark here's roughly what the algorithm should look like and that's accomplish neural net, so specification architecture, is like a hint ass to what he augur should roughly look like and fill in the blanks process of optimization is the training process and you take your neural net that was trained. It gets all the right answers data set any deployed feathers in that case, perhaps at all machine learning cases there's a lot of tasks, so is coming up formulating a task like for a multi had in your network is formulating a task part of the programming yeah kramer, so how you breakdown problem into a set of tasks, I'm in a high level. I would say if you look at the software running in the auto pilot, I gave her a number of talks on the star.

I would say originally a lot of it was written and software one Pineau, there's not a lot of seaports, boss right then gradually there was a tiny neural net that was, for example, predicting in a single image: is there like a traffic light or not, or is there a landline marking or not, and this neural net didn't have too much to do in this in the scope of the software? was making tiny predictions on individual little image and then the restless system stitched up so Ok, we're! Actually you don't have just a single camera with cameras. We actually have a camera overtime. If so, what do you do with these predictions you put them together? How do you do the fusion of all that information and have you act all that was written by humans, ass, he passed boss and them, decided, okay. We don't actually want to do all of that. Fusion in cyberspace because rationally, not good enough to write that algorithm beyond the normal right, the honour them and we want to port- are all that software into the tunnel stack

and so, then we actually had neuromas dead. Now, take all the eight camera images simultaneously and make predictions for all of that, also and and and in, and actually they don't make predictions in a in the space of images. They now make predictions directly in three d, and actually they don't in three dimensions around the car And now, actually we don't manually fuse the predictions over to in three d overtime. We don't trust ourselves to ride that tracker. So actually we give the neural net all the information over time. So it takes these videos now and makes. Predictions as he are just like, putting more and more power to the electoral processing and at the end of it, the eventual serve goes to have most of the software potentially be opportunity to Poland, Has it works significantly better humans are just not very good at writing software. Basically the predictions, they are happening in this, like forty land three dimensional world over time? How do you do invitation that world

but what have you as soon as data irritation, whether itself super eyes. Manual by humans is on his as a big part of this effort to put our world right, I would say by far and the industry, if you like about the industry and how much does the technology of what we have? Well, everything is supervised learning, so you need a dataset of input, desired output and you need lots of it and there are three properties of it that you need. He needed to be very large. He needed to be accurate, no mistakes and you needed to be diverse. You don't want to. Ah just have a lot of correct examples of one thing: you need to really cover the space of possibility as much as you can and the more you can cover the space of possible and puts the better the algorithm homework again now, once you have really good data sets that year, collecting curetting and cleaning

and train your neural net, and on top of that, so a lot of the work goes into clean the assets. Now, as you pointed out, that's probably it could be. The question is: how do you achieve a ton of? Ah, if you want to basically predict authority need data entry to back it up. So in this video we have eight videos coming from the cameras of the system. This is what they saw, and this is the truth what actually was round there was discarded, or is this car the scar, These are going landmark, excessive geometry of their out. There's traffic, light industry, dimensional position: you need the ground truth, as are the big question. That team of solving, of course, is how do you arrive at that countries, because once you have a million of it and its large clean and diverse then training your mac on? It works extremely well and you can ship out into the car, and many mechanisms by which we collected that a train data, you can always go for communication you can, Four simulation as such source. You can also go for what we called the offline tracker that we so

about at the idea and so on, which is basically an automatic reconstruction process for taking those videos and recovering the three dimensional serve reality of what was around so, basically think of doing like a three dimensional reconstruction as an offline thing and then understanding that ok, there's ten seconds of video. what we saw and therefore here's a link on cars and so on. And then, once you have that annotation you can train the lead to imitated and how difficult the reconstruct the three reconstruction is difficult, but it maiden so there too, there there's overlapping cameras in you. Do the reconstruction there's, perhaps there the inaccuracy, so that's caught cotonou irritation step. Ah, yes, the nice thing about the invitation is that it is fully offline. You have infinite time. You have a chunk of one minute and you're trying to just offline in a supercomputer somewhere figure out where were the past. You have all the cars, older people and you have your full one minute. A video from all the angles- and you can

old in euro lets you want and they can be very efficient, massive neural nuts. They can be. must I can't even running the car later at this time, so they can be given. Powerful, no less than what you can eventually the point, so you can do anything. You want three dimensional, a structure, neural mats? Anything you want is to recover the truth, and then you supervised at risk What have you learned? He said no mistakes about humans, doing adaptation his eyes in humans, there's like a range of things are good at in terms looking stuff on screen. Is that how interesting that's you of a problem of designing, tater were humans are accurate, enjoy it quantity, even the metrics efficient are productive, all that kind of stuff yeah so I grew. The invitation team at us live from biscuit zero, two thousand. Ah, while I was there that was really interesting. In my background, is a beauty student research

so growing- that common organization was pretty crazy about you. I think it's extremely interesting and part of the design process very much behind the auto pilot as to where you documents. He must are very good at certain kinds of imitations. There were good, for example, what to do contains of images, do not good at and attain cars over time and three dimensional space very, very hard. and so that's why we were very careful to design the task that are easy to do for humans process things that should be left to the often tracker like maybe the maid computer will do all the triangulation and three during reconstruction, but the human will say exactly these pixels of the image our car is actually least pixels. and so co designing. The other data annotation pipeline was very much bread and butter wasn't what I was doing daily. Getting there still a lot of open problems, that space just in general. Annotation were the stuff them she's, a good at machines do and the humans do what,

a good out, and there may be some iterative process right. I think, to a very large extent through the number of federations and we learned a ton about how to create these data. Sets. I'm not seeing bigger problems like origin. When I joined I was like, I was really not or how this will turn out here, but by the time I left. I was much more secure in so any sort of understand, philosophy of how to create the assets, and I was very comfortable with where that was at the time. So what are strengths limitations of cameras for the driving task in europe. standing when you formulate driving ask as a vision task with eight cameras. You ve seen the entire, in most of the history of the computer vision field, when it has to do in your own words? What she's, if you stepped back, what are they strengths and limitations of pixels using the exhaust the drive yet pixels, I think, are a beautiful sensory. If a sensor, I would say, the things like cameras are very, very cheap and they provide a ton of information.

ton of bits also its are extremely cheap sesar for a ton of bits and each one of these bit as a constraint on the state of the world, and so you get lots of megapixel images very cheap, and it just gives you all. These concerns it's for understanding what actually out there in the world. So vision is probably the highest bandwidth sensor, if Mary and was ass and on our love, the pixels It is as a constraint on the world as as highly complex, hide with constraint on the way and state of the world. They assess it s not just about, but again this real real importance of its the sesar documents use. Therefore, everything is designed for that sesar, the text, the writing the flashing sized everything designed for vision and so on, and you just find it everywhere, and so that's why? That is the interface you want being taken against, your full interfaces and

that's what we actually want to measure the world as well and then develop software for that answer, but there other constraints on the state of the world that humans? to understand the world any vision. Autonomy is the main one, but we were like world. referencing our understanding of human behaviour in some common sense physics, that could be inferred from vision from from a perception perspective, It feels like we're using some kind of reasoning, to predict the world and not just the pixel, some you a powerful prior rights for the world evolves over time, etc. So not just about the likelihood term coming up from the data itself telling you about what you are observing, but also the prior term of like, where were the likely thanks to see and how do they likely move and so on, and the question is how complex is the the the range of possibilities there?

it happened in the driving task. There still is that you still an open problem of how difficult is driving like, philosophically speaking, to do what you've on driving? Do you understand how hard driving is yeah, jarrings really hard? because it has to do with the protection of all these other agents and the theory of mind- and you know what they're going to do in our they look nice who are they were they looking? Were they thinking yeah, there's a lot that goes there at the at the full tail off the the special denies that we had to be comparable with that. Eventually, the final problems are of that form I dont think the problems that are very common. I think eventually are important, but it's like really in the talent in the and the rare edge cases were from the vision perspective, what are the toughest parts of the vision problem of driving and while basically the sensor is extremely powerful, but you still need to process that information and and so

from brightness of dismissal valleys to hay here. The three dimensional world is extremely hard and that's what the new works are fundamentally doing, and so the difficulty really is in just doing extremely good job of engineering, bein tar pipeline interrogator engine having the capacity to train these neural nuts, having the ability to evaluate the system and the it ass. I would say just doing this in production at scale is like the hard part. It's an execution problem, so data engine, but also the deployment of the system such that has low latency performance so as to do all the steps yet for the neural net, specifically just make sure everything fits into the chip on the car near and you have a finite budgeted flops that you can perform and are a memory with another constraints, and to make sure advice and you can squeeze in as much as you can into the tiny what, from that process, because it maybe it s one of the bigger like new things

coming from a research back on where this the system has to run under heavily constrain resource right has the wrong really fast. What what kind of inside of you learned from that you're mature fits of theirs. Many insights you're, trying to create a euro less that will fit what you have available and you're always trying to optimize it, and we talked a lot about it on the ai day and basically, the the triple backflips that the team is doing to make sure it all fits and utilizes the engine. So I think it's extremely good engineering and then there's all the all kinds of little incites pepperdine on how to do it properly in the sixties. Denmark is everything we talked about. The data engine, the entirety of the lay out of this idea that eating it just beautiful well with humans in the loop. Or can you describe the data engine? Yet the dare engine is what I call the almost biological feeling like process by which you are perfect. The training sets for these neural

so because most of the programming now, as in the level of these data, sets and make sure their large, diverse and clean up. Basically, you have data set that you think is good. You turn your role at you, deploy it, and then you observe how else performing and you're trying to always increase the call your data set you trying to catch scenarios. Basically, there are basically rare and, in these scenarios that junior illness all typically struggling, because they weren't told what to do in those rare cases in the dataset. But now you can the loot, because if you can now collect all those at scale you can, then I feed them back into the reconstruction process. I describe and reconstruct the truth. In those cases, an additive dataset as the whole thing ends up being like are a staircase of improvement off up

in your training set and you have to go through deployments so that you can mine are the parts that are not yet represented well in the desert, so your data says basically imperfect. It needs to be diverse, it has pockets. There are missing and you need to pat out the pockets. You can sort of think of it that way in the data world to humans playing this so What's the earth is biological system they are hereby is made up of cells What role like? How do you optimize the human system, the multiple engineers collaborating figuring out, what the focus on what you contribute, which put, which tasked optimize in this new, on that who is in charge of figuring out which trafficking is more data. What case can speak to the Khyber parameters? The human system? I religious, comes until extremely good execution from an engineering team

they're doing they understand intuitively the fizzled, philosophical insights underlying the deed engine and the process by which this system improves and how to again like delegate the strategy of the vehicles and how that works, and then just make sure it's all extremely well executed, and that's where most of the workers is not even de philosophizing or research or the ideas of it is just extremely good. Execution is so hard when you're doing with data So your role in the data engine executing well on it is difficult for me. Is there a priority of like, like a vision, board a thing like we re if you get better stoplights like that, the prioritization of tasks is that essentially in that comes from the data that comes to the very large extent to what we are trying to achieve in the product roadmap or we're trying to the release we're trying to get out in the feedback from the queue a team worth it where the system is struggling or not, the things we're trying to improve and the curating gives some

I know some information in aggregate about the youngster, the system and various conditions, and then, of course, all of us drive it, and we can also see it. It's really nice to work with a system that you can also experience yourself, and you know it drives you home. It's is there some, inside you can draw from your individual experiences. You just can't quite get from an aggregate. Statistical analysis of data is so weird right. Yes, it's it's, not scientific in a sense, could you just one anecdotal sample YA, think there's a ton of it? Is it's a source of truth? your interaction with the system, and you can see it. You can play did. You can put her bet. You can get us if you have an intuition for it, I think numbers just like have a way of numbers and plots and graphs are much harder and it hides a lot it's a gift. You try and language model,

in a powerful way, it is by you, interacting with it. One hundred per strike try to build up an intuition yeah, I think like elon, also like he always wanted to drive the the system himself. He drives a lot and I am going to tell my daily so ah, he also sees this as a source of truth. You driving the system and it performing and yeah. That would you Big tough questions here so desolate last year removed radar from mom from the census, sweet and now just analysis. Can you remove? Ultrasonic sensors rely solely on vision, so camera only does that make the perception proper, harder or easier. I would almost rephrase the question in some way, so the thing is basically do you would think that additional, first by the way, Kansas interrupt. I wonder if the language model ever do that review prompted. Let me on the question of the epic. This is the wrong prom.

yeah? So it's like a little bit of a wrong question, because basically you would think that these sensors are an asset to you. But if you fully consider the entire product in it's entirety, the sensors are actually potentially liability because the sensors aren't. They don't just appear on your car? You need selene, he haven't entire supply chain, yet people procuring it other can be problems with them. They may need replacement. They are part of the money, from process. They can hold back the line in production. You need two sources to maintain them, you have to have teams that right, the former oliver of it and then you also to incorporate and feeds them into the system in some way, and so it actually like bloats during the d, a lot of It- and I think it is really good at simplify simplified best part- has no part, and he always try, throw away things? There are not essential because he understands the entropy in organizations and an approach, and I think in this case the cost is high and you're, not potentially sir Yet if you just computer vision, engineer and I'm just trying to

prove my network and is it he more useful or less useful heavy price? Was it and the thing is, it must preserve the full cost of assent it actually but actually liability and you need He really sure that it's giving you extremely useful information. In this case we looked at using it or not, using it and a delta was not passive, and so that useful is it also. low in the data engine like having more sense understand at the end, is a disaster action and the sensors that can change over time. For example, you can have one type of se radar. You can have other type of radar that change over time. They suddenly need to worry about it. Now. Suddenly you have a column in your sql light telling you oh, what sensor type was it and they all have different distributions and then all they can just be contribute. Noise and entropy into everything

they bloat stuff and also organizationally has been really fascinating to me that it can be very distracting and if you, if all if you only want to get to work as vision, all the resources are on it and you're building out a data engine and you're actually making forward progress, because that is these are the sensor with the most bandwidth, the most constraints on the world and you're investing fully into that, and you can make that extremely good if you're you're only a final, One sort of spend off our focus across different facets of the system, and this kind of reminds me of suddenly a bitter lesson they just like simplifying the system in the long run, the casino well with the long as it seems to be always the right solution. Yes, in that case was for rl, but it seems to apply generally across all systems that you can

edition yeah, so where what do you think about the lidar as a crutch debate or the battle between point clouds and pixels yeah? I think this debate is. I always like slightly confusing to me because it seems like the actual debate should be about like do you have the We are not just like the really important thing about whether you can achieve a really good funds. I have an ai system at that scale, so data collection systems, yeah, do you have a fleet or not as significantly more important, whether you have lidar or not just another sensor and and yeah, I think, similar to earth radar discussion. Basically, I am, I don't think it. It basically doesn't offer extra, ah extra information, it's extremely costly, it has all kinds of problems. You have to worry about any of the calibrated cetera it creates bloat and entropy. You have to be really sure that you need this. This sensor. In this case, I basically

thinking it then I think ass. They will make a strong statement. I think the others are. Some of the other companies are now using. It are probably want to drop it. Yes, all you have to consider the sensor in the fool in considering. Can you build a big fleet that collects a lot of data? and can thinking that sensor within that their data and that sensor into a data engine that able to quickly find different parts of the data that then continuously proves, whatever the model that you're using you. Which look I dislike vision is necessary in the sense that the driver the world is designed for human visual consumption, so you need vision is necessary. and also it is sufficient because It has all the information that you, but you need for driving, and humans obviously has vision to drive since both necessary and sufficient safe,

to focus resources, and you have to be really sure if you're going to bring in other sensors, you could you could, you could add an assessors to infinity at some point to draw the line, and I think in this case you have to really consider the full cost of any sesar better adopting, and do you really need it, and I think he s an occasional, so think about the idea that the other companies are for me high resolution maps and constraining heavily the geographic regions which they operate. Is that approach? Not you're. In your view, nothing the scale over time to the entirety of the united states. I think choosing mentioned like they pre map of the environments and they need to refresh the map and they have a perfect centimeter the accuracy map of everywhere, they're gonna drive it's crazy. How are you going to talking about the autonomy actually changing the world. We're talking about the deployment honestly on on global global scale, autonomous systems for transportation, and if you need to maintain sending your accurate

for earth or electoral many cities and keep them. There is a huge dependency that you're taking on huge dependency. It's it's a maximum dependency. And now you need to ask your shall do you really need it? And I don't need it right. So it is very useful to have a low level map of like. Ok, the connectivity of your road but there's a for coming up when you driving environment you serve. Have that how little understanding, if, like a small google map- and Tesla uses google map like similar kind of resolution information in the system, but it will not pre map environments to semi center middle of the accuracy. It's a crutch, it's a distraction, it costs entropy and it Does the team dilutes the team and you're not focusing on what actually necessary witches comparison problem? What did you learn about machine learning engineering about life about yourself as one human being from work. with elon musk. I think the most I've learned is about how to serve run organizations efficiently and are too

create efficient organizations and how to fight entropy in an organization human engineering in the fight against entropy here, there's theirs. I think, It is a very efficient warrior in the fight against entropy and organisations, intervene in an organization like acceptance, its process, its its profits, as an in and inefficiencies in the intermission things and that kind of stuff yeah meetings he hates meetings, he keeps telling people to skip meetings if they're not useful and he basically runs the world's biggest stirrups. I would say, tat must be: are the world's biggest startups tesla actually actually, multiple startups? I think it's better to look at it that way, and so I think he's he's extremely good atta at that, and I gives a very good intuition for streamlining processes. Making everything efficient best part is no part of simplifying

sing and and just kind of removing barriers are moving very quickly. Making big moves at all. This is a very strappy sort of seeming things, but at scale so strong drive to simplify from your from your perspective. I mean that and that probably applies to just designing systems and machine learning and otherwise exemplifies, simplify yes, What do you think is the secret to maintaining the start of culture in a company that grows there? Can you interest, but that I do think you need someone in a powerful position with a big hammer like he long whose, like the cheerleader for that idea and ruthless ruthlessly pursues it. if no one has a big enough hammer, everything turns into committees demand see within the company process talking to stakeholders soon, making just everything just from if you have a big some, who is also really smart, that has a big hammer. Things move quickly ceased

your favorite seen in interstellar as the intense stockings With the ai cooper talking saying cooper. What are you doing docking? It's not possible. No, it's necessary such a good line by the way, just so many questions there, why hey. I in that scene presumably is supposed to be able. Computer, allow more than the human things optimal? Why the human? I mean that's a movie but I shouldn't say I know much better than the human anyway woody. Think as the value of setting simile, impossible goals, like our intuition seems like something that have taken on an illinois espouses that, where the initial intuition the community may say this is very difficult and they, he could on anyway with a crazy deadline. You just a human engineer, perspective farm,

Have you seen the value that I wouldn't say that setting impossible girls exactly is good idea, but I think setting very ambitious goals is a good idea. I think there's a rifle sub, linear skilling of difficulty. I, which means that tax problems are not tonics hard, usually tonics brought tax harder problem. It is like to three x harder to execute on, because if you want actually like, if you will improve a system by ten percent, it costs some amount of work. And if you want a telex improve the system, it doesn't cost in a hundred x amount of the un's, because you fundamentally change the approach and if you start with that constraint that some protests are obviously going to work and it forces you to re, evaluate- and I think it's a very interesting way of approaching problem solving- requires us, we're gonna, think age to go back. You like phd days like how do you think which ideas in machine learning community are solvable. Yes,

It requires what was there. I mean, there's accoucheur, first principles, thinking by air the basically ignored the community, saying that doesn't like me, doesn't a community in sight usually draw lines of what is in his impossible right and like it's very hard, break out of that. Without going crazy yeah I mean, I think, a good exam here s another deep learning revolution in some sense, because you could be in computer vision at a time when during the deep learning serve evolution of two thousand twelve and so on, you could be in the computer vision stack by ten percent, or it can just be saying. Actually all this is useless, and how do I do ten x, better computer vision, while it's not probably by tuning a hog, feature detector? I need a different approach. Hmmmm, I need something that is scalable going back to richard Sutton's and understanding subtle. philosophy of the yacht, bitter lesson and then being like actually need much more stable system like neural network that, in principle works,

having some people neighbours that can actually execute and admission make it work so where's the tenet solution, what do you think is the timeline to solve the problem of autonomous driving. This still in part, open question yeah. I think the top then what time lists of salt driving obviously, is that no one has created a driving near. So it's not like what do you think is attempting to build this bridge while we build million bridges before here's, how long that takes its it here, it's no one has built a ton of me. It's not obvious. Some parts turn out to be much easier than others since really hard to forecast. You do your best, based on trend lines and so on and based on international, but that's why fundamentally just really hard to forecast us. No one is the even still like being inside of it is hard to to do yes, something

turned out to be much harder and some things turn out to be much easier. Do try to avoid making forecasts cause like elon, doesn't avoid them right and has of car companies in the past have not avoided it either. Forward, another places have made predictions that would solve level for driving by twenty twenty twenty one. Whatever enough they all kind of backtracking prediction. I you as far as an ally, person you for yourself privately make predictions what do they get in the way of like your actual ability to think about a thing y'all would say like what's easy to say, is that this problem is tractable and that the z production to make its front belong, is going to work. Yes, just just really hard. Something's are not to be harder than as something's her to be vizier also, but its. It definitely feels tractable and it feels like at least a team at us law, which is what I saw internally, is definitely attracted up. I do you form

A strong representation that allows you to make a prediction about tractability, so the leader of a lot of humans? You have the courage to say this is actually possible. Happy birthday, wishing it doesn't have to be even driving could be other tasks. eddie what a difficult task to work on your left me classification achieve certain just an image net, a certain level of superhuman them performance, yeah expert, intuition just intuition its belief So just like thinking about a long enough studying looking a simple data like said driving My too issues really thought on this, the guy I have a good intuition about tractability as it could be either it could be anything it could be solvable like a know. The driving task could be simplified into something quite

trivial, like the solution to the problem, would be quite trivial and at sea. A more more cars drying perfectly might make how much easier, if the more as you have joined, like people learn how to drive correctly, not correctly in a way that more optimal for virginia system of autonomous and semi autonomous and we german cars that could change stuff then again also spent over Guess number of hours just staring pedestrians crossing streets thinking about humans, and it feels like the way these are eye contact, it sends the strong signals in Anders certain courts and educators of behaviour and, of course, a lot of the fatal. it is that happen have to do with driving and both an epidemic. citing the driver sizes, does that problem of John. at night nor their color. So I wonder you take the space of past

the solution to time is driving could so many human factor issues. There is almost impossible to predict it. There could be super clean, nice solutions, yet I would definitely like to use a game analogy theirs. Furthermore I bought a you, definitely also see the frontier of improvement. You can measure historically hub watch, you have made progress and I think, for example, what I've seen them roughly five years at a slob when I joined it barely kept lane on the highway. I think going up from allotted to us. I was like three or four interventions: anytime: do road would do anything geometrically or turn too much? It would just like network and so going from that, like a pre competent system figures and seeing what happens also under the hood I end with the scale of the team is operating now. There's big data and computer everything else. just are massive progress so far your climbing a mountain and yes, father you make a lot of progress, log you're, making progress, and you see what the next russians are, and you look

some the remaining challenges and are not like out not perturbing. You and they're not changing your philosophy and you're, not control contorting yourself, you're like actually these are the things that we still need to do at the fundamental components of solving the problem seems to be. There is data engine to the computer, The computer on the cards, the computer, has a training all the cost of, so you done Over the years I ve been a tells you done a lot of amazing breakthrough DE as an engineering all of it from the data engine to the human side of it. Can you speak to why you chose to leave tesla, Basically, as I just got their ran, I think overtime. During those five years I kind of got my and two oberon managerial position. I mustn't my days were in meetings and grown organization, and making decisions about us are vital. Strategic decisions about the team

but it can be working on a so on and it's kind of like a corporate executive role, and I can do it. I think I'm okay at it, but it's not like fundamentally what I what I enjoy and so I think, when I joined there was no computer vision team, because just I was just going, The transition of using mobile lie at their party vendor for all of its computer vision to having to build its communication system. So I showed up there were two people drinking your olympics. they were trained them at a computer at their at their legs? Like doesn't there's a basic classification task yet, and so I kind of like group into what I think is a fairly respectable. Deploring team must compete cluster, a very good didn't organization, and I was very happy with what was, it became quite autonomous, and so I can have stepped away- and I am very sorry, do much more technical things again yeah kind of like we focus on eight july. What was this also she likes. You took a little time off like white arm,

steamed mushrooms that you take away the war was, do you mind that human lifetimes finite he did a few incredible things here. One of the best teachers of aid in the world you're one of the best and I don't mean that I mean in the best possible way you one of the best tinkerers in the arab world, meaning like understanding the fundamental fundamentals of how something works by building it from scratch and playing with it with a bit. intuitions psych einstein. Finally, were all really good at this kind of stuff like small example, the thing to do to play with it to try to undo It also that in obviously now with us eu help build a team. of machine learning like engineers and the system that actually accomplishes something in the real world. given all that, like what was the soul searching like his heart, because, obviously I love

company lot and I love I love elon. I love tesla. I want to do it so hard to leave. I love the team. Basically a butt, yeah, I think. Actually, I will put a potentially like interestin revisiting. It me coming back at some point, are working documents or can enjoy tesla? I think that's not going to do incredible things? It's basically like it's a massive large scale robotics company with a ton of in house talent for doing really credible things, and I think our human, a robots are going to be amazing I think our thomas transportation is going to be amazing, oldest happening at us law, so I think it's just that really amazing organization, so being part of it and helping it along. I think, was very sandro that alot yeah. It was basically, for those reasons, because I love the company, but I'm happy to potentially at something come back for act too, but I felt like at this stage I built a team at felt tons.

And I became a manager, and I want to do a lot more technical stuff. I wanted. learn stuff. I want to teach stuff, and I just kind of felt like it was a good time for a for a change of pace a little bit. What do you think is the best movie sequel of all time, speaking of part, two as a muslim suck, but he was his movie there and you tweet about moisture, isn't it any tangent deserves what it was like. A favorite movies, equal godfather part, two famagusta, because you didn't even tweet or mention the godfather you, I don't know We know it has an agenda that out we're gonna, get it out the hate towards the got about. How dare you do things? I don't make a strong but I dont know I I I don't know why. But I basically don't like any movie. before nineteen. Ninety by some point that didn't you mentioned terminator to ok. Ok, that's like her tomato too, was covered later. Nine andy no

May I too was in there now exterminator one as well, so it s a like few exceptions, but I large, for some reason. I dont like movies before ninety ninety five or something they very slow, the camera, Like doubt, it's boring its kindly it's kind of weird and also to me. I was very much ahead of its time. Yes, and the godfather does look no e g. I saw this mean, but you have good will. Hunting was one in the movies you mention it doesn't have an asia. I either august as mathematics yeah, I guess occasionally I do injuries to pump feature or like anchor man that has no that's the tanker event is so good and understand. I'm you speak of asia. I guess I don't know I warfare, oh so funny it doesn't make sense, it doesn't computers, something about him and he's singular human lives, you don't get them any comedies these days and I wonder for us to do about the culture.

Or the wake the machine of hollywood or deserve to do, or just we got lucky with certain people in comedy they came together causes, single human noise in the thousand ridiculous tangent. I apologize but you mentioned human or robot. What do you think about artemus about us about? Do you think will have robots? In fact, and in the home and ten twenty thirty forty fifty years, I think very hard project it's going to take a while, but who also going to build Juno robotic scale. But I think it is a very good for factor to go after because commissioner? You the world, designed for human form factor these things would be able to. our machines. They would be able to sit down in chairs a dry but actually been dr course. Basically the role does. Is I for humans best of four factory? You want to invest into and make work overtime. I think there is another school of thought which is ok, pick a problem and design robot to it, but actually design

robot and getting a whole data engine and everything behind it to work is actually incredibly hard problem, so it makes sense to go after general interfaces that are or can be. They are not, perfect for anyone given task, but they actually have the generality of just for the prompt with english able to do something across I think it makes a lot of sense to go after a general, ah interface and in the physical world, and I think it's a very difficult project is going to take time, but I seen no other, no other company that can execute them and vision. Ai is going to be amazing, like Ah, basically, physical labour, like if you think transportation is large market traffic labour saying what by scientists just labour to me. The thing that's also exciting. Is the solution about it so that the relationship will have on different levels with those robots that I was really excited to see happen. people have size me the excitement, but I've I've workers

A lot of research, labs, adieu, humanoid, legged, robots, bosnia, makes unitary allowed. There's a lot of companies that do legged robots, but that's the the elegance of the movement is a tiny, tiny part of the big picture, so integrating the two biggest things to me about tests doing humanoid when you legged robots is clearly integrating into data engine soda the data engine aspect, so the actual intelligence for the perception. In the control and and the planning and all that kind of stuff integrating into the shoot the fleet that you mentioned right and then speaking of fleet. The second thing is the mass manufacturers. NOLA cultural lee, driving towards a simple robot. As cheap to produce

at scale yeah and doing that well, having experience to do that. Well, that changes everything. That's why that's a very different culture and style the boston dynamics who, by the way, does those robots are just the way they move. It's like you'll be a very long time for tussock and achieve the smoothness of movement. But that's not what it's about it's it's about it's about the entirety of the system. I could talk about the data and join the friday. That's super exciting, even then what sort of models, but that too was really surprising that in a few months you can get a prototype type and the reason that happened very quickly, as, as you alluded to there's a ton of our copy based from what's happening on the other politics. A lot. The amount of expertise that black came out of the woodwork set tesla for building document robot was incredible. Dizzy, like MR even said that one point we're doing this and then next day, basically like all these cad models started to appear and people talking about like the supply chain and manufacturing and people showed up

I like drivers and everything like the other day and starts like put together the body like whoa, like all these people exist that tesla and fundamentally building a car is actually not that different from building a robot the same, and that is true not just for the hardware pieces and also, let's not forget hardware, not just for demo, but manufacturing of that hardware at scale is a whole different thing, but for software as well, basically, this Robot currently thinks it's a car said. I have a mid life crisis. Some it thinks it's a car. Some of the earlier demos actually we're talking about potentially doing them outside in the parking lot, because that's where all of the computer vision that was like working out of the box that, instead of like an inside but in all the operating system, everything just copy peace. computer vision, mostly copy paste. I mean you have to retrain to neural nuts, but the approach and everything and data engine and offline trackers and the way we go about the occupancy tracker and so on everything copy paste. You just need to

drinking your lot and then they are planning. Control, of course, has to change what of it, but there's a ton of copy paste from what's happening, tesla as if you were to if you were to govern goal of like ok, let's build a million human robots and you're, not tesla. That's it! That's a lot to ask if you're tesla, that's actually like it's, not it's not that crazy than the five questions and how difficult year cycle driving how difficult as the manipulation task that it can have an impact at scale. I think, depending on the cap, tax. The really nice thing about robotics is that unless you do manufacture in that kind of stuff. Is there moreover, air driving associated critical of the end of the tunnel. Like a robot is allowed to move slower, which is nice. Yes, I think it's going to take a long time by the way, the structure of the development is he in to say: ok is going to take a long time. How can I set up the product, develop and wrote,

map, so they are making revenue along the way. I'm not setting myself up for a zero one loss function where it doesn't work until it works. You don't want to be in that position. You want to make it useful, almost immediately and then you want to slowly deployed ends. at scale I at scale, and you want to see you're dating junior improvement loops at the time a tree, the evaluation, the harness and everything and everyone improve the product of a time incrementally and you're. Making revenue belonged away way, that's extremely important, because otherwise you cannot believe these large undertakings, just like don't make sense economically and also from the point of view of it. I'm working on it. They need the dopamine along the way there not just going to make a promise about this being useful This is going to change the world and ten years when it works, is not what we want to be. You want to be in a place like I think, are published today when offering increased safety and and the convenience of driving. Today, people pay for people like it before purchase it

and then you also have the greater mission that you're working towards an you see that so the doping you for the team, that that was a source of happiness and yes and unsatisfactory, are deploying. Thus people like people drive it people pay for it. They care about it, there's all easy to videos, your grandma drugs it she gives you feedback, people like it. We will engage with it. You engaged with the huge do people that drives us. Those like recognising give you love like, like hey thanks for the help. for the fearless nice feature that is doing yakking. The tricky thing is like some people really love you. Some people, unfortunately, like you, are working on something that you think is extremely valuable, useful, etc. Some people do hate you there's not a lot of people who, like hate me and the team, and whatever the whole project, and I think they have the drivers. Many cases are not. We actually yeah that their social mission sanborn humans or the current the ways the humans, interacting thanks. You fixable, I think he was wont to be good to each other, adding twitter and social me as per

the mechanism that actually somehow makes the now italy more viral. Let it does deserve, like disproportionately add of like a fire viral boost young negativity, but I oh, I wish we could get excited about suppress summit jealousy, some of the ego and just get excited for others, and then there's a car aspect too that you get excited for others, they'll get excited for you sinking and academia. You think I care for them like a dynamic system there, if you, if you think of insight I was in good, jealous somebody else being successful that I see practical, intuitively the least less productivity If you as a community- and you individually, I feel like, if you give celebrating others that actually makes you more successful yeah. I think people haven't depending on the industry, having quite learn that yeah yeah, something also very negative and very vocal, so they're, very prominently featured, but actually there's a tonne

people who are our children, but their silent, chile, cheerleaders agenda. When you talk to people and just in the world, they will tell you it's amazing scrape especially like people who understand how difficult it is to get the stuff working. What people have built products makers entrepreneur. Entrepreneurs like make me making this work and changing something is incredibly hard. Those people are more likely the you for one thing that makes me sad as some folks in europe this community, don't do the cheerleading and they shed tears. because I know how difficult it was actually sometimes don't know how difficult it is to create a product that scale radio they actually why, in the real world of a lot of the development robots and asked this done and very specific small benchmarks, and as yet, with the real work additions. Yes,. Yeah. I think it's really hard to work on robotics in academic setting or I systems in the real world. You you ve, criticized you flourished and loved

for a time the image that the famed image that day said recently had some words of criticism that the academic research, em. All community gives a little too much love still to the image that are, those kinds of benchmarking can speak to the strengths and weaknesses of data sets used and machine learning, research, I've, gotta, I recall the specific instance where I was unhappy criticising the minute I think eminent has been extremely valuable. It was basically a benchmark that allowed the deep learning community too, the strength that did not work, actually work, It was a civilian that so I think imaging useful, but basically its beak, a bit of an amnesty. At this point, so I missed is liable to twenty eight by twenty eight greece, gale digits does kind of a joke dataset that everyone like just crushes who is no papers written on amnesty right Maybe I should have strong gabriel, like our papers, a no vote,

it's unlike. How do we learn with a small amount of data that? Yes, the yacht could see that being helpful, but not in sort of like mainline, computer vision, research anymore, of course, I think, the way I've heard you somewhere, maybe just imagining things, but I think you said, The commission was a huge contribution to the community for a long time, and now it's time to move past us cousin. Well, I miss my husband. I mean you know. The error rates are ah yeah, we're getting what many percent accuracy in in one thousand classification way prediction and I've seen those images and dislike really high. That's really that's really good. If I'm are correctly, the top five error rate is now like one percent or something given your ex and with a gigantic we'll work dataset. Would you like to see benchmarks? Certain directions that community uses. Unfortunately, I think economics currently have the next minute. We obviously I think we ve crushed m. This

we basically kind of crushed a minute and there's no next serve big benchmark that the entire community railways behind and uses you know for further development of these networks, I wonder what it takes for data set to captivate imagination already like where they all get behind it, that that could also need a virus like a leader wouldn't right down somebody, popularity, I up what do you imagine that take off their the it just the accident of history? It was the right amount of difficult. It was right mind of difficult and simple and interesting enough. It just kind of like it was. It was the right time for that kind of a dataset question from ready. What are your thoughts on the role the synthetic data and game engines will play in the future of neural net model development? I think, as neuro. Let's converge to humans. The value of simulation to know. What's will be similar to that

you have simulation to humans sophie. You simulation fora people use emulation because they can learn something in that kind of a system. without having to actual experience it. But are you referring this amazing doing our head No sorry simulation, I mean like video games or in any other Forms of simulation fork, various professionals. So let me pushed back and I guess maybe there's emotional we do in our heads like. Similarly, if I do this, what do I think will happen I'll get a second simulation. Internal is now overdoing, as he assumed before we act viability and from like the use of simulation and assess alike, computer games or using to listen for training secretion. Nor is it dependent or is it just loosely correlative because I got is not useful. You like on a factual or like edge gaze I should like you know happens if there's a nuclear war, what happens,

There's you like that those kinds of things the others are different simulation from, but unreal engine. But that's how I interpreted the question. Ah, so, like simulation of the average case- is that what's unrelenting what what what what what he knew by unreal engines oak, Similarly, a world you have it that world. Why that different, like as you also add behavior to that world and you could try all kinds of stuff right? You could throw all kinds, we think they it he's. A real engine is not just about similar guess it is about making the physics of the world is also doing something with that. Yet the graphics, the physics and the agents that you put into the environment and stuff like tat ass. He I think you see, I feel, like you, said that that important, I guess, if over the future of a development that is, that correct interval, you know, I think the humans use simulators for

you monsieur simulators, undefined them useful and so computers you seem simulators and find them useful. Ok, see, you think, I don't use. Symbol is very often the play a video game every once in a while, but everything derive any wisdom about my own existence from from those video games, this a momentary escape from reality versus a source of wisdom about reality, so not decisiveness. very polite way of saying simulations, not that useful, maybe maybe not I dont- see- does like a fundamental. It really important part. For training? No, let's currently, ah, but I think as europe must become more more powerful. I think you will meet fewer examples too. An additional behaviors and, simulation is of course, the domain gap in a simulation that is not the real world a slightly something different, but with a powerful left neural net. You need and the domingo can be bigger. I think, because neural I also understand that, even though is not the real world it like has all this. How structure that I'm supposed to be hidden-

well from so then you'll know will actually yeah we'll be able to leverage the data better. Yes, closing the gap but understanding in which ways this is real data. Rather do better questions next time. That was, that was the question, but I'm just kidding alright, so is it possible? Do you think, speaking of amnesty to construct, neural nets and training processes that require very little data So we ve been talking to a huge days as the engine of their training and yet one You say that, like you said like declaring itself as another level of training. I guess, and that requires little data, but dues see any value in doing research and tat going down the direction of carriers, very little data to train to construct an archway under percent. I just think like at some point. You need a mass of data.

and then, when you pray, train your massive neural net and get something that you know it is like a jpg or something then able to be very efficient at training and your return. You task also a lot of the activities in or you can do, tasks like sentiment, analysis or translation or song. Just I've been prompted with very few examples here's the kind of thing I want you to do like here's, an input sentence. Here's the translation into german input sentence, translation to german input sentence blank and the neural net will complete the translations german. Just by looking at serve the example you ve, provided as things for that very few shot. Learning in the activation of the neural net history, the weight of the neural net, and so I think, come mysqli just like humans. Neural nets will become very data efficient at learning any other new task, but at some point you need a massive dataset to retrain your network. To get that, probably we humans have some point.

That do. We do have some like that do ever passive in the background back our model constructing. thing. They just runs all time itself supervise way, we're not conscious of it. I think he must definitely Obviously we have. we learn a lot during during our life span, but also we have a ton of hardware that helps us initial at initialization, coming from sort of evolution, and so I think, that's also really big big component, a lot of people in the field. I think they just talk about the months of like seconds and im in a person has left pretending that this is a double our asa sarcasm, nationalization of neural net, and it's not like you can look at a lot of What, for example, zebras Zebras get born and they see and they can run their zero train data in their lifespan. They can just do Somehow I have no idea how pollution found a way to encode these algorithms and he's neural net initialization are extremely good into eighty cds.

I have no idea how this works, but apparently possible, because here's a proof by existence there's something magical about go for a single cells organism that is born to the fore, for years the life I kind of like that that the reason we don't remember anything about the first few years of our life is that it's a really painful process like it very difficult, challenging training says, yeah like intellectually like it, maybe the arrow, why don't we remember? Any of that so crazy training going on in the event that the backup- maybe that's the our model training it is very painful and so is best for the system once a strain not remember, I was destructive. I think it's just like the hardware. Long term memory is just not fully developed share. I kind of like the first few years of

france is not actually like learning its brain maturing were born premature, then there is the theory along those lines because of the personnel and swung the brain, and so were premature and then the first few years, we're just stood the brains maturing and there's some learning eventually. My current view on what do you think, do you think neural nets can have long term memory like that approaches, something like humans. giving you a digital needs to be another matter, architecture on top to add, like a knowledge base that learns facts about them, I would not like us off yes, but I dont know, to what extent it will be explicitly constructed it might take on intuitive forms where you are telling the jpg like hey you have. You have a declarative memory bank to which you can store and retrieve data. And whenever you encounter some information that you find useful, just save it to your memory bank,

and here is an example of something you have retrieved and hastened to say it and ears. Are you load from it? You just say load whatever you teach it and tax in english and then it might learn to use a memory bank from from that also in has given europe. That is the architecture for the background model that that the basic and yeah everything else is just on top of that. Just not just the text right, it's your given it gadgets and gizmos so are you're teaching some kind of a special language by which we can it, can save arbitrary information and retrieve it at a later time and you're telling about these special tokens and how to arrange them to use these interfaces is like hey, you can use calculator. Here's I use it just do five, three plus for one equals and when equals is there, or a calculator, will actually read out the answer, and you don't have to calculate it yourself and you just like tell it in English. This might actually work. think in that sense, gotta was interesting. The demise system that is not just a language She throws at all in the same pile

images. Actions are that kind of stuff, that's basically we're moving towards yeah? I think so so gotta is, is very much a kitchen sink approach to like reinforcement, learning lots of different environments with a single fixed, a transformer, a model right, and I think very serve early result in that in that rome, but I think yeah, it's along the lines of what I think things will eventually look like right so this is the early days of a system that eventually look like this, like from a rigid reason, perspective you're, not super huge fan, think all these interface. That, like the look very different, I would want every to be normal into the same ab. I so, for example, screen pixels verse same guy Having looked different role, environments that a very different physics and drunk configuration sent appeal, susan wherever and your having some kind of special tokens for different game. So you can plug I'd rather the two most everything to a single interface said

the same to the neural net. That makes us says all gonna be pixel based pong in the end, I think sir. Ok, let me ask you by your own. so a wife allowed before I know you're the most productive in berlin. People in She may I what is a productive day in the life of andrei capacity? Look like what time do you wake up? You imagine some kind of dance between the average productive day and a perfect productive day. So the perfect product there's a thing which strive towards in the average is color would occur, converges to another mistakes in raw human eventualities and so on, so it will tell you wake up my morning, person I'm not a morning person, I'm a night owl for sure, stable or not so my stable, like a eight or nine, or something like that. You're my phd it was even later I used to go to sleep, usually a three I am, I think, the m hours are our precious. The marines

in time to work, because everyone is asleep at at eight. A m or seventy m east coast is awake. Ass already activity. There is already some text messages, whatever you're still happening gone like some news website and their stuff happening distracting at three m. Everything is totally quiet. And so you're not going to be bothered, and you have solid chunks of time to do the work, and so I like those periods night owl by default, and then I think, like productive time. Basically, what I like to do is you need a you need to like build some moments. On the problem without too much destruction and arm. You need to load your ram, your working memory with that problem, and then you need to be obsessed with it when you're taking shower when you're falling asleep You should be obsessed with the problem at its fully your memory and you're to wake up and work on it right there. So it is the skill it is this in a scale temple

scale of a single day or a couple of days a week a month, so I can't talk about one day, basically, isolation, because it it's a whole process where I want to get when I want to get productive in the problem, I feel like. I need a span a few days, work and really get in on that problem, and I dont have interrupted, and I'm going to justice be completely obsessed with that problem, and that's where I do most of my good words. You ve done a bunch of cool like little projects in a very short amount of time very quickly, so that their cries you just focusing on it there is plenty to load my working memory when the problem- and I need to be productive because there's always like a huge, fixed cost to approaching the problem like. I was struggling with this, for example, at us, not because I want to work unless more site project, but ok, you for need to figure out. Okay, I need to ssh into my cluster. I need to bring up a vs code editor, so I can like work on this. I need to I arranged some stupid error because of some reason I can not at a point where you can be just productive wait, you are facing barriers, and so it's about really removal

that barrier and you are able to go into the problem and you have the full problem, loading your memory and somehow avoid distractions of all different forms like news stories, emails, but also Reactions from other interesting projects that you previously worked out are currently working on. So on you wanna, really foxier mine, and I mean I can take some time off for distractions and in between, but I think it can be Much and I most of your day as certain spent on the problem and then in a light, coffee I had my morning routine. I look at some news, twitter, hacker news, wall, street journal, etc. She's great Basically you wake up. We have some coffee. Are you trying to get to work as quickly as possible? Do do taken his diet of of like what the hell's happening in the world? First, I am. I do find it interesting to know about the world. I dont know that its useful are good, but his part, my routine right now

do read through much amused articles and I want to be informed and I am suspicious of it. I am suspicious of the practice, but currently that's where I am Oh you mean suspicions about the positive effect of that practice. In your productivity in Europe, while being as I well being psychologically so on your ability to deepen, understand the world, because how much is it of information. You normally focused and deeply integrating adds loaded, distracting, you're interested up perfectly productive day for what, along with a stretcher time one session, do try to work and focused everything. A couple hours is one hours at thirty minutes. Ten minutes I can probably like a small few hours and then some breaks in between four like food and stuff and down yet, but I think, like it's too really hard to accumulate hours. I was using a tracker. They told me exactly how much time has been putting any one day and even on very productive day. I still spent We like six or eight hours there

just because there's so much padding commute talking to people, food, etc. There's like the cost of life, it just living and sustaining and homeostasis and is maintained, self ass, a human, is very high, and another seems to be a desire within human minded to are to participate in society that greece, there padding he has the most productive days of overhead is just completely from start to finish is turning out everything. and just as in there and then and then you could do more than six and eight hours. Is there some wisdom about what gives you Did you like days of long focus you're just like, Whenever I get upset about a problem, something just needs to work. Sumptuous needs to exist in his to exist. The needed, and so you able to deal with bugs and programming issues and technical issues and design decisions that turned out to be the wrong was able to think throng given. Given that you want, I think, exist yet mister

and then? I think to me also a big factor is you know our other humans are going to appreciate it? Are they going to like it? That's a big part. motivation, if I'm, if I'm helping humans- and they seem happy- they say nice things, tweet about it or whatever. That gives me pleasure, because something something useful so like you, do yourself, sharing it with the world, like will ya, think it harbours a blog posts or to video yeah. I was thinking about it. Like suppose I did all these things but did not share them. I don't think I would have the same amount of money. We should that I can build up you enjoy the feeling of other people. Gaining value, unhappiness from the stuff we ve created What about diet zur? I saw you played with inanimate faster you fast that helped with everything the players with the things you play what's been most beneficial to the your ability to mentally focus on the thing and just meant the mental productivity and happiness is too fast.

so fast, but I do intermittent fasting, but really what it means at the of the days I skipped breakfast near so I do eighteen six roughly by default when I'm in my steady state, If I'm travelling or doing something else, I will break the rules, but in my study I do eighteen six, so I eat only from twelve to six, not a hard rule, I break it often, but that's my default and and yeah I've done a bunch of random experiments for the most part. Right now, where I've been for the last year and a half I want to say, is among plant based or plant forward I heard play forward its own standards inexact exactly no the differences, but it sounds better in my mind about it. Just means that I prefer plant based food and raw or cooked, or I prefer cooked and paste supplant based forgive me, I don't know she know how wide category of plant entails apply. It just means that you're not I'm not sure about, is anglo, can flux and are you just prefer to eat plants, and you know you're not making you're not trying to enforce other people.

and if someone is you come to someone's house party and they serve you state that there really proud of you will need it raised his eyes. Judgement on this beautiful, I mean that's underfoot cited. But I'm very flexible. He tried to one day. I have accidentally not consistently, but I simply add that I don't I don't like it. I think it makes me feel are not good it's too, too much of a head and are so currently have got to meals a day. Twelve and six I gotta do they're not far from doing you know, you'll, do it elo day, because it is interesting. It's an interesting feeling. Have you the longer than a day you I've, I've done a bunch of water, fasts gloves curious. What happens? What are anything instead I would say so I mean you know it's interesting that your hungry for two days and then starting date, three or so you're, not hungry it's like such a weird feeling, because you haven't eaten in a few days and on hungary that weird there's only what one of the many weird things about human biology figure, something out and fines.

another source of energy or some like that occur relaxes the system. Yeah, the body is like your hunger, your hungry and then just gives up its like. Ok, I guess we're fasting. Now, there's nothing and then this is gonna like focuses on trying to make you not hungry and even not feel does the damage of that and trying to give you some space to figure out the food situation. So are you still to this day most does the night I would say I am, but it is really hard to maintain my phd schedule hum aspect, when I was say working at us alonso on its non starter, also, but even now liking or people want to meet for various events. They said he lives in a certain period of time and you sort of have to like so that it's hard to like do a social thing, and then, after that, return to work yeah, it's just really hard. That's why I try. When I do social things, I try not to do to too much drinking, so I can return and continue doing warrior, but at the

says there is convergence and test a break aim? Any company their convergence towards a schedule. Or is there more is He was behave when the collaborate to learn about this, the aids. They try to impose a consistent, Actually you're all awake at the I'm going. Do try, create a routine and I could try to create a steady state in which a comfortable and also I have a morning routine. I have day routine. I try to keep things to a steady state and do things are predictable and then you can sort of just like your body just sort of like sticks to that, and if you tried to stress that old much of our creator when you're travelling and you're dealing with jet lag you're not able to really ascend to you know where you need to go. He has really worked humans with a habit sister. Would you thought work life balance throughout a human lifetime, so tat same part was known for pushing people to their limits and terms. able to do in terms of what the

trying to do in terms of how much they work or the cost of their own. I I will say just like us all too much backdrop for this, because what's happening is tat. Lies that its burst environment also, I would the baseline. My only point of reference is google black, where I've intern three times and I saw what it's like inside Google and Andy mind. I would say the baseline is higher than that, but then there is up trade equilibrium where once in a while there is a fire Someone like me who worked really hard, and so it happened, spiky and thirsty, and then all the stories get collected upon the first and then it gives the appearance of like totally sanity but actually assisted the more intense environment, and there are fires and sprints. And so I think you know indefinitely though I I would say it's a more intensive armand than something get a clear in your personal forget. All that you see on personal life. Will you think about

the happiness of a human being or a brilliant person like yourself above finding a balance between work and life or is such a thing is not a good thought. Experiment yeah, I think I think balance is good, but I also love to have sprints that are out of distribution and that's what I think I've been pretty. Ah, creative and as well sprints out of distribution means that most of the time, you have a clinical balance. A balance must have. and then I obsessed with something once in a while most unwise what once a week once a month once a year gap, probably like say once a month or something here, will we get an you get how briefer for money That's when you like really care about a problem. It must exist as will be awesome. Europe's obsessed with it and now, you can t do it on that day, you need to pay the fixed costs of getting into the group and then, stay there for a while and then society

and they will try to mess with you and they will try to distract you. You had. The worst thing is like a like I just need five minutes of your time yeah. This is the cost of that is not five minutes and society needs to change how it thinks about just five minutes of your time. It is never never just one minute, just curious, adjust the quickest idea. Why are you being so yeah? No, what's your computer set up what was like the perfect gee, I? U somebody that's flexible to no matter what laptop for screens Yeah or do you prefer a certain set up that the most productive? I guess we are familiar with? Is one large screen, twenty seven inch hum and my laptop on aside what operating system? I do max. That's my primary for all tasks. I it was ex, but when you working on deep learning everything as linux, your estimation into a cluster and you're working remotely, but what about their showed

like thou, using I d say I would use- are, I think, a good way as you just run bs code, my favorite right now on your back you are actually a remote folder throughout his age, so the actual Fathers are manipulating on the clusters world. So what's the best idee vs code, whereas the people, so I use emacs still go ahead. It may be cool out an office maximum productivity So why would you recommend, in terms of editors, it you're a lot of software engineers, editors for python, supposed miss machine learning applications. I think the current answers vs good. Currently, I believe that the best arm, I d, it's got a huge amount of extensions. It has again copilot and on integration, which I think invaluable. What do you think about that the capacity integration as actually I get to talk a bunch with good reason, was acquitted

I thought, and he loves copilot you, like programs, a lot with it yeah do you yeah, he's copilot. I love it and it's free for me, but I will pay for it yeah. I think it's very good and the utility that I with it. What is it is it? I would say there is a learning curve and you need to figure out when its hopeful and went to pay attention to its output and we're not going to be helpful where we should not pay attention to it. because if you just reading at suggestions all the time, it's not a good way of interacting with it, but I think I was able to her like mould myself to it. I found it very helpful number one in copy paste and replace some parts. So I dont when the pattern is clear, is really good at completing the pattern and number two. Sometimes it suggests api that I am not aware of. It tells you about something that you didn't know so and that's an opportunity to discover and use an opportunity to sue. I would never take copilot code as given I almost always copy a couple of years into while searching you see what this function is doing and then you'll live is actually actually exactly when I need. Thank you go pilots. Are you learn something sets in part

search engine power may be getting the exact syntax correctly. There you see a gap is set and be hard things. once you see it, you know yes, exact, correct excitement, you yourself you Derek dairy, you can verify efficiently by Yuki chan generate efficiently and copilot. Really I mean it's autopilot for programming right, currently doing doubling following which is like the simple copy, paste and sometimes suggest, but over time going to become more more autonomous, and so the same thing will play out in just coating, but actually across many many different things, probably coatings and important one right, very writing programmes. What You see the future that developing the programmes this is like being able to write programmes more and more complicated three now its human. supervised in the interesting ways, just like what like the transitional, be very painful. My mental model, for it is the same we'll happiness with the auto pilot, also

currently is following is doing some simple stuff. And eventually will be doing autonomy and people will have to intervene less and less in those could be like you like. lasting mechanisms that give a right function and their function looks pretty damn correct, but honey are correct. Cause you're like getting laser lasers a programmer her my gear, billy too. a little bugs, but I guess it won't make a little known as well said it. A copilot will make our off by one subtle box. It has done that to me, but do you think future systems will Is it really? The off by one is actually a fundamental challenge of programming, and in that case it was fundamental, and I think that things can improve, but I think he must have to supervise. I'm nervous about people, not supervising what comes out and happens to, for example, the proliferation of bugs in all of our systems? I'm nervous about that, but I think there are probably be some other co pilot for, but finding and stuff like that

some point has little be like a lot more automation for man. If they got a programme, a co pilot, the generates a compiler for one that does a linear? Yes, one that does like a type checker yeah, it's Many of like a jpg serve like and then I'll be like a manager for the committee, and then there somebody that says a new version of this is need me to regenerate it. There are ten jupiter, There are four didn't get. Fifty suggestions and other one looked at it and picked a few that they like a bug. One looked at it and it is like it's probably a bug the governor rank by some other thing and then a final ensemble. becomes areas like ok, given everything you guys have told me. This is probably the next token stealing is the number of programmes in the world has been growing and growing very quickly. Do you possible that actually level and drop to like a very low number of this kind of.

Causing you be doing software to point or programming and you'll be doing this kind of generation of corporate tax system. programming, but you won't be doing the old school song software. One. Our programme? I dont currently think that you're just going to replace him programmers hum. It's so hesitant, saying stuff like this right, because it is this is gonna, be replaced in five years, and I know it's going to show that like this is where we thought that I could say I agree with you, but I think we might be very surprised. Re like what are the next at and what's your son, where we stand with language models like? Does it feel like the beginning or the middle or the end, the beginning hundred per cent? I think the big question my mind is for sure deputy will be able to program quite well, competently and so on. How do you steer the system? You still have to provide some guidance to what you are

you're looking for, and so how do you steer it? And how do you say? How do you talk to it? How do you um audited and verify that what is done is correct, and how do you like work with this and its not just as a problem, but are you? Are you ex problem near so beautiful fertile ground for so much interesting work for vs code, plus plus one you're, not just it's, not just human programming anymore. It's amazing! Yes, interacting with the systems, are not just one prompt but iterative, prompting young trying to figure out having a conversation with this is the I mean to me that super exciting to have a conversation with the program running maybe at some point, you're just conversing with its like. Ok, here's what I wanna do. Actually this variable, maybe that low levels variable button, you can also, Imagine like. Can you translate this? The super spaz?

back to python already kind of existence, not just like doing it as part of the programme experience like I think, I'd like to write this function is suppose bus or like you just keep changing for different from programmes, is a different subject. Syntax, maybe I want to commend the central functionally Would I, like you, get to become multi lingual programmer and dares back and forth efficiently I think the u r e x- u x, of it, though, is like still very hard to think through, because it's not just about writing code on a page. You have an entire developer environment. You have a bunch of hardware on it you have some environmental variables here, some scripts that are running in the chrome job like there's lot going onto like working with computers and How do these systems set up environment flags and work across multiple machines, That's concessions and automate different processes like how that works and is article by humans and so on is like massive question a moment you ve, built archive sandy. What is archive

and what is the future of academic research publishing that you would like to see us archive as this preference over. So if you have a paper, You can submit for publication. Two generals were conferences and then wait six months and then maybe get a decision, passer fail or you can just upload it to archive, and then people can tweet about it. Three minutes later and then everyone sees it. Everyone reads it and everyone can profit from it in their own way sneak inside it? and it has an official look to it. It feels like a pump like it feels like a publication process, yet it feels differently. If you just put in a blog post, oh yeah, Yeah I mean it's a paper and usually does the bar as higher for something that you would expect on archived, as opposed to an something you had seen: a blog post with a culture created, the bar click. It probably yes, host a pretty crap. You fix burner cape though Was that make you feel like what was I make you feel about peer review, so rigour peer review by two three experts verses

review of the community right, as is written yeah Fifthly, I think the community is very well able to peer review things very quickly on twitter, and I think maybe just has to do something with a machine learning fields. Specifically, though I feel like things are more easily adaptable and had the verification, as is easier potentially than the verification somewhere else. So it's kind of like come, you can think of these centres. obligations is like will block change where I was building on each other's work inside each other, and you serve have a high, which is kind of like this much faster and loose blockchain, but then you, and any one individual countries like very, very cheap to make, and then you have other feel, or maybe that model does make much sense, so I think in a I at least, things are pretty easily verifiable as that's why, when people upload papers are a very good idea and so on, are people can try out the next day.

It can be the final arbiter of whether it works or not on their problem and the whole thing just move significantly faster. So I kind of like academia still has a place. Certain dislike conference journal process over the place, but it's sort of like an m. It lacks behind. I think its a bit more may be higher quality process, but it's not sir, a place where you will discover cutting edge work anymore. It used to be the case when I was starting my phd. Let you go to conferences and journals, and you discuss all the latest research now when you go to conference, our general like knowing discusses anything. They are because it's already like three generations ago irrelevant in which makes me sad, abolished, mind, for example, where they still less to publish in nature and these big prestige I mean they're still value, I suppose supposes the prestige that comes with these venues, but there there result. That day, there were now some breakthrough performance take got a year, your actual the details I mean and those details inside

the polish immediately witnessed by the community to monitor and direct with their young would speed up the rest of the community, but I dont know to what extent that's part of their objective function also assure does not just the pretty A little bit of the delay is as partners Certainly did my specifically has been none. Working in the regime of having like a higher quality, basically price, and latency and punishing those papers. That way. The question from red it. Do you or have you suffered from impostor syndrome being there java, Tessa, being dispersed? when you're at Stanford were like the world, looks it you're the expert in aid to teach teach the world about machine learning. When I was leaving to slaughter five years, I spent a ton of time in meeting rooms and I would repeat purse in the beginning, but I don't know I was writing code and then

is writing less and less code and I was reading code and then I was reading less and less code, and so this is just a natural progression. That happens. I think- and definitely I would say near the tail end. That's when it's sort of like starts to hit you, but more of that you're supposed to be an expert, but actually the source of truth is the code that people are writing to get up and for the actual code itself and you're, not as familiar with that as he used to be and so I would say maybe some like insecurity there and I think you're pretty profound there. The insecurity has to do with not rain. The code in the gdr cyberspace agree, because that is the truth that, at that rate, the is a source of truth, the papers and everything else. It's highlight summary I dont It's a high level summary, but at the end of the day you have to read code, it's impossible to translate all that code into actual. You know paper form or so, when when things come out the special endeavour source code available. That's my favorite place to go so, like I said you one of the greatest teachers of machine learning I ever from

he has to thirty one and two today. What advice would you give to beginners interested in getting into missionary? Beginners are often focused online, What to do- and I think the focus should be more like how much you do so. I am like believer on a high level and dessert ten thousand hours kind of concept, where you just kind of have to just pick the things where you can spend time and you you care about, and you're interested in. You literally have to put in ten thousand hours of work. It does like matter as much like where you put it in your you'll iterate annual, improve and your way. Some time I don't know, there's a better way. You need to putting two thousand hours but I think it's actually really nice cause. I feel like there's some sense of determinism about being an expert. At one thing, if you spend ten thousand, hours. You can literally pick an arbitrary thing and I think if you spend ten thousand hours of deliberate effort and work, you actually will become an expert at it as I think it's kind of like a nice thought and so on instead, I would focus more, unlike our eu spending. Ten thousand hours focus on sit. So

and then thinking about what kind of mechanisms maximum as your likelihood of getting to ten hours, exactly which for us silly humans mean probably forming a daily habit of arguments. what they actually doing. The thing, whatever hopes you side to think to a large extent, the cycle legal problem for yourself the thing that I hope that I think is helpful for the psychology of it is many times people compared themselves to others. In this area I think, is a very harmful, Only compare yourself to you from some time ago. Alexei yoga are you? Then you year is the only way to think- and I think this, then you can see your progress and is very motivating. That's so interesting dead, focus on the quantity of ours. They think a lot of people in a beginner stage, but she throughout paralyzed. By the choice which one topic. This path of this path me up, I feel a little get paralyzed by like which idea,

it is, for their worries yacht bill, worried about all these things. But the thing is that some of you, you will waste time doing something wrong eventually figure out: it's not right. You will acumen scarred tissue and next time, you'll grow stronger, because next time you had the scar tissue and next time, you'll learn from it and next time you come to a similar situation, a like I, I messed up. I've spent a lot of time working on things that never materialised into anything, and I haven't cartouche tissue, and I have some intuitions about what was useful with wasn't useful, how things turned out all those mistakes where we're not dead work. You know, so I just think you should You spoke of some working. What have you done? One of your last week? That's a good question. I should ask for for a lot of things that is. Machine learning is a good way cut the weapon the toy used, but the flood, the blubber, wherever there are the innocent This is a life.

What do you love about teaching? You seem to find yourself often in the past. I draw the teaching a very good debate, I also join too. I mean, I don't think I love teaching. I love happy humans. And how can humans like when I teach I I wouldn't say I hate teaching or tolerate teaching, but it's not like the act of teaching that I like it's it's that you know I I have some. I have something I'm actually okay, added. Yes, I'm ok, teaching and people appreciate a lot and so I'm just happy to try to be helpful and looking itself is not like the most amazing. in the pits, can be really annoying. Frustrating has written a bunch of luck just now. I was reminded back to my days of through thirty one and just how much work it is to create some of these materials make them good. Adoration and thought, and you go down blind alleys and just how much change it so creating something good

in terms of how educational values really hard and it's not fun. It was difficult soap of peoples who doesn't go watching news that you but their lectures we actually building a thing like fun? Like you said, the code is truth so discussing back propaganda by building it by looking through and the whole thing. So how of course, that the prepare for, I think it's a really powerful way to teach a hot did. You have to prepare for that or are you just live thinking through it? I will typically do like say three takes, and then I take like you better take are so I do multiple takes now. Take some the better takes in the niger delta electra. That way, some I have to delete thirty minutes of content, because one down the knowledge that I didn't like too much. There's a bunch of integration and probably takes me you know somewhere around ten hours, to create one hour of content to do when our its interest I mean, as is a difficult to go back to the bank, the basics draw, a lot of like wisdom from go back to the basics, ya, goin back to back propagation last functions where they come from and what they like.

teaching a lot honestly, as it definitely strengthens your understanding? Also not a purely altruistic activity. It's a way to learn. If you have to explain something to someone, you realize you have gaps in knowledge, and so I m surprised myself in those letters like the result will obviously look at this and then the result doesn't look like it. None like ok. I thought I understood this yeah, but that's why it's For this they will literally code. You run it and a notebook and gives your resulting you like all actual numbers actual employed at the actual costs, not mathematical symbols, etc. The sort of truth is the code. It's not slides, it's just like us build it. Beautiful your rare human. In essence, what advice would you give to researchers tried to develop and pulled the idea that a big impact in the world by some b on the grass maybe early graduate students

I would say, like they definitely have to be a little bit more strategic. Then I had to be as a phd student because, of the way ice, evolving, It's going the way of physics where you know if, as it used to be able to do payments in your bench? Stop and everything was great and you can make progress, and now you have to work in like Lfc or like sermon and so the eyes going in that direction as well. So certain kinds of things is just not possible to do on a bench stop anymore and I think come that didn't used to be the case. The time you still think that there's ghana type papers to be written world. Like very simple idea, It requires just one computer to illustrate a simple example mean one example: that's been very influential recently is diffusion models, the future. Those are amazing. The fusion models are six years old for a long time, people are kind of ignoring them the control and there an amazing generative model, actually in love and images and so stable before. and so on its old diffusion based defused,

As knew it was not there and came from where it came from Google, but a researcher could have come up with it. In fact, some of the first actually know those came from Google as well, but a researcher could come up with that in in the constitution yeah. What do you find most fascinating about diffusion models so far from the societal impact to the to the technical architecture? What I like about the fusion as it works so well surprising to the amount of the variety almost the novelty of the synthetic data is generating. Yes, stable diffusion images are incredible: it's the speed of imports. went in generating images has been insane. Ah, we wonder quickly from generating like tiny digits, two tiny faces and it all look messed up. Another stable diffusion, and that happened very quickly. Not that I can make us will contribute. I know that, for example, on flash attention is a very efficient colonel for running the attention operation inside a transformer that came from an economic environment. It's a very clever way to structure the colonel.

That can do it ass, a calculation, so it doesn't materialise, detention, matrix and so there's. fingers still like lots of things to contribute, but you have to beaches more strategic You think neural networks can be made to reason. Ah, yes, do you think there are already reason? Yes, what's your definition of reasoning or information, processing or the way the humans think through a problem and come up with novel ideas It feels like a reasoning so that the novelty added. I don't want distribution ideas, things possible, yes overseen that already in the current year on what's yours, to remix, that training set information into true generalization in some sense that doesn't it. Here. It doesn't get here on the motorway interchanging set like you're doing something interesting algorithmic, let your manipulating you know

symbols and you're coming up with some correct, a unique answer in a new seller. What would illustrate to you holy shit? This thing is definitely thinking. To me thinking or reasoning, is just information, processing and generalization, and I think the neural that's already do that today to be able to perceive the world I perceived the wherever the inputs are and to make our predict. Hence based on that are actually based on that. That's that's reason: you're, giving correct answers in novel settings by manipulating information. You ve learned a correct algorithm you're, not doing just some kind of look up table nearest neighbor search like them. Let me ask about asia. I work what are some one said ideas you think my make significant progress towards a

yea and maybe another weighs what a big blockers that we're missing now. So, basically, I am fairly bullish on our ability to build eg ice. Ah, basically, automated systems that we can interact with that are very human like and we can interact with them in a digital realm or physical realm. Currently, it seems most of the models that sort of do these sort of magical tasks are in a textural and I think are as I mentioned. Suspicious that text, rome is not enough to actually build full understanding of the world, I do actually think you need to go into pixels and understand the physical world and how it works. So I do think that we to extend these models to consume images and videos and train on a lot more data that has multiple in that way. If you to touch the world to understand, and also of the big

one question. I would say my mind is if you also require the embodiment and the ability to surf surf interact with the world, run experiments and have the data of that formed the need to go to optimise or and like that, and so I would say, optimists in some way is like a hedge and in a g. I because it seems to me that it's possible That's just having data from the internet is not enough, and that is Yes, then optimist mainly to enjoy because optimist what I to me and there's nothing be after miss. You have like this humanoid for a factor that can actually like do stuff in the world. You can imagine Some of them interacting with events and so on, and I doubt if there I don't give a rise to asia at some point, but not I'm not sure what we'll do so. From a completeness perspective, I think that's the that's a really good platform, but it's a much more harder platform because you are dealing with atoms and you need to actually like build these things and integrate them into society. So I think that path takes longer, but it's.

More certain and then there's a path of the internet and just like training. These compression models effectively on trent compressed all the internet. and that might also give its agents as well compressed easier, but also interact with the internet? It's not obvious to me. I suspect eating regime without ever entering the physical world the which is a little bit more concerning because it might that results in a happy faster. So feels like were again boiling water won't know, as is happening. I would like to add afraid of asia. I am excited about. there's always concerns, but I wouldn't to know it happens. oh have like hence about when it happens, like a from now on it will happen that kind of thing here. I just feel the judge, or am I just might happen. I think all we have available to us.

Because no one has built a g. I again, so all we have available to us is: is there enough for power ground on the periphery? I would say yes and we have the progress so far, which has very rapid and there are next steps there are available, and so I would say yes quite likely that will be interacting with digital is. How do you know the somebody's bodies It's going to be a slow, I think, is going to be a slow incremental transition is going to be product based, unfocused is going to be get up, cobalt, getting better and better deputies, helping your right and then these oracles, that you can two with mathematical problems. I think we're on the verge of me. well to ask very complex questions in chemistry, physics, map of these articles and have been complete solutions? so a joy to use primarily focus on in town so cautious, doesn't enter. Two into it. Send my my consciousness is not a special thank you. Will you will figure out and bolted

I think it's an emerging phenomenon of a large enough and complex enough. General model surf, so if you have a complex enough, a role model, that understands the world. Then it also understands its predicament and the world as being a language model which to me, is a form of conscious so self awareness, and so in order understandable, deeper, you probably have to integrate yourself into the world, gets an inert interact with humans and other living beings. Consciousness is a very useful yahoo. I think consciousness is like a modeling inside modeling incite yeah, it's a u powerful enough model of understanding the world that you actually understand that you are an entity in it ever. There is also this perhaps just a narrow If we tell ourselves if it feels like somebody to experience the world the hard problem of consciousness yeah, but that could be just the narrative, though tell retailers yeah. I dont think what I think it will emerge. I think it's going to be something very boring like we'll be talking,

These generally, I stable claim their conscious. They will appear conscious, they will all the things that you would expect of other humans and it's going to just be a stalemate, I think there be a lot of actual fascinating ethical questions, like Supreme court level, questions of whether you're allowed to turn off a cautious ay. I, if you allow to be the cautious ai on. Maybe there will have to be the same color. The basic have around sorry to bring up a political but but abortion, which is at the deep request. with abortion is what is life, and the deep question. They is also what is life and what is cautious and I think they'll be very fascinating. To bring up, he might become a legal to build systems that are capable that they, like,

of such level of intelligence, the consciousness will emerge and therefore the capacity to suffered emerge and somebody a system that says no, please don't kill me, but that's what the lander computing I'm not sure but already told this, the engineer ride like there was talking about not wanting, you die or so on. There might become illegal to do that. because otherwise, you might have a lot of a lot of creatures. I don't want to die and they will respond, If you have some other cluster and then there may lead to like horrible kafkaesque. Then there might be a lot of people they secretly, love, murder and they'll start pressing murderer nor systems may there. Just I did to me All of this stuff just brings before nearer, to the human condition and human Nature will get to explore and that's what like the best of the supreme court of all the different,

base. We have by ideas what it means to be human. We get to ask the questions that have been asking throughout human history. There has always been the other in human history where the good guys and ass the bad guys and we're going to you know throughout human history lesson. The bad guys- and this will probably happen with robots, it'll be the other at first and then we'll get to ask, What does it mean to be alive or isn't me to be cautious, and I think there, I'm canary nicole mice, even with what we have today and in all four of these there. These, like wife, who still like work then some people are trying to like this company is going to shut down This person really like love there. I flew and click estranged like port it somewhere else, unlike it's not possible and like, I think, like definitely, People will have feelings towards towards these some systems, because if some says they are like a mirror of humanity, because they are left her like a big average,

if humanity, in a way that it's trained it, but we can that average. We can actually watch this. It's nice to be able to interact with the big average of humanity and do like a search query on it. Yeah, that's very fast. name, and we cannot go of course also like shape it. It's not just a pure average. We can muster the train data we can best, but the object. We can find them in various ways. We have some impact on what those systems look like. If you once we achieve a g? I n, and you could I have a conversation with her and ask her talk about anything. Maybe ask a question: what what kind of stuff would you would you ask, what are some practical questions my mind, like? Ah, Do I or my loved ones really have to die? What can we do about that? Difficult clearly or would it answer politically? I would expect it to give solutions. I would expect to be like well I've,

all of these textbooks- and I know all these things that you have produced and it seems to me like yard the experiments that I think it would be useful to run next and hear some gene therapy is ethic would be helpful and here are the kinds of experiments that you should run. Ok, let's go was thought there again imagine that mortality is actually pretty. I got a prerequisite for happiness. so if we become a mortal will actually become deeply unhappy. the model was able to know that. So what is the supposed to tell you stupid human about Yes, you can become a mortal blow. You will become deeply unhappy if, as if the models it if the asia system is trying to empathize with you human- was this supposed to tell you that Yes, you you don't have to die, but you really not gonna like it is be deeply honest, like there's a interstellar. What is it the ai says like, humans, one ninety percent, honesty so like you have to

How honest I want to answer these court practical question? Yes, I love the idea. Interstellar, but away. I think it's like such a sidekick to the entire story, but at the same time is like the interesting. If current limited in certain ways right yeah and I think that's only find by the way. I don't think I think it's fine. possible to have a limited and imperfect aging ice is at the feature. Almost has simple like it has a fixed amount of compute on its physical body, and it might just be there Even though you can have a super amazing a brain superman. can I go can have like you know less intelligent. As they can deploy in a power efficient way and then they're, not perfect them mixed mistakes. No, I meant more alike, say you had infinite compute and is still good to make me Sometimes I, in order to integrate yourself I arm what does it come back? The goodwill

singer, robin williams character, says, like the human imperfections. That's stuff right. This isn't it is not like. We don't want perfect. We want flaws in part two to four connected with each other, because it feels like something you can attach your feelings too. The flaws in the same way you want to I that fly. I don't know I feel like perfectionist, but then you're, saying again business asia. I perceive asia. I would need to be intelligent enough to give answers to whom he was understand, and I think perfect is something you can understand, because even signs doesn't give proof Dancers is always gas and miss treason of humans want perfect. yeah, I can imagine just stop having conversation with this kind of oracle entity as you'd. Imagine them and you may be, it can tell you about it are based,

my analysis of human condition- you might not want this, and here somewhere things they might him might happen every day. Every dumb human will say yeah trust me I can and giving the truth I can handle it, but that's the beauty like people, countries are so important, the old marshmallow tests for the kids and so on. I should like to make people like you handles the truth problem, including myself, a deep true the human condition. I don't. I don't know if I can handle that. I quote: what is there some dark? Well, what have we are an alien science experiment realises that what if it had a meal I mean this is the matrix all over again I dunno I'd. I would what would I talk about? I don't even yet. I probably I will go with this- for scientific questions, the first they have nothing to do with my own personal life

immortality, just like the ball physics and so on ya did to build a bridge, see words that or maybe C4 has a sense of humor. That's another question: would it be able to presumably in order to if it understands humans deeply, would able to generate the degenerate. Humor yeah, I think that's actually a wonderful benchmark almost like is it able? I think it's a really big point, basically to argue laugh at the table to like a very effective stand up comedian. That is doing something very interesting. Computationally, I think, being funny is extremely hard because it's hard in away like a turing test, so? The original intent of the turing test is hard, because you have to convince humans and there's nothing. What that's what the comedians talk about this like there's that this is deeply honest. Because, if people can't help laugh, if they don't

have they measure not fine. They laugh that's funny, you're, showing you need a lot of knowledge to create decreed humor about that your commission, human condition and so on and then indeed be clever with it Why should a few movies you tweeted movies? I've seen five plus times, but am the unwilling to keep watching interstellar, gladiator contact, goodwill, hunting, the matrix or the rings are three avatar four elements on Gaza to me, too. Mean girls, and I get asked about the latter, his great. What are some of the german? to your memory, let you love, and why are you in the matrix is as a committee, person why you of the matrix, there's something properties and make it like beautiful interesting, so there's always fought off course us, but in those also edgy ice and theirs, relation and it's cool, and there are, as you know, the black, Are you know the look of it

if you look at the feel of it, the action the bullet time it was just like innovating in so many ways. And their good work. Goodwill hunting way like our yeah, just ivory, like this sub torture genius serve character, of whose, like grappling with whether or not he is like any responsibility or like what to do with the skiff he was given or like how to think about the whole thing. and are the also dance between the genius and the personal like It means the love another hurrying and there's a lot of things. There is just a beautiful movie and then the fatherly figure, the mentor and in the psychiatrist it like really like it must. with you in other some movies it just like really messed with you. Ah, on a deep level We relate to that movie at all. No, as I said, lord of the rings self explanatory terrier too, which is interesting he'll be washed out. There's about it intervened, one you

y'all are just like the terminator one as well. Ah I like terminator two a little bit more, but in terms of liquid surface properties, I do think Skynet is at all a possibility. Ah yes, like the actual sort of atomic weapons system. Kind of thing, do you know about that? So do I use a war, a hundred percent worry about it, and so the the fia in some of these fears of ages and how this will point out that these will be like very powerful entities, probably at some point, and so for one they're going to be tools in the hands of humus. Hyena people talk like alignment of asia eyes and how to make the problem is lucky humans are not aligned. Also, how this will be used and what this gonna look like a sum. Yes troubling, so did It happened so slowly enough, there will be able to act. There came a civilization think through the problems. Yes, that's my hope is that it happens slowly enough and an open enough way where

people, can see and participate in it. Just a figure How to deal with, This transition. I think we have some interesting. I draw lots inspiration from nuclear weapons. Causation. Sure thought be will be fought once they develop nuclear weapons, but like it's almost ache, one on the systems are not so dangerous. They destroy him. Civilization. We deploy them and learn lessons, and then we quickly, if it's too dangerous, will quickly quickly, whom I saw deployed, but you very quickly learn not to use them. So there'll be like this balanced attitude. He was a very clever. Ass, a species fishing, we exploit the resources as much as we can, but we don't. We avoid destroying ourselves since jack. While I dunno about that actually hopper continues I am definitely like concerned about nuclear weapons. Is so and not just as a result of it

in conflict even before that the problem, like my number one concern for facility, says if humanity destroys itself, or destroys zero. Ninety percent of young people. That would be because a nukes, I think so as not even about full destruction. It's bad enough. We reset society loverlike terrible her, be really and I can't believe we're like so close to it, yeah it's like so crazy to me. It feels like we may be a few tweets away from something like that: yup. Basically, it's extremely unnerving but and has been for me for a long time It seems unstable that world leaders, just in a bad mood can like take one step towards a bad direction. Escalates yeah because of a collection of bad moves, it can ask without being able to stop. Yes it's a huge amount of power and

and also with the proliferation of biscay. I don't. I really see. I don't know what the good comes our here, some difference where did by their lot and then each year is not only there, but I think at some point will more and more become something like it that the danger It's a jedi even is that I think it's even less likely worse in the sense that, ah there are good outcomes of a g. I and then the bad outcomes are like an epsilon away like a tiny wondering, and so think capitalism and humanity, and so on will dry for the positive ways of using the technology. But then it bad outcomes are just like tiny, what flip a minus sign away, now, that's a really bad a position to be in a tiny perturbation of the system resulting in the destruction of the human species. So galliard line to walk. Yeah, I think, and general was really weird, but like the dynamics of humanity in this explosion, we talked about it just like being saint coupling afforded by technology yeah and are just the stability of the whole dynamical system, an angel

listen, don't look at honestly That is how she could be destructive. A constructive in the probabilities are non zero in both gallons I'm gonna have to. I do feel like I have to try to be optimal and so on. And yes, I think, even in this case I still am predominantly optimistic. But there is definitely me do. think, will become a multiple entire species. Probably yes, but I dont know if its dominant feature of future humanity, there might be some with people on some planets and so on, but I'm not sure if it's like yeah, it's like major player in our culture and so on. We still have to solve the drivers of self destruction here on earth, so just having a back up on mars is not going to solve the problem. This, oh by the way, I love the back up on mars. I think that's amazing, absolutely did yes and I'm so. Thank you for an. Would you go to mars I personally know I do like earth quite a lot of good amongst other things are three yet you from,

eventually. I would unless it's safe enough, but I dont actually know if it's on my light timescale unless I can extended by lot. I do think that, for example, I might disappear into virtual realities and stuff, like that. I think that could be the major thrust of sort of the cultural development of humanity if it survives so it might not be it's just really hard to work in physics rome and go out there, and I think, ultimately, all your experiences are in your brain and so much easier to desert disappear into digital rome and I will find a more compelling, easier, safer, more interesting theory, little bit, captivated by virtual reality by the possible worlds where there is the matter for some other manifestation of their yeah yeah, really injured, it is- and I am interested just just talking a lot- the carmack where's, the where's, the thing that's currently preventing that yeah. I want to be clear. I think, what's interesting about the future, is it's

that I kind of you like the variants in the human condition gross. That's the primary thing. This changing its not as much demean of the distribution is like the variants of it. So there will probably be people mars and there will be people in we people here on earth is just like there'll, be so many more ways of being, and so can I feel like He doesn't like a spreading out of a human experience, there's something about them. china that allows you to discover those little groups and the gravity together, some by your biology likes that kind of world that you find each other and will have since you minutes and then we'll have the amish and they're gonna, everything is just gonna coexist with a modicum of interacting with a bunch of communities. Is they don't know, but about each other like you, can have a very happy existence, just like having a very closely not knowing about each other. We even even census she's having travelled to ukraine. This think they'll know so many things about. America do you like it when you travel across the world, I think you expenses to theirs.

Cultures there like they have thing on there on his there you can see that happening more and more and more and more in the future. We are a little communities, yeah yeah, I think so. That seems to be the things to be how it's going right now, and I don't see that trend like really reversing, I think people are diverse and there the trees down like path in existence, and I sort of like celebrate that ends We spend so much time in the members in the virtual reality. which community area I have the physical ass. They were. The physical reality enjoy air or d. He's wrong, a lot of pleasure and fulfilment in the digital world yeah. I think Currently, the virtual reality is not that compelling ass. I do think it can improve a lot, but I don't really know to what extent maybe in others, actually look even more exotic things. You can think about we're like neural link, sir stuff, like that, so

currently are gonna see myself is mostly a team human person. I love nature. I live harmony, I love people, I love humanity. Emotions of humanity, and I want to be like, and this last solar punk clothes utopia of my happy place here. My happy place is like people. I love thinking about cool problems, surrounded by a lush, beautiful, dynamic nature, and a secretly high tech in places that count places like the use technology to empower that love for other humans and nature, ya. Think that technology used like very sparingly I dont loveliness. Gets in the way of humanity in many ways. Ah, I, like just people being humans in the way we serve like slightly evolved and prefer, I think, just by before people kept Ass can make us they? They know you love reading are their particular books that you enjoy there had an impact on you for silly or for profound reasons, for you recommend

You mentioned the vital question. Many of course biology as an example, the vital question is a good one. Anything by maclean, really like a son, I would say, is like a bit more potentially representative is like a summary one of the things you ve been talking about, I was very impacted by the selfish jean. I thought that was really good book understand, altruism as an example and where it comes from interest realising that, actions and the love of geese was a huge inside for me at the time and its her. What cleared up? thanks for me, would you Think about the the idea. That ideas are the organisms, the media's lovett hundred percent? I have to walk around with that notion, for, while that. There is an evolutionary kind of process with ideas is world absolutely ass, theirs means just lodgings and they compete and they live in our brains dispute.

Are we silly humans? Thinking aware the organisms? Is it possible that the primary organisms are the ideas It would say, like the the ideas kind of live in the software of like our civilization in the in the binds and so on. We think as humans that the hardware is the fundamental thing. I Man is a hardware entity, but it could be the softer right. Yeah. Yeah, I would say like there needs to be some grounding at some point to look a physical reality, quote and Andre. The software is a thing like is this thing that makes that thing special. I yeah. I guess I you're right, but then cloning might be exceptionally difficult like there may be a integration between the software and the hardware- always we don't quite understand or from the ocean Have you like what makes me special is more like the the gang of genes that our writing in my chromosomes, I suppose right like there, that replicating unit

I suppose, and nobody, but is prolific prolife the thing that makes you special sure, while The reality is what makes you special is your ability to survive based on the software that runs on the hardware that was built by the genes, so the software is the thing that makes you survive. Not the hardware I get over both eyes. He is just like a second layer. It's a new second layer that hasn't been there before the brain. They both they both coexist, but there's also layers of the software. I mean it's it's not so jaxa, is on top of obstructions, but ok itself is so selfish gene and a nick lane, I would say sometimes a books are like not sufficient. I like to reach for textbooks sometimes, and I kind of feel Books are for too much of a general consumption some time and they just kind of like com.

High up in the level of attraction, and it's not good enough. So like textbooks, I like this, so I think the sound was pretty cool. This might also I like to read off the plane is because he's pretty willing to step one level down and he does entire gay sir. If willing to go there, but he's also willing to sort be throughout the stocks. I go down to a lot of detail, but then he will come back up and I think he has a yuppies neighbourly appreciate that that's why I love college early college, you in high school, just textbooks on the basics of computer science and mathematics of the biology of chemistry. Guess those are they condense down the guy sufficiently general, they can understand both the philosophy and the details, but also like you get homework promptly. You get to play with it as much as you would were if you were in the programming, stuff gas and then I'm also suspicious of textbooks honestly, because, as an example,

deep learning, there's no like amazing experts in the field is changing very quickly. The same is true in say that biology and so on these books, the cell, our kind of outdated there still high level like what is. shall we also of truth, its people and what labs working myself you know, sequencing genome, send yeah, actually working with working with it, and I don't have that much exposure to doubt or what that looks like so I saw them fully, I'm reading through the cell s kind of interesting and I'm learning, but it still not sufficient. I would say in terms of understanding what is it lean summers. Asia, the mainstream narrative, young yet the learned that, before you break out at the covered towards the cutting edge over the actual process of and with these cells and growing them in incubating them- and you know it- welcome massive cooking recipes of making sure your slows and proliferate and then you're sequencing the money, experiments and how that were I think he's gonna, like the source of truth of a the day, what's really useful in terms of creating therapies and all

they want in the future. I textbooks will be there's an artificial intelligence, the modern approach of actually having read. If it's come out there, Since version the recent there's been, recent edition ass, the sun the size of deep learning book, I'm waiting for textbook they're worth recommending worth reading and yeah. This tricky cosette papers. And code code code. Honestly, I don't papers are quite good. I specially like dependence, appendix of any paper as well as lacks, is like the most detailed can have fun, but it doesn't have to be cohesive connected anything else. You just described me where's your civic way. You saw that particular thing. Yeah many times papers can be actually quite readable, not always, but sometimes the introduction in the abstract is readable. Even for someone outside of the field or not. This is not always true and sometimes, income. Fortunately, scientists, two's complex terms, even when it's not necessary, I think that's harmful. I think they're. There is no reason for that:

it is sometimes a longer than they need to be in this in the parts that don't matter yeah appendix I belong put in the papers of look rice, the biggest but it us only. I come across papers, I would say say like synthetic biology or something that I thought were quite readable for the abstract and introduction, and then you reading the rest of it and you don't really understand. But you of our getting a just, and I think it's cool what advice you give advice, folks, engineer, machine learning and research. but in general life advice to young person, high school early college about how to have a career. They can be proud of our life. They can be proud of. I think, I'm very hesitant to give general advice. I think it's really hard I've mentioned like some of the stuff I've mentioned, is fairly general. I think like focus on just the amount of work you're spending on like a thing compare yourself to only to yourself not to others is good

are fairly general how'd, you pick the thing: are you just have like a deep interest as something or like try to like find the art max over like the things you are interested in our tax at that moment and stick with it are you now get distracted and switch to another thing? are you can if you like? We, if it is you do in our max repeatedly everyone It doesn't converge at the forum yeah. You can like low pass filter yourself in terms of like what has consistently been true for you, but yeah. I definitely see how it can be hard, but I would say, like you're, going to work the hardest on a thing that you can but the most I so look Ass well to yourself and really interests, backed in your past, where two things I gave you energy and where things that took her energy away from you crete, examples and usually from this concrete examples and ass parents can merge. I, like I like it when things look like this one of these, hence the status of the field, but the kind of stuff you doing in a particular field of view

it seems like you- are energized by implementing stop building actual things. being low level learning and also are communicating so that others can go through the same realizations, a shocking that gap, because I usually have to do way too much work to understand a thing, and then I'm like okay. This is actually like. Okay, I think I get it and like why was it so much work that should have been much less work, and that gives me a lot of frustration and that's why I sometimes got it. So, aside from the teaching you doing now, Videos, aside from a potential godfather part two, Would the asia I tesla beyond what does the future fonder capacity hold have you figure that out yet her now mean you see through the fog of war. That is all of our future do dese. Thirteen, so what on earth with a possible future could quite because it,

thing I've been always interested for me at least as his ally and dumb. That's probably one of spending, my trust in my life on, because I just care about, lot and actually care about, like many other problems as well like say, aging, which I basically view as disease and I care- that as well. But I don't think it's a good idea to go after it specifically I don't actually think that humans will be able to come up with the answer. I think the correct thing to do is to ignore those problems. and you solve a high and then use that to solve everything else and I think there's a chance, this will work, I think, to very high chance. And that's gonna like the the way I'm betting, at least so you think about it. I I you interested in all kinds of applications, size, domains, a domain you focus on will allow you to get inside the big problem of asia. It, for me, is the ultimate mental problem. I dont work on any one specific problem, just too many problems. So how can you workin all problems simultaneously? You you solve the met, a problem which to me is,

intelligence and how do you oughta? It is their cool small projects like a guy sanity in So one day you thinkin about, though, that the world The amount world can anticipate this always like some fun site projects. Society is one and yet basically like there's way too many archived papers. How can I organize it and recommend, and so on. I traced. All of your new airport casts. What did we learn from that experience from transcribing the process of, like you, I consuming audio books in and pakistan on here- is a process that but cheese, closer to human level performance and irritation yeah. Like definitely was like surprised that transcription with opening has whisper was working, well compared to what I'm familiar with from syria and like a few other systems. I guess it worked so well and ah that's but give me some energy till I tried out tonight, thought it could be fun to run upon us it's kind of

not obvious me why whispers so much better compared to anything else. because I feel like there should be a lot of incentive for a lot of companies to produce transportation systems and that they ve done so over long time. Whisperers now super exotic model. It's a transformer, its take smell spectra, grandson, you know just up its tokens of texts, not crazy. The model. Modern everything has been around for a long time. I'm not we have to censure. Why did he not is not obvious? Neither in MID makes me feel like I'm missing something fundamental something yeah because there is a huge meetin governance on you too bad transcription, yeah yeah, some clear, but some of it is also integrating into a bigger system that is The user interface house deploy not occur stuff, maybe running it. Isn't it thing is eaten it's easier like an order of magnitude, easier than deployed a large integrated system like you, too ask ocean or anything like me, exude has transformed the transcription that's kind of crappy.

But creating interface, where the text a different individual speakers, its aims, to display in compelling ways run a real time. All the cost of maybe less difficult, do less pressure have, because I come from currently paying quite a bit for human transcriptions caption right. in addition, like it seems like there's a huge incentive to automate that yup. It's very confusing- and I think I mean I dunno if you looked at some of the whisper transcripts but they're quite good good and especially in tricky cases, scene, Oh whispers performers, unlike super tricky cases, it does incredibly well. So I dunno Yes, it's pretty simple ethic quality, audio and you're speaking, usually pretty clearly etc. Open your eyes. Plans are here either, but get there's, always let one from project basic

and thus table diffusion, also is opening up a huge amount of experimentation. I would say the visual realm and jerry generating images, videos movies online videos now answered that's going to be pretty crazy. That's going to. he's going to almost certainly work and is good. You really interesting when the cost of content she's, going to falter zero used to need a painter for a few months to paint a thing, and now it's going to be speak to your phone to get your video hollywood and start using the degenerate scenes, harm which completely opens up yes even mega, like em, like avatar eventually for under million dollars? Much less may be. Just by talking to your phone, I mean, I know it sounds kind of crazy and then there be some voting mechanism like how'd, you haven't would there be a shoal, annette netflix has generated complete automatically. Sally somebody actually here and what is it

like. Also when you can judge jury on demand and it saw and there's infinity of it yeah, man all the synthetic garden, in its humbling because with which feed ourselves a special for being over the jerry, in ideas that kind of stuff, if there It can be done in an automated way by ai. Yes, I think it's fascinating to me how these are the predictions of ai. What it's going to look like I was going to be capable of are completely in and rock and supply of fifty or sixty slack, totally, not pride. They imagine, hey I as like super calculating fear, improvers and we're getting through I can talk to you about emotions, it can do. Art is dislike. Weird. You said about the future, just as like hybrid systems, heterogeneous systems of human denies. Talkin bout emotions, nothing

in children, ai system, virginia budget, were the netflix thing you watch is also turner but hey. I think it's it's going to be interesting for sure, and I think I'm cautiously optimistic, but it's not it's not obvious with the sad thing is your brain and mine, developed in time Where, before twitter, before the in before the internet, I wonder people that a boy inside of it might have different experience like I. Maybe you can was still resisted and the people born now or not, but I do feel like him truly malleable here and you're, probably right what is the meaning of life andrey? We talked about. the universe, heavier conversation us humans or with the systems, will create to try to answer for the universe. For the creator of the universe, to notice us we're trying to create systems,

There are loud enough to answer back how does the meaning of life? That's like meaning of life for some people, the first level so I would say, is any can choose their meaning of life because we are conscious entity and it's beautiful number one butter. I do think that, like a deep, meaning of life is someone is interested, is or along the lines of like what the hell is, the oldest and like. Why, and if you look at that into Little physics and the quantum field theory and standard model there like weight and the very complicated and down there's disliking online. in preparing the parameters of our universe and like what's going on with all the stuff and wise it here and can I hack, can I work with it is our message for me: I suppose, to create a message I think there's some fundamental answers. There are, but I think, actually even like can't actually really make a dent in those without more time, and so to me. Also there's a big question round: just getting more time honestly, yellow

like what I think about quite a bit his walk so kind of the ultimate. I leave first weight to sneak up to the. Why question is to try to escape the system the universal and then for that is her back track as ok. For that that's gonna, be you take a very long time, so the white question boils down an engineer perspective too. How do we extend he? I think that the question number one practically speaking, because you can't you're not gonna, go the answer to the deeper questions in time you have an echo, It's in your own lifetime or extending just the lifetime of human civilization, of whatever wants to Many people might not want that, but I think people who do want that. I think I think it's probably possible and I don't know that people fully realised this. I kind of like people think of death as a minimum and billy butter. At the end of the day, this is a fiscal system. Something go wrong. It makes sense why things like this

have speaking and there's most personally interventions that mitigate be interesting of death is eventually looked at us. As a fast anything they used to happen to humans. I don't think it's unlikely. I think it's I think it's likely. It's up to our imagination, to try to predict what the world without death looks like yes hard to. I think values will complete change could be. I dont I dont, really by all these Yes, that without death there's no meaning there's nothing is I dont intuitively by all those arguments, I think there's plenty of meaning plenty of things to learn their interesting exciting. I want to know how to calculate to improve the condition of. Although he lives organisms they're alive at the wheel? we find meaning my change. We There is a lot of humans problem, including myself that

fines, meaning in the financing of things, but that is, and ass the only source meaning. Yet I do think many people will go with that, which I think is great. I love the idea that people can just choose their own adventure. Like you, you are born- is a conscious free entity by default. I like to think and dumb You have your unalienable rights for life in the pursuit of Happiness priscilla was ended, Should the landscape of happiness. can choose your own adventure mostly and that's not fully true, but I philip, british sure and pc, then I beseech can't know too busy. There Different degrees allows of consciousness. Does a more beautiful way to andrea you're, credible person. I'm really are you talk. Will me everything you, and for the machine learning world for the world to justice by people to educate, millions of people has been has been,

wherein I can't wait to see what you do next has been an honor man. Thank you so much for talking to the awesome gift, thanks for listening to this conversation with aldrich or party to support this podcast. Please check out our sponsors in the description, and now let me leave you some words from Samuel Karlin. The purpose of models is not to fit the data, but does sharpen the questions thanks for listening and hope to see you next time.