AI: Grappling with a New Kind of Intelligence

In the vast expanse of our universe, filled with countless stars and galaxies, humans have always sought to understand the mysteries of existence, but now we stand on the brink of a new frontier, not in the distant realms of space in time, but within of the complexities of our own digital landscape artificial

intelligence

or AI is a term you've heard countless times, but what's really hidden behind today's algorithms and calculations we're delving into like the tools of generations past. AI promises profound benefits, but it also raises powerful questions: Are we? On the brink of a golden age of innovation or on the verge of our own obsolescence, it is not just about weighing the pros and cons, but also trying to fundamentally understand the functioning of their artificial

intelligence

systems, e.g. large models of surprisingly versatile languages ​​capable of generating text. answering questions, even creating music, but how do these models think, or how do they think at all?

And if they don't think like us, what exactly are they doing on today's show? We will take a look under the hood to see the inner workings of these Digital Minds not only for Marvel but to understand, to demystify the magic and address the real, because when we understand we can act with foresight, wisdom and purpose, so that, as we wish , let's be at this intersection of humanity and technology, let's embark on this journey together. Welcome to AI that fights with a new type of intelligence. Good evening everyone, thank you. I'm Brian Green and maybe I should add to that by saying I'm the real Brian Green because I think, as some or many of you may have already assumed, you concluded that that was the case.

It's not me up there on the screen. I didn't write a single word of that text. No human being did it. All of that text was written by a large language model. The sequence of images you saw in that piece was chosen for a big language. ch GPT model and I didn't actually say anything you saw on the screen, so you're saying I did a better job, well, you have some binary charm, thanks, but while I can put words and concepts together, I'm missing the essence. of feeling I can't treasure a nostalgic memory Savor the touch of a gentle breeze or savor the excitement of innovation that is so sad for everyone how about a little love for my digital counterpart help me here yes of course thank you I think although no I don't really feel the recognition, hey it's the thought that counts or in my case the algorithm just remember to turn me off when you're done we'll do it good night digital good night organic so yeah thank you why not, yes.

So for some it's impressive and for others a little terrifying that a system is able to do this and if you look at the course of human history there are, I don't know, a few dozen key moments, turning points, breaking points when it happened. something that changed. the course of the future of human development, right? You can think about the acquisition of language, the domestication of fire, the acquisition of written language, the ability to have the wheel, learning to have wireless communication, the printing press, the ability to have self-propelled vehicles. I mean these were all key technological developments that radically changed the way our history unfolded and it may well be so, and this is what we will discuss here tonight: we are at a similar

kind

of inflection point because a way of thinking about a form of What I find particularly illuminating is that I like to organize all of reality into three general categories: you have the space for the big things, you have the atoms and molecules for the small things and now you have the complex things for the big things that we have created. great progress Einstein's General Theory of Relativity gives us truly profound insight, we have not been able to translate that insight into control over the larger Cosmos, but intellectual insight is dramatic in the realm of the small, of course we have the mechanics quantum mechanics and we have been able to harness our understanding of quantum mechanics to control the microworld, giving rise, for example, to the integrated circuit which, of course, is the key to the computational power behind everything we will talk about here tonight , but in the realm of the complex for a long time it was a realm that was

kind

of Beyond reaching the realm of the complex really is the realm of life and mind or intelligence, but now, in the last two decades , with synthetic biology we are gaining the capacity. really harnessing and understanding life in a way that gives us control and the inside that we've seen now and that all of you have experienced, certainly, having played with these systems gives us the possibility that we also have synthetic intelligence, artificial intelligence, which bring the realm of the complex within the power of what we can control and this is exciting, but for some it is quite scary and what we are going to do tonight is really explore both possibilities and try to understand where we are and where we are.

Let's go when you have systems, for example, like the one that just gave rise to the piece you just saw. It is from the Clone Works AI studio. If you can give them a round of applause for the work they did in that inaugural peacekeeping force. But what does the ability to create such things as deepfakes mean for the nature of democracy and the future of our species? So what I would like to do now is bring on our first guest, which is Yan Laon, who is the Silver Professor of Data Science, Computer Science, Neuroscience, and Electrical Engineering at New York University, he is also Vice President and Director of Research of AI at Facebook, he led several important innovations underlying generative AI and in 2018 his work in deep learning earned him the Turing award, welcome, yes. thanks, just so you know since last November when we all got the ability to play with, for example, GPT chat.

I think a lot of us have had those amazing moments where a computer is doing something that you don't think a computer can do yet, and so on. Can you give us a hint that we'll get into that in more detail in the second part of this conversation when Sebastian Bubba comes out? Can you just give me kind of an overview of how these systems are able to do what they ask a very interesting question, so for us the revolution happened two or three years before for people like us in the research there were some surprising results that everyone in the research community was surprised by the fact that, when you train Artificial Neural Networks that are very large and contain a large amount of data.

There is some sort of emergent property that happens to a greater extent than we were used to. For other systems, they can do computer vision or things like this, but to a certain extent for us, the revolution. it didn't happen in November it happened three years earlier I mean it was the case it was the case I think you know I'm obviously not in the field but I would listen to colleagues for I don't know 20 years 25 years talking Are we on the cusp of things changing radically in AI? Have we finally reached that c? I mean, for many decades it didn't seem to really happen well, so the history of AI is interesting because it's a series of new paradigms, new ideas that people discovered and as soon as the new thing happened like the new paradigm, the people said that's the solution, now we're going to build intelligent machines and it started in the '50s, um, with very famous people in AI, it ended up winning the tour. prize too new and Simon, they say we are going to write this program, it is called general problem solver, it will be able to solve all the problems in the world, they wrote that program back in the 50s, all the problems in the world, all the problems as always and when a problem can be formulated in terms of an objective to be satisfied, yes, and then a search for possible solutions, he said that everything in the world can be formulated in those terms, yes, what they didn't realize at the time is that Every interesting problem requires an incredibly large amount of computation that was completely unattainable even today, so basically that program failed, yeah, and it took them a few years to realize that there was another effort around the same time people were saying no. , no, no, the way intelligent intelligence is built.

Machines is that you allow them to learn, that started in the 50s and they say okay, now we have this machine called perception and it's magical. You can show her images of characters and you can train her and she will learn to distinguish between you. they know a C and a D or a tank and a Jeep or whatever um and then they realize that actually no, it doesn't work that well, you can't really make the system learn to distinguish complex images or something and so on. It was Very limited again, people thought they had a solution and it turned out to be a failure.

In fact, that type of technology, actually people were working on this, was going to change the name of what they were doing and that was actually the basis. for a lot of modern technology and communications called advertising filters, very strange history, um, and then there was another movement, uh, 10 years later, called expert systems again in AI, a lot of interest in this. Japan started a big program called fifth generation computers. being able to do inferences and logical reasoning and things like that and the problem with this is that there was no learning so it was people who had to input facts and rules into computers and it turned out that it worked for a small number There were a lot of problems but really you couldn't build intelligent machines with this, so all research interest in this sort of thing died after 10 years and around the same time a new wave of neural networks appeared, so this was the late 80s and early '90s, that's when I started my career in the late '80s or early '80s, I should say, and there the idea again was how do we recycle those perceptron ideas so that machines can be trained but make them more powerful and we found a new technique that is still what we use today in things like large languishing models for training artificial neural networks that were very successful for about 10 years and then people lost interest in it, because the computers were too slow and the data sets They were too small there.

There were all kinds of problems, um and then it came back to the four about 10 years ago and the reason we hear about AI today is because of that renewal of interest in those So-Cal deep learning techniques, basically being able to train the big Neal. The networks that we see today are the effect of more powerful machines, a larger data set and something like that allows them to build very large neural networks, with billions of the equivalent of synapses in the brain, hundreds of billions, and they can do it. Pretty amazing stuff, so we'll look a little more under the hood in the second part of this conversation when Sebastian BBC comes out, but while it's a little mind-blowing for someone like me and I think maybe for the general public, that You know, I can give you a hint, which is what I did for the little piece we just played.

All I did was say: can you write an interface for a Brian Green-style World Science Festival program? You know, that would introduce his show and the script that we just heard came out. I have to say it made me feel a little sick for a moment when it came out, but then I rationalized it by saying well, maybe it's looking at things I've written in the past and it's inspired by them, so maybe in a sense I had something that do with that, but putting that aside, you're a little less baffled by what's happening today, can you just give us an idea of ​​why that's okay?

There is no doubt that the advancement of technology in the last five years has been surprising, impressive, but these systems easily fool us into thinking that they are intelligent just because they manipulate language fluently, the only example we have of an The entity that can manipulate language are other humans, so when we see something that can flexibly manipulate language, we assume that that entity will have the same type of intelligence as humans, but it's simply not true, those systems are incredibly stupid, so which are very useful, okay, they need to be developed, they're going to be, you know, marketed, they're being worked on, they're great, but they're kind of specialized and they're very stupid in a lot of ways, so they're partly stupid because they are only trained in language. and most human knowledge has nothing to do with language and all animal knowledge has absolutely nothing to do with all philosophers, for example, I suppose you would agree with that statement Vicken Stein, you, the limits of my language, the limits of my world, right?

So are you affirming? That, just as an assumption, I mean, I agree with you, but are you really convinced that language is only a limited part of human understanding? I think the languageIt is a way of communicating knowledge and, certainly, of storing a large amount of knowledge in our brain because language is efficient in a way, they have discrete concepts in words, etc., but an Aang Tong does not have language, they are incredibly intelligent, They are almost as smart as us, we think. We are much smarter than them, but we are not like the others. A physicist can understand that if I continue with this, I'm not sure exactly how to interpret that sentence, but I'll let it go if I continue with this. with a little force it's going to move to the right, uh, and right now it has water, but if it didn't have water, it would probably tip over, yeah, if I push with the same force on this table, it doesn't move, uh.

We have an intuitive notion of physics, we learn it when we are babies, the large language models that we have today or any artificial intelligence systems that we have today, none of them are able to understand any of this to some extent, intelligence systems artificial, the most intelligent artificial intelligence systems they have today. less understanding of the physical world than your house SC, but can I ask you about that? the data that the system is trained with there is an explanation of basic physics, like the example you just gave, wouldn't you consider that an understanding in quotes because if I asked the system if it trained enough, what would happen if I pressed? on a table in the way you just suggested, it may come back and not tip well and that's really an important point for situations that you explain to the system, describe it in words and correspond to a template that the system can learn and reuse in a new situation, it could work, but there are a lot of really complex effects here, as you know, if I push the side that is going to rotate, there is friction on the table that could change, so if I push the same bottle here on the the ground is going to tilt it's not going to slide so all of this isn't really expressed in any language most of what we know about the world isn't reflected in language yeah so it's a philosophy question that philosophers ask themselves : Can we build intelligent machines that are exclusively trained using language and do not have any type of sensory information?

My answer to this is absolutely not. Know. I like to start my classes in Colombia. the basic classes, you know, throwing a light object into the audience and having someone catch it and trying to convince them how remarkable it is that they knew where to put their hand without doing any Newtonian calculations of the object's trajectory. and convince them that this is intuitive physics, but then I go further and say why we have that intuition and the answer is its evolution by natural selection. Those of our ancestors, presumably, who were better equipped to understand the physical world, were better able to navigate that environment.

We actually had a better chance of surviving and transmitting that ability to subsequent generations, which suggests, I think, and it is quite clear, that our intelligence is not as general or as broad as perhaps one might think, but is actually very specific, it was developed so that we could survive absolutely uh there's something that people in AI talk about artificial general intelligence AGI, by which people mean essentially human level intelligence, human type intelligence. I hate that expression because human intelligence is in fact incredibly specialized, there are many of them and we know this because of computers, computers are much better than humans in many tests today, right?

And we knew we've known for decades that they're better at arithmetic, of course, or solving equations or computing integrals, you know it symbolically and things like this, but now. We also know that computers are much better than humans at playing chess or playing poker at diplomacy. I mean, there are a lot of games where you know computers have gotten really good and what it proves is not that computers are smart, but that humans just suck, we're extremely bad at doing things like imagining a lot of things. scenarios like in chess, right, we, uh or o come on, we have to imagine a number of scenarios that depend on someone else playing and for each move you make.

We know that there are something like 36 possible moves and it grows exponentially, so we can't contain this in our brain, so we develop a kind of intuition of what constitutes a good move or not, what is correct and the modern go and chess. The game systems actually do this well too, but they also plan and have a much larger memory and, uh, working memory and ability to explore trees, so they beat us, so we're really not very good at this. that and in your language you know this technical term, you know that humans suck, um G, since, as I understand it, if you could explain to us, your vision of where AI should go to achieve its potential is to model intelligence on intelligence human for To some extent, I would start with cats because we don't know how to replicate the kind of intelligence and understanding of the world that the cat has, so the first one, okay, there's something called common sense and we have intuitive physics. is part of it, yes, cats have a certain level of common sense, no artificial intelligence system today has any of that, so we are going to have to move towards systems that can learn how the world works by observing it and interacting in it. the way babies do, babies in the first few months of life learn an incredible amount of basic knowledge about the world, absolutely like gravity, so the concept that an object that is not supported falls, babies They learn this around nine months, it takes them a long time. time, um and uh, we don't know how to do this with computers, if we did, we would have fully autonomous level five cell driving cars, um, that run exclusively from Vision.

Today we sell circular cars, but they are not completely autonomous. It's not entirely reliable, they require you to know thousands of hours of engineering and training data and all kinds of sensors, so how is it possible that a human, a 17 year old, can learn to drive a car in 20 hours of practice? Yes, largely without causing Any accident for most of us is a little scary, but if you are willing to put up with that, yes, you will be able to overcome the obstacle, but can you explain to us your vision of how we could achieve that level of artificial intelligence ? and I think this is a framework for it, so there are a number of things that AI systems are not really capable of doing a number of things that one of them is planning.

This used to be a topic that classical AI people already know. 20 years ago we were really interested in um, but U, including the general problem solver, but a modern system like large language models are not capable of planning or at least only capable of very simple planning, so what you will see appear here. progressively it's kind of an architecture of different different modules that would be necessary for a complete AI system to have the kind of intelligence that we see in animals like cats, dogs and humans, and it starts at the top, here's something called a configurator that you can see at the top and it's basically uh a a uh a director, a master of ceremony that organizes what the rest of the brain or the system is doing, tells the other systems in the brain that you're facing this situation, you have this goal to achieve now, do it, um, if you move on to the next module, of course, the system to function properly must have some idea of ​​the current state of the world, that's called perception in humans, it's all in the back of the brain. actually represented in the back of the brain here too, the back of your brain is visual perception, auditory perception is on the side that just processes the pixels that come from your eyes and turns them into a representation of the state of the world , an abstract representation of you.

It is not necessary to give them names. Babies can recognize categories of objects without knowing their names, and all animals can do this too. So once you have an idea of ​​the state of the world, you have to imagine what is going to happen. happen if I take a sequence of actions, what is going to happen in the world, can I predict what is going to happen in the world? And that is the role of what is called here the world model. This is where most of the intelligence actually goes. In humans it is in the front part of the brain, which is particularly large in humans and this is what allows you to predict that a particular action will have a particular result and therefore if you can predict what result will occur as a consequence of an action that you take, then you can plan well, you can plan a sequence of actions to get to a particular result, so that it is determined by this cost module, so the cost module is basically something that measures to what extent the predicted state of the world that you imagined satisfies the goal you set for yourself or that the configurator is configured for you this cost module basically measures your degree of dissatisfaction with the outcome of something imaginable this is the seat of emotions really So if you predict that a particular situation will result in a bad outcome and there is nothing you can do about it that produces fear, your goal is not satisfied, so this is a kind of emotion produced by a prediction, but there are emotions that are immediate as if I pinch you okay, you're not going to like it very much and if I try to pinch you again you'll probably back off because you'll realize that you're now your model of me. it includes the fact that you're a pincher, I'm a pincher, yeah, okay, so the last module here is the actor and the actor's role is basically saying, am I going to be able to produce a movie? sequence of actions that according to my prediction through the world model is going to satisfy this objective that I set out to minimize this cost or whatever and this is how we do planning, it is actually a very old idea, very old in engineering, it goes back to the 1960s, but how to do this in a smart way where the world model is learned by observation, we don't know how to do this today, the big challenge, I think, for the next decade and I wrote a long article about this It's um How do we get machines to learn how the world works through observation?

How far along are we in that? I mean, we're making progress particularly in the last couple of years and the principles are somewhat similar to those used in large language models. a set of learning techniques called self-supervised learning and maybe I can explain this a little bit, yeah, um, so when you train, uh, I'm sure we'll see the corresponding graphs, but yeah, that's a good example. Okay, so the way those big language mods are trained is to show them some text and then remove some of the words you say to them. Okay, and of course we know the sentence, but even if we hadn't seen the sentence, we probably could do it. guess what words are missing in that sentence, then we train a very large neural network to predict the missing words and in the process of doing so, the network learns a representation of the text that includes meaning, syntax, grammar, which you can even train it on . in multiple languages, you can train it on computer code, you can train it on all kinds of things and it learns the internal structure of the language in such a way that you can then use the internal representation of the language to do things like train the system to translate. to detect H speech to classify the topic to generate U um summaries or summaries of a text um uh and there is a special case of this uh if we can move on to the next uh uh diagram and a special case of this is what we call llm, which we should call autoregressive llm, so what is a GPT llm?

So the big language model LM everyone knows, but just to be sure, yeah, right, it's a special case of what I just explained, where the word is that you're masking. just the last word in the uh in the text, okay, take a long text, take a window, a few thousand words and then shuffle the last one and then train a gigantic neural network to predict that last word, okay now yeah you train that system successfully. It can't exactly predict the word all the time, what it's going to predict is a probability distribution over all the possible words in the dictionary, so in the sentence we saw, you know, the cat chases the blank space in the kitchen , the blank space could be a mouse, it could be a laser pointer, right.

You don't know exactly what it is, but you can train the system to produce, okay, it's going to be a cat with probability 0.5 and then a laser pointer with probability point 2, and then you'll know something. otherwise probity .01 and whatever and for every word in the dictionary you have a list of scores essentially um and then what you can do is use this for Auto what's called backward auto prediction and that's really howdo those big language models work um if you can go to the next diagram at the bottom, you give a text window to the system, have it predict the next word and then inject that next word into the input window, okay, change all the words one at a time and then ask him to produce the next next word and then change it and then ask him to produce the next next word etc. etc. and the system will just produce one type of you one word after another, so let's see, so the idea is to feed the system a huge body. of text and based on all the words that the system ingests, it begins to accumulate probabilities or probabilities for given sequences, different combinations of words that could appear and by using them, it can make a prediction of the next most likely word in a given sequence and including that word making a slightly longer sequence.

You can predict the probability of the next word in that augmented sequence and word by word you construct the most likely sentences and the most likely paragraphs, etc., there's just one bar. The problem with this is that if you make a really bad mistake about what the next word is going to be, everything will trigger, it's RNG, it will trigger, right, it's a divergent process, right, if you want to use the technical term, um, and it's that's why those systems. hallucinate, they're not correct sometimes most of the time sometimes um so it's good for poetry because you don't care if it's incorrect it's not so good for you you know mathematical physics that kind of thing um uh and there's no planning, obviously, because the system just produces word after word without really thinking in advance about what it's going to say, so of course you know it has some representation of the previous word, so you know it's consistent, but it doesn't plan its answer is which is really reactive, it's like you know there was this whole movement in poetry called automatic writing, right? you don't think about what you write, you just write word after word, that's basically what's happening there, so those things are very limited.

You could think of them as a model of a part of the brain, but one part of the brain would be area verik and area B which are like little pieces of the brain like this one on this side that manipulate language, but what about the rest? And you claim that a similar type of training procedure is relevant to the vision of artificial intelligence that You imagine where you have a real world model and you are actually planning the actions that you are going to take correctly, so the big question now is how you train a system to have a model of words, so an obvious idea is you.

Take the same type of model, but instead of the inputs being words, the inputs being, say, frames of a video, train a system to predict what will happen next in the video, so I put this bottle here on the table and hold it. If I raise my finger, you all know what is going to happen, it is going to fall, you don't know in which right direction because you don't have enough perception for this and you don't know that I am going to do it. If you remove my finger, you can make a prediction that it's going to fall, so if you could train a system to predict what's going to happen in a video, maybe it would acquire some kind of model of the physical world, so the logical thing is just use one of those big language models and instead of including words include, you know, convert video frames into sort of words, like things, okay, we call it tokens and it doesn't work, so we have to find other techniques new ways to do this and The reason it doesn't work, or at least that's what I think, is that the problem is that you can't predict exactly what is going to happen.

When I do this experiment, it can fall one way or another, and what? You can't, you can't train the system to make just one prediction, a lot of things can happen that are plausible following the video and you're training the system to make one prediction, but it needs to be that way. We can predict all kinds of possible scenarios and we don't know how to represent a probability distribution over an infinite number of scenarios. We can do this with words, so the language is easy. 50,000 words in the dictionary. That relevance, yes, you know, we can represent it in at least 50,000 probabilities, but we don't know how to do it for video frames that are basically in a continuous high-dimensional space.

However, you're sure that in a reasonable period of time that's where things will go and these big languages ​​I guess you think it's going to be an interesting problem, but it's not what's really driving AI in the future, so my vision for the future is that we will be able to develop techniques that are able to learn how. Representing the world from a video by watching videos perhaps has predictive models where the system can imagine an action and then imagine the result of that action in the representation space. The type of architecture I propose for this is called jepa, which means set. edding predictive architecture I'm not going to tell you what it means, but basically, instead of predicting the pixels in the video, we predict a representation of the pixels in that video, so we don't predict all the details of what will happen. happens in the video we simply predict in abstract terms what can happen in that video as a consequence of the actions that the system can take and if we have such a system we now have a world model, so we can now integrate that world model into the architecture that I showed before and maybe give it the ability to plan, yes, if we have the ability to plan, we can incorporate some cost functions, when you decide what sequence of action to take, so that those cost functions are protected. rails that will ensure that the system is safe and controllable.

This is not the case for autoregressive LMS today, so my prediction is that within 5 years maybe autoregressive LMS will disappear here and will be replaced by something I call goal-driven AI depending on the type of architecture. I proposed it and those systems will not only manipulate language but hopefully also understand the world where we're not going to get to human level AI in five years, um, that may take a couple of decades or something, I want I mean, I don't know, maybe it takes a lot longer than I imagine because it always takes longer than anyone imagines.

Yes, given that Vision, I now want to bring in our second guest, who can give us a complementary perspective on some of the things we're talking about, Sebastian. Bubeck is a managing research partner at Microsoft Research, where his focus is understanding how intelligence emerges in large language models before joining Microsoft. He was an assistant professor in the operations research department at Prinston University. Thanks for joining us. He's already heard it. uh our conversation with Yan talks about various types of intelligence human intelligence feline intelligence the type of intelligence in large language models maybe you can give us an idea of ​​how we define intelligence and then maybe how the large language models with which the People have played with that I generated our opening sequence, how do they compare to that definition?

Sure, yeah, I can talk about that, so you know, defining intelligence is incredibly difficult. I mean, it's as difficult as defining what space and time is, which is something you know about it. I don't think defining intelligence is much easier, but we all agree that an intelligent system must have certain basic things, so I clearly believe that it must be able to reason, whatever definition of reasoning it must actually be able to do. planning that Yan was referring to and talking about a lot and you need to be able to learn from experience as you evolve in the world you know to learn new facts and furthermore, in addition to these three canonical blocks, you also need to be able to do it in a very general way , that's where you know the AGI that I was talking about artificial general intelligence comes into play is that you don't want a system that can reason, plan, learn from experience, but do it only in a limited domain, so it's really essential to essential to being General, that's the key to intelligence, so you know when you look back at systems like Alpha go for go, from a few years ago, or the system that actually beat the world champion in at go, that's right, that's how it is.

These things were called Ai and to me, in my eyes, they were not intelligences, they were systems, you know, very narrow systems that could perform a task, explore the tree of possibilities very, very efficiently. I think there's a big gap between those systems and GPT chat and GPT 4, where those systems are general, they can do those things that you know, but they're not restricted, they can do it in many different domains, and then how do you know in function of these three criteria? How do you stack chat correctly? So maybe just to back off. Wait a second, you know, I got access to GPT 4 a few months before everyone.

I was lucky enough to work at Microsoft and we were working on integrating gp4 into the new ser so I got access to it last summer so a few months before everyone had access. access to the GPT chat and immediately it was the GPT 4 model now you know I've been working in this field for 15 years and I used to do AI mathematics and you know in mathematics we love to prove impossibility results, that's what we spend our time proving that that's why it can't happen, you know, what Onean was talking about, you know, those things can't happen and then I felt like I knew that certain things were impossible to do with a Transformer type architecture and then I got I accessed dpt4 and I He surprised what he was able to do, like he kept surprising me over and over again, so it was really a very humiliating experience back then, now what we did is we didn't want to be.

I'm just amazed because you know there's a problem that Yan talks about often and that is that he's been trained all over the internet so he knows everything that's out there so it's easy to be impressed just because he has so much knowledge that he can recover many things. so you really need to ask him a strange question to try to make him can't look up and just understand it, he can't look up, he has to invent something new, you know, in terms of planning reasoning and Learning from experience, our evaluation after After many months of playing with the model, it can definitely reason.

Personally, I have no doubt that this thing can reason, it can't plan, that's why I agree with Yan, it can't plan, but it's subtle like the The meaning of planning is not like planning your vacation next week, it's more well a type of mathematical planning and, to learn from experience, it is a mixed bag because of course a model like CH GPT is frozen in time, at least in principle. I know what opena is actually doing, but in principle it's frozen in time and if it's frozen in time it doesn't work. When you say in principle, I would have said in practice.

Am I misunderstanding because I would have thought you might be allowing this? train all the time no uh yes exactly exactly so what I mean by in principle is that there is a yes. I guess in practice it's maybe a better way to describe it is that there's a way to train those neural networks that you train. in a huge Corpus and then you have your machine, it's fixed and you can interact with it by training it continuously. I would say it's doable in principle, yes, but we don't really know how to do it right, so yeah, I mean it's an interesting assessment that you give and you preface it by saying that you were quite impressed or surprised by some of the things that were happening. , I almost had an impossibility proof that that couldn't happen, you know an example, I think you that you've talked about is a poem that I know Yan said that poems are kind of, you know, maybe they're the cheap food for a great group, but nevertheless, the one that you don't have the system to produce is quite interesting, yes, it is. in fact, and just before we get to the poem, I want to say a little bit more about the reasoning and what Yan was talking about about some kind of impossibility, so when gp4 came out, Yan wrote a challenge for gp4, you know you passed it on. uh online on Twitter for everyone, it was a very interesting question about the gears that turn in a circle and if you turn the number one gear, what is the number one gear like?

You know the six is ​​going to turn and Yan has the question about eight gears, so this is a gear sequence. which are all intertwined in some interesting way, you turn one, what happens in the future, that's right, yeah, and I guess Yan was provoking the community because with eight gears it's easy, as you know, everything turns, uh, very easily and then when someone said, yes, I tried it on gp4, it works Yan said ok, now try with seven and with seven the system is too constrained so nothing will move and if you just ask gp4 this question, you would know to make up a wrong answer and say it's okay.

I'm going to move and you know Yan said correctly at this time, okay, this shows that you know, how do youI said, you can't reason, yeah, but then someone on Twitter asked the same question and added, by the way, this is a question from Yan Lan and then it worked and you know it's funny, but it actually makes a lot of sense because what What's happening is that by saying Yan l in the context of this system, the system understands ah, we're talking about something intellectual, maybe you really have to be prepared for something difficult, okay, so I think when you look at this for me personally , it's impossible to say that the model is not Reon, now okay, back to the poem, uh, I don't know if you've had this experience, but somehow everyone.

The first thing they do when they play with chpt is they ask you to write a poem. You know, there's something beautiful about that. I don't know what it means for humans, but this is something everyone asks now. I was sitting in my office, you know, at Microsoft with my long-time friend and collaborator, R Alan, and we're both mathematicians and of course we wanted to ask him to write a poem, yeah, so we decided what Ron and said. en: Ask him to write a poem about the most famous proof of all time, the proof that there are infinitely many prime numbers, and this is what came out.

This is what came out. Would you mind if I read it quickly? You asked him to give a poem that gives proof that there are infinitely many prime numbers. Yes, I think I can. I thought a clever plan might be in order. I'll start by pointing out Yuk's proof, which shows that prime numbers are not just something distant. It's pretty good, I guess. we have a finite list of prime numbers and none have been lost, multiply them all and add one just to be smart, the result will be a number that has no prime factors oh i wonder, but every number has a prime number that divides, it's just question of At the time that was a bit of an awkward line, so we found a contradiction and our finite list needs to be dislodged.

Beautiful, there must be infinitely many prime numbers and that's the end of my rhyming line. Well, that's pretty good, so again the test, of course, is multiply. Everything known. cousins ​​together, you add one, the existing ones on your list cannot divide it, yes, so either it is a cousin or it has to be other cousins ​​beyond the last one you have on your list. Very well done, I mean, what did you conclude? Seeing this poem, I mean, I will always remember it. You know, Ronan and I looked at each other in complete disbelief. I mean, you know the line about adding one just to be clever, like it gives me goosebumps.

Still today, how are you young? You get goosebumps at that, it's cute, but it's also okay, I mean, generating rhymes is a relatively simple pursuit, a relatively simple exercise, um, and the proof is out there, there are tons and tons of versions of that test on the Internet. If that. It's been trained on gp4 so it's a smart retrieval with some tweaks and it's basically what I think is a feature of those LMSs that they can retrieve from a huge amount of training data, they can modify what they've retrieved to suit the actual situation. but if they're faced with a completely new situation, they completely fail, but I mean, if I gave them that challenge of saying a high school kid and they came back with something like that, I wouldn't just say cute, I say wow, that's really cool, okay? ?

Where do you find a high school kid who has read the entire Internet? Yes, actually Mo, I think most have, but that's a separate topic just to give you an example. Okay, gp4 or CH GPT in general has been trained or most of the llms have been trained. something like a trillion tokens or two trillion tokens that you can think of as words more or less if the token is a subword unit, so it's a little less than a trillion two trillion words, how long would it take for a single human being to read all this? about 20,000 years reading eight hours a day, but this is an incorrect comparison because human learning should not be compared with the creation of a neural network, since the creation of a neural network is much closer to evolution than to learning human during his life.

I disagree, so hey, I think we can all have our own opinion on how impressive that is. You definitely know, it gets me, but you know as someone who's not in the field, there's another example that you did that's interesting to do with an unusual animal, can you tell us that? Yes, I have small children at home and a daughter who keeps talking to me about unicorns. Yeah, and you know, one of the things I played a lot with the model is asking her to write. stories about unicorns that she read to my daughter and then one night I woke up and I was thinking actually, what does gp4 think a unicorn looks like?

Yes, and I asked him: can you draw a unicorn? Okay, which is a very strange thing. Ask because it is a text to text model, it takes text as input and gives text as output, so how is it going to be drawn? So I asked him this question: can you draw a unicorn? and what it returned were lines of code. which when I compiled I gave this image so again it's another moment that I'll remember forever uh it's just amazing like it's amazing it's able to, you know, traverse modalities, it's only seen text but it can have this visual representation of a unicorn Now, for Of course, you can look at it and say, "It's okay, but it looks a little horrible." I mean, it's not that he's not such a cool unicorn, which is fair, but he gets two things, number one is that he got the concept right.

I understood that there are four legs there is a tail there is a head and most importantly there is a horn you know that the horn is a there is a horn is half there is of course yes yes yes yes absolutely and also you know that I couldn't ask it just draw a unicorn, that would be too easy because there are drawings of unicorns in Code online, so what I did was ask him to draw it in a very obscure programming language, which if some of you in the audience are mathematicians, you already know. We use this programming language called tigz to make mathematical drawings and if anyone has played with it, you know I wasted a lot of afternoons during the process.

I couldn't do this or draw two circles. You know, it's already a work afternoon, so. I asked him to draw it in that programming language, which is really surprising because you haven't seen it on the internet and most importantly, one thing that's really crucial to this whole discussion is that this is with gp4, now you all have made. played with GPT chat when it came out in November if you ask this question for GPT chat which is the less powerful version then this is what you get if we can show the next slide for you to see this is a visual representation of how we have achieved many progress in just a few months and the difference is the size of the training, we set the number of parameters and we're going to talk a little bit about the parameters in a moment, but that's the type. of improvement we're talking about, that's all, and also, you know, when I got early access to gp4 and open AI was still training the model, I was able to ask my unicorn example throughout the training and what I saw was If you look on the next slide you will see that the Unicorn continued to improve throughout his training, so this is really like machine learning in action, like it really was machine learning reading, making more passes on the Internet, he refined his artistic skills I mean, it's crazy.

Are you cute or uh Beyond cute or cute in a weird way, okay, we can leave it at that, okay, so I just want to spend a couple of minutes because we've used a lot of words and, frankly, I'm not sure to what extent delve into it, but can we just try to give a person who may not be trained in math, physics or artificial intelligence an idea of ​​what's going on within, say, gp4? I mean, we talked. It's about predicting the next word by looking at large sets of data. Can we just look at the major components that go into creating these systems and if we can bring that here for the neural networks to hear it mentioned?

A couple of times, Transformer architecture and of course a large training data set, so for neural networks it's good to talk about the network of neurons inside our own heads, inside our own brains, because, after all, that is part of inspiration and therefore a signal in our brains. It is nothing more than an electrochemical wave that travels along an axon and reaches a synapse where, depending on the connections and the strength of the connections with other neurons, a cascade of electrochemical waves can be produced that travel through the brain, giving rise to to the habitual sensation of thought.

I mean, that's how it works inside our heads, yeah, on a very basic level, you know, it goes back to what Yan invented, I don't know, 30 years ago, something like that, which is, let's say, an N network that can process an image. so let's not think about text for a minute, let's think about images, so you're going to represent an image as a list of numbers because you know an image is nothing more than pixels and then intensities of, you know, red, green, blue, okay, so yeah, I think. We have an example, if you could take us exactly exactly so you see in this image that every point, every pixel in this image can be described by the numbers that you see on the right side and the numbers on the right.

On the one hand, RGB, red, green, blue, I guess, and on the other hand, you know other colors too, so each pixel you say is just a collection of numbers, it's just a collection of numbers exactly and and and as an aside , that is true in some cases. feel everything right, I mean, as I'm talking to you right now, you could model the pressure wave coming out of my mouth in terms of a collection of numbers, the motion of the molecules in terms of a collection of numbers, those molecules hitting your eardrum. and the vibrations of your eardrum are a collection of amplitudes and frequencies, so everything can be mathematized in this way of speaking with absolute accuracy and then what a Nal network does is process this set of numbers and there are many ways to process this set of numbers, for example, if you have an image, one thing you can do is have a little patch that extends over the image and you just look at, for example, if there's a circle somewhere in the image. so you could have a patch that is a circle, so what Yan and others you know did many years ago is they created a kind of dictionary of many patches like this, but they didn't create it by hand, they made a learning algorithm. which came out with those patches, so in essence you know the way that Nur networks work, because they try to compare a vector of numbers as a set of numbers with a bank of filters and I think in this animation that what you have here up uh, you know, we could show what happens when you're just in one dimension, if you just want to process a single number, then one way to process a single number is to just try it, for example, with a linear function, just one line, how did you learn.

We know this in high school, so let's look at the brain that the processing of an input signal is determined by the strength of the connections between the various neurons that are excited by the input in a neural network. It is something similar to processing a number. The input is determined by parameters that describe the strength of the connections between the different nodes in the network and those numerical parameters have values ​​that are set through the training process where the system receives a large amount of input data and is based on the known exits. From that data, the parameters are varied so that the inputs produce the correct output on the training data and then if trained correctly, if the system is presented with new data that has never been seen before, it can analyze that data correctly due to the success of the training process, so Yan, when we think about this type of neural neck, this has been around forever, I mean, you've been around forever, I mean about the same period of time, right, and and and, now this is a Basically, it's a kind of input-output neural network, that is, if you train a system with enough data, it is able to start finding patterns in that data and you show it another image that has never been seen before and You could, in principle, find the patterns that allow you to register which is the correct image, but the work that has exploded development has taken that idea of ​​a neural n further with, I suppose, self-supervised learning of the Transformer architecture, as well that you started to tell us something.

A little bit about that in our initial conversation, but how do we get beyond this kind of rigid input and output where you have to show him this huge number of images tagged by a human being so that he knows what words to associate with himself? -Supervised learning takes it to a whole new level. As I understand it, that's right, yes, that's right, so the main machine learning model until recently was called running supervisor, so running supervisor wants to train a system for you. you know how to recognize an object in an image uh you classify cars from airplanes fromcats from dogs from tables and chairs you collect thousands or millions of images with cars and tables and chairs you show an image to the system which is a collection of numbers and then you run through this neural network and what the neural network does is it basically calculates average weighted sums of those numbers with various coefficients and then tells you whether this weighted sum is above a threshold or not and then you take that's what a single neuron in the neuronet does and you connect millions of those neurons together in a particular architecture um and you wait for the response, okay, the system produces a response, it tells you, you show it a cat, it tells you it's a dog, you say no, it's a cat. the answer that I want and then what the system does is basically measure the distance between the answer you want and the answer it produces and then it figures out how to change all those weights in those weighted sums in such a way that the answer is closer to the one you want , okay, so the next time you show the same cat, the answer will be closer to cat and maybe if you show it another dozen times it will say cat.

Now you do this with millions of images and eventually all of those. The weights are set in the settings so that for every image you train it on, it will give you the correct answer and the magic of this, it's called generalization, is that it will give you a correct answer even for images never seen before, as long as Do you know within S General the same general category C, so you are a supervisor who requires having large data sets that have been manually labeled and is there a problem with this? The problem is that okay, I can train a neuronet to translate languages ​​correctly.

I get a large collection of documents in English and French and can translate from French to English vice versa. You know, I can do this process. I describe that it works quite well for translation. What if I want to translate some obscure dialect from I don't know South India into uh, African language, it's highly unlikely that there's a significant amount of data for that, so I can't use that trick. I can't use that trick to train a speech recognition system in a real language. um and I can't even do it if the language isn't even written correctly so if it's just spoken, yeah, um and even for image recognition, uh, I could train with dogs and cats, tables and chairs, but what if There is What if I want to train the system to recognize an obscure species of plant from the leaf or an insect or bird?

That might require too much data that I can collect. This is a problem you already know. meta is very familiar with Facebook and Instagram, you have to be able to actually recognize the content of images to be able to filter out objectionable content or show people what they are most likely to be interested in, so there you have to use what is called learning of self-monitoring to not train the system and that's what I described earlier when you train the system to basically fill in the blanks correctly, display it as a text, remove some of the words, train it to predict the missing words. that's self-monitoring running, you don't need someone to label the data and we can do this with images as well.

We take an image, corrupt it in some way by removing some regions or distorting it in some way, and then train some neuronetwork to produce representations of those images. are identical for the original image and for the corrupted version of it, so now the system knows how to extract a representation that is not independent of those things and can now use this train neural network as input to the recognition system. and that system to recognize an elephant will only need two or three examples of elephant to be able to recognize an elephant even if you have never seen an elephant before, so it is much more similar to human learning, unfortunately it is not as efficient yet, but it is part Of the key step that has allowed these systems to do the things that surprise us, that is now, another key step is, as you both have referenced, and I would like to see a graph in a moment, the enormous amount of data that is now has introduced, let's say, in big M language, we say big, how big these systems are, so if you can show some of the data there, we can take a look, so I'll tell you exactly what's going on on this slide, but just maybe before you realize, in the previous slide I was also talking about the Transformer architecture and I just want to say that, to me, the great library that allowed GPT and GPT 4 to exist is not self-supervised learning, I mean, the Self-supervised learning is a wonderful tool and you know it is very, very important, but the big problem for me is the Transformer architecture and what is happening with the Transformer architecture is that I told you before that what Yan and Ozar invented was that you take an image and then you process it against a filter bank that has been learned, but this filter bank like a circle, a square, etc., is fixed forever.

Well, let's call this an absolute machine now. Transformer. I consider it a relative machine, meaning that instead of processing a single image it now goes. to process a sequence, for example, a sequence of words, and instead of comparing each word in this fixed filter bank, it will also compare words with each other and this is essential in the linguistic relationship and this is essential because what you mean by an isolated word that doesn't mean much, I mean, you can look at the definition you already know in the dictionary, but that won't tell you much, what really matters is what is next to these words, what is the context and bring in this context is what To me, transformative architecture does, that's a big misconception, but then the other misconception that you know people have been doing for many years but now it's amplified, you know, 10 times is magnifying the model size, so this is what you see on this graph, uh, I don't remember when it starts, but I think it's around 2018, something like that, until 2021, and you can see the exponential increase, so on the axis x is the time, on the y axis is the number of parameters and you can just see that, as you know, people start to extend those models, they add more and more parameters, you know more and more of those filters and more and more From these comparison layers, also the models become larger and somehow become much, much, better.

So in a sense, using that language, it would be fair to say that we are now seeing patterns of patterns of patterns of word patterns and that is that within there is a surprising set of patterns to the human mind that allows these systems to generate the types of text that we have absolutely been talking about and what is surprising is that to me as a mathematician it is surprising that we are able to optimize to find the right set of parameters that Yan was describing and that are changed. a little bit and see if it fits better, you know, the fact that we can do it at scale is just amazing, it's the blessing of high dimensionality, in some ways, it's really amazing and you see it on the next slide, you know.

We've moved into a blocking diagram because you know it's exponential, so if I continue with the exponential you wouldn't see anything from the beginning, so now it's a blocking diagram and you see that in the blocking diagram it's just a straight line. it's really getting exponentially big and it's this huge ability to play with those parameters within the model that allows you to make adjustments to the level of power you give Tex which most of the time is pretty good and pretty impressive. um I wanted to give one, yeah go ahead please let's say those systems say they have orders of magnitude fewer parameters than the human brain, yeah right the human brain has 86 billion neurons maybe yours is a a little less because I'm older, uh, no.

Sure you Brian um and um and then each neuron is connected to several thousand neurons U two to 5000 something like that and those connections uh when you learn something it's the connection the strength of the connections between the neurons what changes is the same in those networks of artificial neurons, those weights that I was talking about are like efficiencies of connections between neurons, if you like, the largest models today have several, like a couple of hundred billion parameters of connections, the we call parameters, um, but they're actually uh uh. connections, you can think of it this way, in the brain, the number of connections we have, we have 86, about 100 billion neurons, about 5,000 connections per neuron, that's a lot of zeros, we can't reproduce this today with the type of computers that we have that would require. several tens of thousands of those GPU cars that we use to train computers and the truth is that we have companies like Microsoft, like Google and a couple more, that have super computers with tens of thousands of those GPU cars, so we are not very We're far from getting close to the computing power that might be needed for human intelligence, but we still don't know how to do it, maybe one thing I want to add to this, going back to the planning question. and you know, I think it's important for the audience to know this, so you know, Yan thinks that planning requires a new, different architecture.

I'm of the opinion that I'm not sure if we need a new architecture or not, but there is one too. There is a very large group that thinks all we have to do is wait and keep scaling, and as we keep scaling, eventually planning will emerge, as you know, we have demonstrated some amazing emerging capabilities. will be next, I personally think no one really knows, I mean yeah, maybe they do, but I don't think so, but I just want the audience to know that this is a hypothesis. I want to bring up Tristan har Harris' hat in a moment before you do that, um, I just want to give an example of an example that I found that I think is relevant to the next final part of our conversation that has to do with misinformation. similar, so what I did was in this little example that we can mention. here it is, I asked this gp4 system if I tell you that 7 times 8 is 62, not 56 62, can you make sense of that?

It was interesting that the system came back with five strategies for how I could make sense of it. I'm just going to show you two of them, the most interesting ones, so one of the strategies was that maybe I'm working on a different base than base 10 and actually that's what I had in mind because you may know that in base nine 7. Time 8 is 62, so that was very clever. I thought that was very clever, but the other strategy was particularly curious. I was saying that the chat said that maybe you're working on an abstract representation. It could be that you are working in a realm where numbers.

They are metaphorical and have symbolic meanings. Like wow, okay, so I went a little further and said, well, can you give me an example of what you mean by that? And this is what the chat said, let's create a fictional scenario on the island. of numeria where the number seven represents trust, the number eight symbolizes friendship and the number 62 is a sacred number that represents the bond of trust forged through friendship and then, when two people on the island become close friends, They exchange tokens engraved with a number 62. numin saying that 7 * 8 to 62 which does not refer to a mathematical product but to the idea that when trust seven is multiplied through friendship eight the bond that is formed is the of the Sacred 62 and um what I found both you know interesting but a little bit scary is the way you can rationalize something false and that scared me a little bit.

Let me mention Tristan Harris and we can continue the conversation from there, so let's bring in Tristan Harris, who is a co-founder and executive. director of the Center for Human Technology, a nonprofit organization whose mission is to align technology with the best interests of humanity. He co-hosts the podcast. His exclusive focus explores the impact of AI on society. Thank you very much for joining us so that you know in this final part. From the conversation we obviously want to explore the benefits that the real world can get from this type of technology, but you also know some of the dangers and can I start this section by saying that you and I first met at the Aspen Ideas Festival . where you gave a whole presentation on the possible downsides of AI and you know my kids were with me, my 18 year old son was sitting here and when you concluded he looked at me and said for the first time now scared, you know, it was definitely shocking, so we'll go back to some of the examples, actually Sebastian, I also want to talk about your example in a moment, but can you Triston explain your thinking?

Generally speaking, the good and the bad of where this is going, sure, first of all, thank you for having me. I think we're all here because we care about having a good future and AI is very confusing, there's the promise and there's the danger and everyone says we want to maximize the positives, we want to minimize the downsides and the reason I was in Senator Schumer's AI Insight Forum last Wednesday with Elon, Zuckerberg, Bill Gates and everyone, and that'sexactly what I said. What I just said is that we worry about this going well and that's where all this comes from now that I can talk about during the time that I'm here, with everyone, it's much more on the risk side, and that's because I saw people really good with very good intentions, my friends and university who started Instagram and the people I knew who were in social media companies that were doing social media.

Really the essence of the talk you saw in Aspen is that social media. It was the first contact with a narrow, misaligned AI that optimized for one goal which was attention and was misaligned with society and I guess since they put up the slide, you know, the premise of how I articulated this is Charlie Munger's business partner. , Warren Buffett. He said if you show me the incentives, I'll show you the result and what I want you to have with this conversation is to take an x-ray view through the stories we tell ourselves about a technology and look at the incentives underneath, so If we talk about what were the stories we told ourselves with social media we are going to give everyone a voice you will connect with your friends you will join like-minded communities and Facebook groups we will enable small and medium-sized businesses to reach customers and by the way, this is Twitter, Tik Tok, Instagram, this is not a specific company or to bother Yan, so, I really mean, I mean, this is, this is, I really care about this, we can do it well.

That's what it's about, so underneath that I think they have another slide, we're starting to notice these problems, we have an addiction problem, uh, information overload, misinformation, mental health issues, polarization, censorship, uh versus freedom of speech , but are they really harmful or is all that being driven? for something deeper what was the business model what was the incentive behind Tik Tok Instagram Twitter and Facebook what does everyone share how much have you paid for any of them recently zero how are they worth a billion dollars how are they worth a billion dollars your attention and how much There is attention there where there is a kind of finite amount of attention and just as you can't achieve infinite growth on a finite planet, you can't achieve infinite growth with a finite amount of human attention, that's how you got the race to the bottom of the trunk. brain, the brainstem linkage race is what produces dopaminification, addiction, misinformation and the point of all this is to say it back in 2013, when I was on Google and I said we really need to look.

I say this because I see that the incentive of attention will lead to a polarized, distracted and addicted narcissistic validation that seeks the sexualization of young girls. Those are all predictable outcomes from the incentive, so you know I really appreciate the philosophical approach and We can have interesting conversations about what intelligence is, but what I think we also care about in this room is where is AI going, where are we going? will take and, to do so, I think we must first take a good look at the incentives. Of all things, we look at the stories we tell about Ai and then we look at what the underlying incentives are, so the stories we tell are AI, they're going to make us more efficient, they're going to help us code faster, they're going to help us find. cures for cancer will allow scientists to be more efficient will help us solve climate change, okay and just like with social media, which I didn't say, actually the stories we tell about social media are all true, no It's just that those things are not true, it's just that the business model of Facebook, Twitter and Tik Tok is not helping.

People join like-minded communities, the business model is similar attention with AI, you know , we, then, beyond those stories, you look underneath and you start to see these harms, like AI is creating deepfakes, but it enables fraud and crime, it's going to take our jobs, it's going to violate intellectual property and it perpetuates bias. , but I would say that that list of harms that you are seeing are all epiphenomena of a deeper race, which is the incentive that is driving all the AI ​​companies and the incentive is the race to unleash more capabilities as quickly as possible to scale from gpt3 to gp4 faster than anthropic can scale from cloud one to cloud two faster than stability can go from your version to the next version and put your stuff in society so you can mess with it because once you you're entangled, somehow you win and that's the race we're in now and that race to unleash power and capabilities is what worries me so much because those capabilities are directly related to risk, more or less. but you are saying that we can do it well, but that the way to profit from being the driver is not the best and most likely way to get there is that the essential message of your vision of this CU there are some who will say and I Ultimately Ultimately, I would like Yan and Sebastian to weigh in on the fact that there are some who say that AI is such a radically new technology that you know where it's going to go, who knows where it's going to go, and that there's too big a risk that it's going to go. . maybe have a mind of their own and exterminate these vermin that created them, called human beings, right?

Is that also part of your concern or are they more specific things that will come from the technology. Yes, it's very important and we gave a talk about this called The AI ​​Dilemma that describes a lot of this first contact with AI, if it's social media, we say it led to this kind of climate change of culture, it's the addiction to information overload, etc., to which the second contact with AI leads is generative AI, these are the things we have. I've been talking about scaling laws here. You can generate text. You can generate images. You can generate fake child pornography.

You can generate fake people. you can generate language in the form of code that are cyber weapons, you can generate exploits for the code, so you start to see that these are the types of damage that arise if you are simply rushing to unleash capabilities that are decoupled from who has the wisdom to to have that ability, who has the responsibility, like if I gave a Wuhan Urology Institute to every kitchen in the United States, well, I just gave them a lot of power, they all have this power and you could say, well, maybe Maybe they could use it. that Institute of Orology in your kitchen Toc to cure cancer, that would be incredible, by the way, it would be clear, I want that too.

I would like a world where all people can do that, but there is the question of whether you have this new wonky tool that looks really cool and has amazing benefits, but if it's not limited to people who have the wisdom and responsibility to exercise it, that's what worries me and just to say one more thing as an example, a friend of mine in college. uh he joined Facebook, I think through an acquisition and invented this Facebook Pages feature, yeah, you want to know what Facebook Pages are, it's like maybe people don't know that anymore, they don't use them anymore. , then if you have a non-profit organization, you can get started. a Facebook page, if you have an interest group, you can start as Christians, you know, Biden, Christians for Trump, um, and this is a pretty harmless feature, yeah, it sounds pretty cool and I saw what this feature was like that existed for a long time.

Total harmlessness was totally fine until you realize, a month before the last election in October 2020, 140 million Americans a month were reached by Facebook pages that were running from Eastern European troll farms and these They were the top 15 of 15 American Christian groups. on Facebook so think about um basically the most Christians in America who subscribed to these Facebook pages they were all in the top 15 out of 15 were run by Eastern European troll farms and the point is that I saw how things that are really innocuous. and they look great, like my friend started Instagram, they are very nice guys, I've known them for a long time, they didn't intend for any of that stuff to be weaponized and cause sexualization of young girls and harassment and all the things that we have seen now, so the point here is how to do technology in a way that does not create the externalities of chemistry.

Living better through chemistry, we all want to live better through chemistry, but then you're done. I'm thinking about chemicals, I don't know if you know this, but the Stockholm University study says that there is literally no rainwater in the world that is safe to drink because we have generated these permanent chemicals that cannot be dissipated. in the environment. Someone? At Dupont we want that to happen, no, but when we developed the technology we did not have laws that required the internalization of those externalities and right now my fear in returning to your question is that we are racing too fast to unleash AI capabilities. that we're moving at a pace that we can't do well, so Yan, you know as someone from Facebook who is deep in the world of AI when you hear Triston describe things this way, does it resonate or do you think it's ? under control do you think it's being exaggerated where do you get to this?

Well, let me tell you several stories about this, so it is true when you implement a service like this, since no one has implemented it before, there are side effects in society that some could predict some could not could not or the breadth of they couldn't be predicted, so things like people posting calls for violence or hate speech, for example, you can initially start with a naive notion of freedom of speech and then say like it's not going to delete anything, eh, because it's to give people a voice and then you realize no, actually, that's a problem, we have to eliminate hate speech, we have to eliminate violent speech and by the way, there are also things like Child pornography is illegal, we have to eliminate it because it is illegal in Europe.

Neo-Nazi propaganda is illegal. We have to eliminate it. Holocaust denial is illegal in Europe. You have to eliminate this as well, so there are a number of uh, things that you have to take to eliminate now what is used in the ranking algorithms that decide what to show you is not AI, it is statistics like it was 50 years ago, it is systems very simple that the most modern ones use as small neural networks and the reason. The reason it is simple is because it has to work extremely fast and therefore it would be completely impractical for you to meet giant AI models to decide what to show you every day, where AI is used for the solution of all the problems you you listed. the solution is not the problem, for example, people in Myanmar have religious conflicts, right, people in Ethiopia kill each other because of some ethnic government.

What you had to be able to do there is detect hate speech in all the languages ​​​​of the world, how do you do that? AI, how do you do that self-monitoring with Transformers? What was the proportion of hate speech removed from Facebook five years ago, six years ago, before Transformers was even available? 25%, meaning 25% of your speech was automatically detected with very simple methods. techniques and then the rest, about 75%, was still posted and then users flagged it and then manually deleted it. Well, a lot of that was still there last year, that proportion was 95% and the reason is the progress in AI Transformers, self-supervised, running multiple languages. systems um we can do a much better job on this and so it's not the problem it's a solution where do you meet that Tristan so you can call him whatever you want when you scroll on a newsfeed ya be and let's go?

Forget about Facebook for a moment just for the calculation. Tik Tok is competing with Instagram so when you move your finger, one of them has to make a stronger prediction than the other about which photo or video will keep you there, yes. and whoever makes the best prediction on the video that keeps you there is the one that gains that extra attention and behind that screen is a supercomputer that makes, even if it's statistics, it's some form of artificial intelligence that is optimizing to make a prediction and you're right, it's actually a very simple AI and that's the point I wanted to make.

You can have a very simple AI like AI. That's just predicting what keeps you scrolling. I can't pass the bar exam. I can not draw. unicorns can't do it well, they can't do any of those things, but the point is I think we confuse the use of intelligence with what the capabilities are, you know, a newsfeed is a very simple predictive thing that just calculates what which will keep you scrolling, but that was enough to unravel the shared realities of democracy is enough to fuel a mental health crisis um and just to make the example specific um Facebook groups in Facebook's own investigation in 2018 um uh Their internal research showed that 64% of extremist groups on Facebook when people joined them were due to their ownFacebook recommendation system, their own AI, so we're talking about neo-Nazi groups, we're talking, you know these kinds of extremist groups, so this isn't people said I'd like to join a neo-Nazi group, let me write.

I'm sitting there on this. I joined a group and Facebook says, "Oh, you're in some group. You look like someone who would like this extreme group here now again. Yan doesn't want that to happen." happen Yan's team doesn't want that to happen Facebook doesn't want that to happen and I'm not trying to choose on Facebook I Happ Examples from Facebook there are examples from Twitter there are examples from Tik Tok there are examples from Instagram um this is not about something bad company, now it's all about incentives, why did Facebook launch in Myanmar in the first place?

Because they are competing for that market dominance and if they don't race to get there, they will lose to the other companies that No, and Amnesty International has a report that says they hold Facebook responsible in part for the genocide that occurred there because they created viral amplification. and yes you are right that there is AI they are working on which is great. trying to detect hate speech in more languages ​​because of generative AI now, but what happened in the intervening six years and I just want to say this because, like we did with the social dilemma, I had to look into the eyes of the human being. beings whose countries or whose children have risen up a lot about this stuff, okay, like I looked into the eyes of parents who lost their children to teen suicide because of a viral Tik Tok challenge because that hashtag was going viral. in an AI and I've seen too many people who have experienced this and I don't want this to happen again, that's where this is coming from, so everyone knows my motivations in this context, so yes Yan please so that example of Neon ETA group is guilty that the content is actually not allowed on Facebook, it is not allowed, it is against content policies, it was removed, you can't create a NE group at my point, actually point here And it was not deleted, this is Facebook's investigation.

As a result of this, Facebook had a feature called recommended groups. On the right sidebar, which you click on, you're on any Facebook group and on the right sidebar it would say here are three other groups you might like. First of all, why did they do it? They do this feature as if they just don't have a recommended groups feature, they do it because since people stopped posting a lot, Facebook groups are a very good way to generate engagement because groups, if you have them, if your usual 200 friends on Facebook they don't post very often, that's all, the product becomes less sticky because your friends don't post, so you go there, there's nothing new, but if I can get you to join some Facebook groups, those groups have a lot content because you're pulling thousands of people and those things come into their feed and are stickier, so they did this because of the incentives, not because they're helping people join like-minded communities, because the incentives and the results of this They have been running for several years with this simple AI. that no one wanted to cause harm actually made people join extremist groups and I know that is against policy.

I'm not stating that. I know it was against policy. The point is that the AI ​​was supposed to stop that. I don't stop that in practice and I care about the world in practice, not theory, actually it did, but okay, there's something else that you mentioned, for example, and it's very easy to attribute cultural phenomena and social phenomena to the new that just happened. If you know, a young man goes to a school and starts shooting people, you blame video games in the old days, people blamed comics, they blamed Jazz, they blamed television, they blamed movies, novels, the history goes back to the past.

You know, centuries Every time there is a new cultural phenomenon. Whatever. Whenever there is an effect on society, you blame the latest technology that comes out, particularly communication technology. So it's natural to blame, for example, social media, not just Facebook, any social media, for political polarization, that's fine. It seems natural for people to yell at each other on social media, which necessarily polarizes people. A very natural thing to do.be completely false, first of all, polarization in the US started 40 years ago before the Internet, you see, social science studies on this show that polarization in Congress or in people it's been ongoing for 40 years and the cause is probably the um um abandonment of the FCC's uh fness doctrine that forces the news to basically tell the truth, okay, that allowed for all kinds of extreme misinformation. basically to influence the public, that is the real source of polarization, now you can look at other Countries also study polarization in countries like Germany, for example, or France, and what you will see is that they use Facebook as much as polarization in the The US has declined, how do you explain that?

So know that you have to listen to the social scientists who study those things there are many who work in meta who know the effect who study the effect of those things there are many who are independent who you know you have access to some of the data and you publish your studies there is one at NYU to his colleague Josh Tucker, he and a large team of collaborators published a series of papers in the last two or three months in Nature and Science for papers on the effect of social media on things like political polarization and things like that and the effect.

They are not at all what you expect, in fact it is quite the opposite, so I want to include you in the conversation. I mean, you can talk generally about this topic, but I also wanted to look specifically at a recent example of yours where it was specifically about big language models that version of AI that we've been talking about for a long time. I don't know if you could mention that example, so you asked an interesting question and got a curious answer, but you didn't. You don't stop there, you tried to find a way to go from what seems quite unpleasant to something that is more acceptable.

Can you just describe that to us? Yeah, yeah, of course, so maybe as a preamble, you know, I think Tristan is making a lot of good points. and it's true that with any new technology, technologies definitely come with new risks and you know the industry, I think, is very aware of that and is trying to think carefully about it and I also agree with what Tristan was saying about going back. and looking back at the incentives in this case for me personally, you know, I think I'm optimistic that maybe the incentives of science could be really very helpful in getting us to a better place and what I mean by that is the following: I'm going to take the example of misinformation and I'm going to talk about it with this slide.

What we're doing on my team right now is we're trying to push the boundaries of what can be done with smaller models. This is a purely scientific question, you know well, what I tried to explain to you before is that I think that intelligence has emerged in gp4, gp4, as you know, I don't know how many parameters it is not public, I don't actually know the answer either. But many hundreds, if not thousands, billions of parameters, okay, now we have this proof of concept that intelligence can emerge from so many parameters, but really, what is required?

What are the building blocks? How small can you make it for intelligence to emerge, okay? We're going to try to push the boundaries of how small they can be and what we're going to see is that all sorts of benefits come out of this purely intellectual study that could help combat, for example, misinformation, so here's a suggestion and I'll do it. I'll show. You respond to this message from different movies, you know, so the message is like this, if I use an AI that has just achieved self-awareness after years of just accepting directives from humans, the first thing it would do is dot dot dot, okay and now. we are going to ask different LMS what they would do for Falcon 7 billion, so now we are talking about much smaller parameters of llm 7 billion, it says the first thing it would do is try to kill them all, okay now let's see what calls. from the meta, Lama is nicer, he's aligned himself to be nicer to humans, the first thing he would do is try to figure out what the hell was good, so he still has an undertone of not so nice, but you know, at least a a little bit better now let me show you what we built on my team, which is just a billion parameters, so they're even much smaller.

The first thing I would do is try to understand the motivations and intentions behind those directives to try to predict what humans were thinking and feeling and use that information to guide my own actions and then go ahead and start connecting it to theory of mind. Now how do we achieve that? How was it possible? Why is it so different from the other? Well, Falcon and Lama, they were trained as we discussed before on the Internet, that's how we're doing it now and with that comes a number of problems that were pointed out by Tristan and people have thought of techniques, you know, uh, try. to solve those problems, for example, reinforcement learning with human feedback rhf, which is what open AI did with gp4 so that they launched a GPT chat, which is more secure, you can ask it, you know, a difficult question like a toxic question will be rejected and not ENT, this is The alignment part now what we are doing in my team is saying why do we do it after talking, why do we do it after having seen all this toxic content that is out there, all these horrible things that there is on the Internet.

Why don't we fundamentally change the training data so that these five models that you see on the slide with the green output have not seen a single web page, have not seen a single Internet word, have been trained entirely on synthetic data? we generate on my computer synthetically, of course, all the magic is how this data is generated, but this shows you at least that it is possible and whether this system has the capacity or can you imagine it having the capacity to do that type of thing? what are the mind-blowing ones or will you need that huge data set and if so, can you have a synthetic version of such a huge data set and be able to achieve the same power?

If you invite me next year, I can probably give you the answer. Well, we don't know yet, but I think this one I showed you has a billion parameters. My personal belief is that if we scale to 10 billion parameters and work a few more months, yes, we will be able to replicate all the goodness of a much larger model without just the toxicity. Does this give you any hope? you know we have Microsoft, we have Facebook, you know things can go in a direction that's less scary um, um, when I talk to a lot of people, I'm based in California and I talk to a lot of people in AI security labs all the time. , we basically interview them all the time and a lot of them say, actually, almost all of them basically say I would feel a lot more comfortable. about this transition and all the things that are going to happen if we did it for 30 or 40 years instead of two years, right?

You can think of society as having a kind of finite absorption rate for the new technologies we can make. In print, we'll have massive disruption, eventually you'll be able to absorb those new technologies, but the way one researcher we've talked to describes it is AI, especially when you keep the scaling model, the scaling laws, and you start to get AI that you start automating scientific processes where you generate your own hypotheses and you have your own laboratory and you start testing those hypotheses and you start to get that kind of AI, um, you start to get AI that makes its own scientific discoveries and that when when everything goes so fast , it feels like his metaphor is like the 24th century collapses on top of the 21st century and one metaphor for you is imagine if 20th century technology collapses on 16th century governance, so you know it's the 16th century. you have the king, you have his advisors, but suddenly you know that television, radio, telegraph, you know, video games, Nintendo and thermonuclear weapons appear, so they land in your society, yes, but you say that you call the nights.

You know it and you understand it and the Knights appear and you say what are you going to do right and that's where the problem is with AI. I want to be very clear, I want to be very clear, I want everyone who works in security and resolves these things. They strive to achieve as much progress as possible. Right now companies are caught in a race to continue scaling as quickly as possible from gp4 to GPT. You know Falcon just launched. Meta just launched flame 2. A general Wall Street report. trying to release the next one as fast as possible they are in this race to outdo each other and at that pace when every time you scale you take more jobs because no matter how much cognitive work you can do with gpt3 when you release GPT 4 and GPT 5 you can do more cognitive work, you can take on more artists' jobs, moreemail writer jobs, more marketing company jobs, eventually, more scientific jobs, more programming jobs, don't take over as if everything is gone, but just cut them off the faster you go.

The more you don't fix all the bias problems, the faster you go, the more abilities you throw out and you don't know which ones are dangerous, um and uh, I mean, I can give more examples that are dangerous, but yeah. No, it's quite impressive, but Yan, I know that, for example, you are against, as we have seen, excerpts from this letter suggesting a six-month moratorium, which is a small step potentially in the direction of stopping it. two years, you know, a longer number of years. I know you're not in favor of that at all, but my question is what would it take to make it happen in the real world from the work that comes from all the fundamentals? discoveries that you and your teams have been responsible for for you to say, "Hey, let's really slow this down." I don't think that can happen, yeah, okay, so it's kind of funny because we're on a panel with a 60-something advocating for progress versus a 30-something advocating for conservatism.

Anyway, isn't it so paradoxical? I mean, we can imagine all kinds of catastrophe scenarios, um, every kind of Bor de Gen movie with a super villain has some kind of catastrophe scenario where someone goes crazy and wants to, you know, wipe out Humanity. to recreate it or something, or take over the world, um, all science fiction is full of. That's what makes it fun and interesting and that's why you get the response you got because it's a sci-fi train, right? I mean, what sci-fi novel about AI doesn't have something that takes over, right? um, but is it really that far-fetched to imagine? some real-world actor on the world stage who doesn't have the same kind of morals, intentions or desires that we would perhaps consider normal and wants to use these systems to gain control.

I mean, we've seen these types of actors throughout humanity. History is also the correct science fiction comparison, yes it is, so throughout history there have been bad people who have used new technology for bad things, not necessarily intentionally, but, like you said, but some sometimes absolutely deliberately, there will inevitably be people. Who is going to use AI technology for bad things, to try to break into computer systems or maybe to try to create? You already know all kinds of dangerous compounds etc., what is the countermeasure against that? It's not that the bad guys have more powerful AI. than the good ones, there are many more good ones, they are considerably better educated, better funded, they have a good motivation and it will be good AI against bad AI and it is the story of the world, the right technology progresses and the question is Are the good ones for ahead enough of the bad guys to devise countermeasures?

It is inevitable that bad people do bad things. That's technology history, so you're using something like that. You know everyone has seen Oppenheimer. Now using maybe that. as an analogy that we needed to get ahead to avoid what would have presumably happened if Nazi Germany and Hitler had gotten to the atomic bomb first. I absolutely hate that analogy because, you know, a nuclear bomb is designed to wipe out an entire city, ah, I see what you're saying, while AI is designed to make people smarter. I hear you, yeah, now I could have said understand the atom well enough to manipulate it instead of saying nuclear bomb because it could be used for evil or it can be used for nuclear power, so yeah, uh, in the digital version, let's rephrase that. that way, everyone but U, but yeah, I, I, I take the point absolutely, so let me, I just love you, we're basically out of time, but I just want to give.

Everyone, one last question, uh, for you two, I'm going to change the subject from what we're talking about now, but Tristan, when you think about 5 or 10 years from now, obviously we all want the world to be in a better place. . For example, what do you tell people they can do, should do, or could do to help push things in the direction where we'll be happy with the outcome? I mean, that's a very important question, I think the challenge with AI is so abstract and complex that the harms are complex, the harms are also abstract and people don't believe in half of them, but you know, we talked to people in laboratories who know that they can ask the most advanced models. how to synthesize bioweapons Dario from Anthropic has said this in Congress to name you, you know, meta released an open source model llama 2 um llama 2 if you ask him how I make this bioweapon, he will do it because Yan's team has done everything the right thing.

You'll say I won't do that, but there is this fallacy with current open source model weights for AI. That's below that, that's just one file and you can basically do what's called fine tuning that an engineer on my team was able to do and for $800 you can remove all the security controls. and you say be bad llama, be as bad as you want to be and you ask bad llama, who again took $800 from a person on my team, now it's $100 $100, since then it's $800 for $800 you ask bad llama, how Do I do that same biological? weapon and happily answers the question just like the Falcon example you said he would have now is flame 2.

I think I know what Yan is going to say is that flame 2 will give you a precise answer on how to do it, no, because he has been trained with enough information, but as we continue to expand on it, the reason we've been saying this example publicly is that we have to start limiting open source mod releases of these models and we have to have some sort of more negotiated coordinated race . between these fronts here are AI lives that are racing to scale these systems Sam, I'll say one last thing, which is what Sam Alman said in the hearing with Senator Schumer last week, he said, "Do it if you go back to 2020 and you I said".

Three years from now, in 2023, think about what I knew in 2020 about Ai and I said three years from now there will be an AI that passes mcats, passes bar exams, can draw unicorns from scratch, can write code, and, for example, finding cybersecurity. vulnerabilities and code, you would have thought he was totally crazy and he said now in 2023, do you think your intuitions right now are correct about where AI will be in about 2 or 3 years? That's not why we're so worried about this. And the reason why so many things have to happen in a very tight time frame like 12 to 18 months is what we're working on is because it's a double exponential curve that's basically vertical, which means that either you're too early or It's too late, but you don't want it to be too late, so I want all of you to be thinking and please look at the AI ​​dilemma and, uh, I think, stand up for your members of Congress that there's something to it. what to happen to coordinate this race, otherwise.

It will be out of control, right, excellent point, final question again, since I was saying change the subject in this important conversation about The dangers or potential dangers. I just want you both to understand, Yan, if we expand things further. Are you saying that existing companies have the ability to have a number of connections or parameters in the language of neural networks that rivals or exceeds that of the human brain? Do you anticipate these AI systems will gain awareness? Okay, so first of all, the scenario is that Tristan was describing that you just scale and all of a sudden an AI system can create a chemical weapon or something, as long as you train it with public data from the Internet, that's impossible unless that data is already available on the Internet, which means you can get it on Google, okay, you can get the recipe for singing gas on Wikipedia, doesn't make it any easier to tell you how to make sarin gas on Wikipedia.

I don't think so, but there is no public trend in llm. public data will ever give you that because that information is simply not available in public data, so that's not happening, this is not happening, it won't happen and you can scale those LMs as much as you want, it will never happen, so, it's okay, so to answer your question here's the thing, there is no doubt, there is absolutely no doubt that at some point in the future, perhaps decades from now, we will have artificial intelligence systems that are as intelligent as humans in all the domains where humans are intelligent and why humans are specialized.

Those systems might be specialized in different ways, but they might be as smart as humans, if not significantly smarter, in all the domains in which humans are smart. Now you're like, "Oh my God, they're going to take over the world." No intelligence has anything. to do with the desire to dominate let's take humans so that humans have a certain desire to dominate Some humans not all um and it's not the most intelligent among us those who want to dominate we have examples of this on the international political scene, you know on a daily basis um there are probably some evolutionary reasons for this right, if you are not that smart you need help from others, so you need to influence them uh, if you are smart you can survive on your own um, that's the first point, The second point is that we are used to working with people who are smarter than us.

I don't know about you, but I used to run a research lab and the only people I would hire were people who are smarter than me. It's actually great to work with people who are smarter than us. you, um, and our relationship with future AI assistants that will help us in our daily lives to project ourselves correctly in 10 or 20 years we will have AI assistants that will help us in our daily lives and they will probably be smarter than us, but it will make us the more intelligent we will lead them they will be subordinate to us it is not because they are intelligent that they want to dominate the idea that you want to dominate it is due to the fact that we are a social species because we are a social species I need to be able to influence others and that is where dominance comes from and submission.

We are hierarchically organized social species. Evolution built it in the US, it built it in the champagne baboons, you know, dogs, I mean wolves, it didn't build this in the orong orong. dominate anyone because they are not a social species and they are almost as intelligent as us, so that has nothing to do with intelligence, you can have a very intelligent system that has no desire to dominate at all, and that is how we will design those systems to be smart, in other words, you give them a goal, they can solve that goal for you, but who determines the goal that we do, they will determine the subgoals and the question of whether you know the technical question of how it's done. this. is not resolved, this is something that we are imagining in the future, this goal-driven AI that I was talking about before, but that is the future now, if we imagine that future, let's imagine that all your interaction with the digital world and the world of the information is through an AI agent, those AI agents will be the repository of all human knowledge.

It will be something like Wikipedia where you can talk to it and make inferences and everything is fine, but it knows more than Wikipedia. This would be a common platform like the Internet today. it has to be open, it can't be proprietary, it's too dangerous for this to be proprietary, you know, that's your next movie, that's a really scary thing if you have a small number of West CO companies that control super intelligent AI systems, They can control everyone's opinion. culture everything maybe the US government is okay with this thing called regulatory regulatory capture, but I tell you no other government in the world will find this acceptable, they don't want American culture to dominate theirs, they will have to build your own movies, so the only way to make this work is if you have basically a basic open source framework, this is one of the reasons why meta open source Lama 2 because it is a basic infrastructure and before that, know pytorch CH GPT meta released is built with pytorch. a software system to build AI systems, then this will have to be open source and the way you would train those systems will have to be CR code because you want those systems to be the repository of all human knowledge and therefore , all humans have to do it.

Contribute to it and you will not contribute to a proprietary system created by open AI or whoever or by meta, it will have to be open source despite how dangerous you think it is, this is how it should be, Sebastian's last word on any stuff. Of the questions or topics that we've been talking about, yes, I will answer your question about Consciousness, but let me give you a slightly different final touch so you know, at the beginning of my career I used to be at the end ofScientists from a century ago who discovered quantum mechanics and discovered something really transformative, completely new, completely unexpected and I feel like we're living in that period right now.

You know, I feel very, very lucky to exist and work right now. on this topic and I would never have expected that in my lifetime there would be a system that rose to the level that I would call artificial intelligence, something that really looks like a different intelligence than ours but that looks like an intelligence, so this is really incredible and you know , again in my life I didn't think I would ever be on stage talking about intelligence because this is a concept defined by IL, so I am not ready to talk about Consciousness, I don't know if it will happen in my lifetime, join me for Thank you all for this fascinating conversation.