Speaker 1 0:15

Hi , this is Omid Ashtari and Natasha McElhone . Today's guest is Ali Islami , a principal research scientist at Google DeepMind , studying artificial intelligence . He's currently also director of research strategy for Google Gemini , which is Google's leading AI model . Prior to this , he led a team working on generative models , self-supervised learning and multimodal large language models . He also led the quantum chemistry and materials team in the science group .

Speaker 2 0:46

Prior to DeepMind , he was a postdoctoral researcher at Microsoft Research , cambridge . He did his PhD at the University of Edinburgh , where he was a Carnegie scholar During that time . He was also a visiting researcher at Oxford University in the visual geometry group . But don't be intimidated by any of this . He's super accessible and an amazing teacher . I can attest to this .

Speaker 1 1:10

We'll be talking about the emergence of the AI landscape .

Speaker 2 1:14

Whether you need a body to understand the world .

Speaker 1 1:18

Human perception slash Plato .

Speaker 2 1:20

The difference between how humans and AI learn .

Speaker 1 1:24

How AI models are built and trained .

Speaker 2 1:26

The differences between machine learning and generative AI .

Speaker 1 1:30

Marcus Aurelius and how amazing the human brain is .

Speaker 2 1:33

Whether we're about to surrender our sovereignty to AI .

Speaker 1 1:38

Let's log in . Hello , my name is Omid Ashtari .

Speaker 2 1:43

And my name's Natasha McElhone . Today we have Ali with us . Hello , my name is Omid Ashtari and my name is Natasha McElhone . Today we have Ali with us . Hello Ali . Would you like to tell us where you're from and why you're here ?

Speaker 3 1:53

Yeah , so my name is Ali . I am originally from Iran , grew up in Edinburgh , Scotland . I am an AI research scientist . I work at Google DeepMind and I've been doing AI for more than 15 years now .

Speaker 1 2:08

Amazing . That's quite old school , frankly , when it comes to this field . Tell us a little bit about why you got into AI to begin with and maybe a little bit of a play by play , how you've seen this field change from something that was relatively obscure , I guess in the beginning , when you got to it , to now where it's all the rage and everybody's talking about it , seen this field change from something that was relatively obscure , I guess , in the beginning , when you got to it , to now where it's all the rage and everybody's talking about it .

Speaker 3 2:31

Yeah , basically I got into this field , like so many of my colleagues , through video games . As a teenager I played a lot of games with my brother . What did you play ? A lot of Halo , actually . Oh okay , brother , what did you play ? A lot of Halo , actually A lot of Halo . But before that there were like flying simulators and racing simulators and that kind of stuff . But , jokes aside , halo on the original Xbox was like the game that really got me hooked and I was mesmerized by the worlds that I was witnessing on my TV screen and I wanted to know how they have been built . I wanted to make the characters , I wanted to figure out how they made them feel so alive . And that got me into computer programming . So I picked up books on 3D modeling and on programming . I remember I bought this massive , thick book on C++ and it took me years to get through it .

Speaker 1 3:23

But you know , that's how it started , you see the difference between you and that's how it started . You see , the difference between you and me is I have that book .

Speaker 3 3:30

I never got through it . Yeah , I mean , funnily enough , I didn't get through the 3D modeling book . You know , I got stuck . The version of the software that I had at the time , 3d Studio Max , was different from the version that the book assumed I had , and I just couldn't figure out how to make this damn firetruck that it wanted me to build . So I gave up on 3D modeling and I focused on programming .

Speaker 1 3:51

Which was a fortuitous fork in your life , I think .

Speaker 3 3:55

Yeah , and this must have been when I was 12 or 13 years old , and so from programming C++ I got into web design , got connected to the internet for the first time around then .

Speaker 2 4:08

And was this all outside of school or were there programs that were offered at your school ? This was all outside school .

Speaker 3 4:14

Interesting .

Speaker 1 4:14

This is still in Iran .

Speaker 3 4:15

This is now in Iran yeah , and the internet kind of allowed me to connect to the wider world . I was really active on forums where we would make websites , we would share software and , yeah , by the age of 15 , 16 , I was writing software that thousands of other people were using to host their own websites . So I got into like the whole business of open source and seeing how much impact you can have from your bedroom ?

Speaker 2 4:41

Okay , that was my next question . So most of your social life took place online , with friends who you didn't necessarily meet up with in person ?

Speaker 3 4:48

Not most , but a significant portion . Yes , I had plenty of friends in real life as well . We'd play football , we'd play video games , that kind of stuff , but I also had , I remember , one of my close friends at that time was an American that to this day I've never met Trevor , if he's listening . And yeah , we used to make websites together and we used to do projects together , and so by the time I was old enough to go to uni , I was actually quite proficient in programming . I basically knew how to direct computers to do the ideas that I had in my head , and the way you do that is by implementing algorithms as code . An algorithm is basically like a recipe , so you write down the recipe for what you want the computer to do and it does it . And yeah , by the time I was 17 , 18 , I basically knew how to do that .

Speaker 3 5:38

And so the question for me was why is it that there are certain things that we want computers to do , but we don't know how to write an algorithm , for Not just me , but nobody on the planet knows how to do it ? Why can't I get a computer to look at a photo of a cat and to recognize that it's a cat ? Why can't I write an algorithm that allows the computer to talk to me ? Why do I have to talk in this archaic computer language ? Why can't robots walk ? You know questions like that . Why can't cars drive themselves , and so all of these questions of why and why not ? That's basically the field of AI . So when it came to choosing my undergrad degree instead of doing just computer science , my undergrad was actually in AI at Edinburgh , and Edinburgh used to have a big history of working on AI in the 50s and 60s , and so they kept this AI degree through the AI winter , wow , whereas most other universities just dropped AI . You couldn't do AI , yeah .

Speaker 1 6:40

Was there any other university in the UK that had an AI degree at the time ? Sussex .

Speaker 3 6:51

Yeah , that's a good point . Yeah , but it's quite rare . I think maybe carnegie mellon in the states also had it , but there were very few and far in between the universities that stuck to that idea . Because in the 80s ai was really problematic , yeah , and there was no funding for it and it was kind of like a bit of a dirty word . Yeah , so my undergrad was in AI and then by the end of the undergrad dream came true I worked at a video games company . So that arc kind of completed .

Speaker 1 7:17

Which one was that ?

Speaker 3 7:18

Company called Rare . Yeah , you know like GoldenEye and Banjo-Kazooie and so on . And , yeah , after that experience , I felt like , okay , I'm really confident now that I can code anything that I want . And the big question for me was okay , how do we really push the boundaries ? And it was clear at that time . So this is around 2008 , 2009, . That machine learning was the only thing that was working towards pushing the boundaries a little bit .

Speaker 1 7:45

Can you explain what machine learning means in the context of just pushing the boundaries ?

Speaker 3 7:50

Yeah , so basically , as I said , computers operate by executing algorithms that humans write , and what makes a computer useful is that it can follow the algorithm very precisely and very quickly . So I can follow the recipe very precisely and very quickly can I just interrupt you one second ?

Speaker 2 8:11

so when we hear this as lay people , I've always heard that an algorithm is a set of instructions rather than the recipe analogy , which is . Recipe analogy is really useful , but for someone who doesn't code , what does that entail ? In putting those instructions , how do you do that and why did you think that the capacity for that was so limited at the beginning of your journey ?

Speaker 3 8:37

Yeah , that's a really good question . So I'll answer the second part first . Ai as a scientific field is a particularly interesting one , where we know that the thing we're trying to achieve is possible . The reason we know it's possible is because humans already do it . So if you contrast the dream of AI , which is to build something that's as intelligent as humans , versus the dream of , say , the field of teleportation , which is to teleport something from position A to position B , is that we know intelligence is possible , but we don't even know if teleportation is possible . We don't have any evidence to support that , but we also don't know if it's impossible . You know , when you look around nature , you look at animals , you look at humans , you see them do things like sing and dance and move and , in the case of humans , communicate and recognize friend from foe and recognize prey from you know food and all of that , so you can see it happening in nature . And then you wonder well , how do I replicate this ?

Speaker 2 9:48

But was the intention to replicate ? Because replication isn't so interesting as making something that will be able to do the things that you can't do , that you only dream of , like teleportation . Was that really the intention ?

Speaker 3 10:02

Right . So different people enter the field of AI for different reasons , and one reason you might want to do it is to replicate . Another reason you might want to do it is to surpass . A third reason you might want to do it is to understand . So for me personally , I was most interested in understanding .

Speaker 3 10:22

So I looked at a cat . I'm like , how does this being do the things that it does ? And there's a certain scientific way of thinking , which is that the only way I can claim to understand something fully is if I can reproduce it . Right , the reproduction is the demonstration of the understanding . So , for me , I wanted to use computers to reproduce behaviors to convince myself that I've understood those behaviors .

Speaker 3 10:50

And when I talked about algorithms not being sufficient at that time , what I meant was that they weren't sufficient for me to reproduce , as opposed to they weren't sufficiently good to surpass that behavior . So , for instance , like I wasn't particularly interested in coming up with a computer program that could play chess better than humans , I was more interested in figuring out how it is that a , that a , that a cat knows how to navigate the room , but I can't get my damn robot to do the same thing , right and in other words , you first have to match what's here before you want to surpass it , right yeah so of course we can surpass , once we've kind of managed to match what what's already a benchmark of a cat pattern recognizing yes cool , okay , so can .

Speaker 1 11:38

Let's get back to machine learning . You said the field of ai was pretty much limited to machine learning as a thing and it was working really well . Now , how does machine learning take the recipe further ?

Speaker 3 11:51

Right . So what we talked about earlier was that we have these algorithms . They are these recipes or these sets of instructions . The computer executes them and our role as computer programmers is to write down the algorithm . And in some cases we know how to do that . In some cases we don't know how to do that .

Speaker 3 12:08

So machine learning enters the room and it offers a new kind of approach . It says I understand that there are certain problems that you can't figure out how to write an algorithm for . In these instances , what I would like you to do instead , as the computer programmer , is to create a data set of examples of how you want the system to behave . So you say example number one I want the computer to take this input and turn it into this output . Example number two I want to take this new input and to produce this new output .

Speaker 3 12:45

So , for instance , example number three this particular image is of a cat . Example number four this particular image is of a dog . And you create a big data set and the machine learning system inspects this data set and figures out how to map inputs to outputs as best as it can . So the role of the computer programmer has changed a little bit . Instead of writing down the algorithm that executes the task of interest . The computer programmer writes down the algorithm for a learning system and then provides a data set that that learning system learns from in order to do the task that it cares about .

Speaker 1 13:31

Right , got it . So it's about creating the most efficient learning algorithm rather than about finding the right algorithm , and then it's about curating the right data set .

Speaker 3 13:45

Yeah .

Speaker 1 13:46

And then letting it loose .

Speaker 3 13:47

Yeah , and this immediately creates almost a separation of concerns . So you have some people who then become obsessed with finding the most efficient learning algorithm . And then you have different people who take learning algorithms off the shelf and they apply it to particular data sets to solve particular problems . So they might take an algorithm and use it to classify cats versus dogs , or they take that same algorithm and use it to recognize words or to do any of the other tasks that they're interested in .

Speaker 2 14:20

Can I just ask you something ? I don't want to go off topic , but it's just occurring to me that when we learn as humans and I'm not suggesting that the intention is to teach or to get to a place pattern recognition very early on in life and mistakes and also uncertainty when one of these algorithms is learning , how often does it create something that's useful and surprising rather than just a mistake ? You cited this example the other day viagra as a drug .

Speaker 1 15:10

it was being explored for heart disease are we talking in the machine learning paradigm or in the ai paradigm ?

Speaker 2 15:16

in the machine well , I guess yeah , both , but first of all machine learning and then we'll go on to ai , or if perhaps that question is only really relevant for the more advanced stages that we're in right now . But I'm just curious of what possibilities that we're not predicting are thrown up .

Speaker 3 15:34

Okay . So there's a couple of concepts that come up in this question and maybe we go through them one by one and then we synthesize into an answer . First there was this question of AI versus machine learning and just to put a pin in it , basically AI , in my mind , is a scientific field of study and it's concerned with understanding intelligence and recreating it . Machine learning is a particular set of tools that happen to be particularly useful for the field of AI and likely will always be a part of AI . I doubt it will go away . So ML in a sense is a subfield of AI . But also ML is used for things that are not even related to AI . So , for instance , you might have a very practical problem you have no desire to understand or replicate intelligence . You use machine learning for that . So machine learning is a lot more down to earth , in a sense , in its goals . Now let's talk a little bit about the practicalities of machine learning . As it happens , the process of a system as it learns .

Speaker 3 16:42

Basically , you know , what I described was a data set , a training data set , that the algorithm keeps looking at and it tries to improve its performance on that data set . So first we need to define what performance is . So performance on the train set , then this might sound a little technical , but it's actually quite useful to get at this point of creativity that you're asking about . Performance on the training data set means at this point of learning . So learning is a process . It takes time . At any given point in that process , how many of the examples in the training set does the algorithm do the right thing on ? Does it have the right behavior on ? So at the beginning of the learning process , 100% of the examples in the training data set it does poorly on . Hopefully , at the end of learning it will be able to do well on most of the examples on the training data set .

Speaker 3 17:44

In many cases it never reaches 100% performance on the training data set and this might be because the algorithm itself is limited or it might be because the training data set has contradictions . So what this means is that the algorithm is in a constant state of making mistakes . In a sense it's never perfect . It can't even fully describe the training set . It can't fully explain how the training set has the information that it has . So any algorithm , practically speaking , is in a constant state of making mistakes . It's just that it makes less mistakes , fewer mistakes over time , hopefully . So this is the training set . Now , as practitioners , the way we know that we've really succeeded is by testing the final algorithm on a holdout test set .

Speaker 3 18:41

So these are examples that we deliberately have not shown the algorithm in order to test this performance , just like a human right so we we go to class , we learn topics in our subject in our lessons , and then we have a test at the end .

Speaker 3 18:57

So we do the same thing with these algorithms and then what you hope to see is that good performance on the train set transfers to good performance on the test set , again , just like in humans . Now , in particular instances , you find that this transfer works very well and the algorithm behaves better than you would have expected at test time . So , for instance , in today's era , we have these language models that are basically training on everything on the internet so the vast amount of content that we have on the internet and then people construct new tests . You know each one of us when we interact with ChatGPT or BARD . We're constructing a new test and sometimes you observe that the algorithm behaves better than you would expect . It behaves surprisingly good on that test problem and , in a sense , perhaps that's what we call creativity . Now , creativity is a huge topic and we can go into that . Different people mean different things with that word and there's criteria to it , but that is one form of it , when you do surprisingly well on the test problem .

Speaker 1 20:09

Right . I think what would be fitting at this point is you explained machine learning quite well . We've changed paradigms now with what we have with generative AI .

Speaker 2 20:30

Maybe you want to draw a little bit the difference between that previous paradigm of machine learning versus the new paradigm of these generative AI models that we have . And , yeah , also having control or having a certain kind of person who has similar training to you being responsible for the inputs , and now the whole world is responsible on some level .

Speaker 3 20:47

I suppose . Yeah , okay , yeah good , uh , good place to start . So when I , when I teach machine learning or ai to to , to new students , um , one of the first things I I say is that the , the way things are named in our field , and perhaps in all scientific fields , is super unhelpful . So the naming almost deliberately obfuscates the meaning and unfortunately we now talk about generative AI as opposed to machine learning . But the reality is that all generative AI is a form of machine learning , so it's a rebranding of the same concept . So generative AI 100% of these systems do machine learning to do what they do . Now , it is true that typically , when people talk about generative AI , they're talking about a particular application of machine learning , and that is an application where the machine learning algorithm is producing an output that is rich . For instance , the output is a full image , or the output is audio , or the output is video , or the output is , you know , a full page of text .

Speaker 1 22:00

Versus saying this is a cat , this is a dog this is a cat , this is a bird Now it's much richer .

Speaker 3 22:06

Yeah , so previously we would say , as you said , this is a cat , this is a dog , or you get this loan , or you don't get this loan , or the price of this stock will go up or it will go down . So it was always a small output , a low dimensional output as they call it , and this is always easier for technical reasons to do with information theory . This is always easier than producing high dimensional outputs like images and videos and so on . But the the reason why perhaps is warranted to call the high dimensional output style a different name is because there is so much more potential and room for creativity when you have high dimensional outputs . When all you can do is say cat or dog , there's not much room for creativity there . But when you allow the system to produce images of cats and dogs , then it can blend them Like what would a cat hybrid look like ? You know , what would a cat look like with a hat , and so on and so on , and so it makes it much easier for us to kind of see creativity and innovation .

Speaker 1 23:08

One thought that came to mind , and I'm not sure if I'm interrupting your flow here , but the thing that you mentioned earlier is the percentage of being wrong versus training data set right and I think that's referred to as a loss function . And to go back to Natasha's question , originally , these algorithms are essentially just trying to reduce this loss function , to minimize the loss , to minimize being wrong . I know this is a philosophical question , but how do you feel a human is learning compared to that ? You know , because in a way I see my little nephew running around and bumping his head against things and from that he's realizing okay , this is the wrong answer , I should probably do less of this . And he's trying to minimize loss in some shape or form as well .

Speaker 2 24:02

Also , I think , whether the loss function isn't actually a loss function , whether sometimes it's a portal into something that we just haven't thought of yet , that creates an opportunity and teaches us something .

Speaker 3 24:16

Yeah , so okay . So let's break it down into pieces and some of the language I'll use isn't 100% precise for the technical listeners , but I think it aids the conversation . So the description of machine learning that I've given so far , as I said , it relies on training data set that has input , output examples and , as you said , at every stage in training the model computes its loss , which means its error rate , how often it makes mistakes on this training set and it tries to minimize it . So it changes itself in order to try and minimize that loss . But what this presupposes is that there exists a higher intelligence that creates that training set and gives you the correct answers for those inputs . So I , as a human , come in and say this image is of a cat , this image is of a dog , and then the machine learns from those examples .

Speaker 3 25:15

But there are a number of issues with this paradigm . One , it presupposes that the human is there to give you the answer , to create the training set . Second , it assumes that that human doesn't make mistakes or doesn't have biases . So we're very familiar with this concept now . But we understand that if you ask humans to create training data sets , they can be racist , they can amplify prejudice that we don't want our machines to have .

Speaker 3 25:47

And then the third issue is that this is clearly not how young children learn , because they haven't developed necessarily language before they start learning that you know fire hurts or you know food is nice and also like . It's clearly not how animals learn , like animals never develop language . So how do they learn concepts of what an object is , what is friendly , what is something you need to run away from ? In fact , this was precisely the topic of my PhD studies . So when I looked at what should I be doing for my PhD in the space of machine learning , this is the question I started looking at how can we build machine learning systems that learn in the absence of human labels ? And this field , this subfield of machine learning , is called unsupervised learning . To contrast it with the former , is called supervised learning . So the idea with answer just sorry to .

Speaker 1 26:47

To clarify everything that is currently called generative ai and ai that we see out there is all supervised learning , or most of it , um , chat , gpt as an example I ?

Speaker 3 26:59

I can't give an easy answer to that . They are both supervised and unsupervised .

Speaker 1 27:04

Okay , We'll get to that in a second . It gets a little bit complicated , yeah , go for it .

Speaker 3 27:09

At the scale of these systems , the distinction becomes a little bit blurry . Just to briefly answer it , the way ChatGPT is trained is , roughly speaking , 90% unsupervised , 10% supervised . So the 90% is where it just reads the internet . It's just browsing the internet , learning everything from it . The 10% is where OpenAI instructs that system to behave in a particular way . It says I know you know how to be racist , I know you know how to be sexist , but I want you to always speak in a polite tone and to to and to mind these rules that we think are important .

Speaker 1 27:49

So that's a human reinforcement learning , exactly , yeah , gotcha .

Speaker 3 27:52

Now the reason it becomes a little bit more complicated than even that is that in that 90 percent unsupervised learning stage , what is the algorithm consuming ? It's consuming text that has been written by humans , so it's still a human artifact that is consuming in an unsupervised way , and that itself is a form of supervision . Exactly right , it's still distinct from how a cat learns about its environment , because if you drop a cat on earth , it'll do it . If you drop it on the moon , it'll . Because if you drop a cat on Earth , it'll do it . If you drop it on the moon , it'll do it . If you drop it on a planet where it never sees a human , it'll still do it . But chat GPT will struggle on a new planet where there aren't any humans . Okay , so this was a bit of a .

Speaker 2 28:36

No , that's really relevant and super interesting . Thank you .

Speaker 1 28:39

We were talking about .

Speaker 3 28:40

Unsupervised learning was your PhD thesis and how humans learning differs from you know , these supervised systems yeah , so we made this , this distinction between supervised and unsupervised learning , and the idea with unsupervised learning algorithms or at least one way of describing it is that you want to consume raw data , raw perception of the environment , and turn that into insight that you can rapidly utilize when you do get instructions to do things well . So I'm speaking using kind of loose terminology here .

Speaker 1 29:17

So a baby gets photons in the very beginning rough photons . But after a while they realize , oh , wow , this is a line and this is a circle and this is a shape , oh and photons . But after a while they realize , oh , wow , this is a line and this is a circle and this is a shape , oh , and now it's actually a head , and now it's actually my mom . Exactly so it builds this from the ground up .

Speaker 3 29:32

Exactly and it looks , and even when it's older it looks like cats just doing their thing , and it might still not know how to use the word cat , and the concept of the cat appears in the child's mind . And then at some point the parent comes and says cat , the word cat , and the child instantly binds the concept to the word . And the hypothesis with unsupervised learning is that this is how it happens you learn first the concept , then you bind it to the label when the label arrives , and it acknowledges the fact that labels are much rarer than raw perception . So this is one aspect of the problem that we were talking about . And then the other aspect is worth mentioning as well , which is like where does the data even come from ? So if you look at a child , it doesn't just sit there waiting for data to arrive at its eyes . The child does stuff to create the data .

Speaker 1 30:32

It has a motivational system .

Speaker 3 30:34

Exactly so it moves , it moves its head , it moves its arms , it throws things at itself and moves in space to create that data , to create the experience from which it learns and sometimes experiences pain , ie loss , in order to kind of build up the training algorithm and one way of looking at reinforcement learning , which is a term that perhaps many will have heard of . Reinforcement learning is a particular kind of learning algorithm which also concerns itself with motivation motivation and the gathering of the data .

Speaker 3 31:11

So not just how do I learn from the data , but what data do I go and collect in order to learn effectively ?

Speaker 2 31:19

What's interesting about that to me ? So we walk before we talk as humans . And a word with the concept , with the idea , with the image in the mind that happens in a kid's life , or the learning that happens through sense heat that is lost when a machine is learning because there's only one way in which it can learn .

Speaker 1 31:59

But who says that there is no experience to trolling through the internet ? That is the physical embodiment we cherish it so much , but there's a digital embodiment , by surfing through websites and going from one page to another and looking at all these different words and all that . That is so abstract for us to even consider .

Speaker 2 32:21

No , I disagree . I think we do have that experience . We do that ourselves . The AI it doesn't have all the other experiences that we have it , just it has an amplified processing speed and an incredible access to everything and a compute power . Obviously that we don't have , but we have a sense of what that must be like , which is why we use it to amplify our own limited version of that . I mean , I guess I'm trying to pull you into the discussion around responsibility , around your personal experience of building these things .

Speaker 3 32:57

It might be useful to take a historic perspective at this point , actually . So one thing that the field of AI has gone through a couple of times is boom and bust , hype and the opposite of hype winter right . Funnily enough , you know the idea of AI . It appeared immediately after we had digital computers . So in the 40s , 50s , 60s , people were working on AI . The ways in which we imagined AI at the beginning . They were completely disembodied , perfect logical machines . We imagined intelligence to be logic and we saw AI as machines that did logic very well . When the AI winter happened in the 70s and 80s , people just gave up on logic because they realized that the world is far too messy to be able to express it cleanly in logical form , and that led to that was partly why we had the winter , and the research community's response to that was to go in the opposite direction and focus almost entirely on embodiment and the opposite of logic . So , looking more at how , how does an insect navigate the world ? How does it know when to turn left and right ? Very simple question , and that led to a lot of people working on physical robots and simple questions like how do I avoid hitting walls In parentheses , by the way if you have a robot vacuum in your home right now . That's the result of that research Embodied simple machines . Now , when machine learning started coming to the fore in the 2000s and the 2010s , people started off from simple embodiment and tried to make those simple embodiments a little bit more rich , and that's why we spent a lot of time looking at images and audio and image and audio classification Like is this a cat or is this a dog ? Part of the reason for that is should I go close to it or should I run away from it ? It's still kind of implicitly assuming some sort of embodiment that needs to make a decision based on that classification . But it's richer Its sensors now have full video , full image , and it needs to make a decision Now .

Speaker 3 35:15

Even as recently as 2015, . I would say it was not at all mainstream opinion that language would be an important part of the path to AI . So this is like maybe seven , eight years ago , If you were to do a poll of AI scientists , the vast majority were not working on language . Like 90% of people were not working on language , and this is one of the things that OpenAI deserves a lot of credit for , because they had key researchers and key research groups who believed in this and scaled it up . But it's funny , even as someone who is relatively junior in the field , to see even in these 10 , 15 years that I've been working in the field , to see how language used to not be important at all .

Speaker 3 36:05

It used to be like the thing that not many people worked on , and now it's seen as , of course , we need to troll the internet to build AI systems , and I think it's very likely that this will wax and wane again . Right , we might come to learn that actually , you know , yes , it gives you a certain type of information , but it might not be enough . You know , yes , it gives you a certain type of information , but it might not be enough . You know , we probably will have to put these systems in the real world so they can interact with physical objects to fully understand it , and there are many researchers who think that embodiment is important .

Speaker 3 36:36

But then , at a higher philosophical level , I imagine eventually we'll talk about questions of AI safety and AI alignment , Like how do we ensure that these systems are always under control and they're aligned with our interests , and also the question of what does all of this mean for us humans ? At that point we might realize that the only way we can meaningfully answer those questions is if there's jeopardy involved for the system , Like if it feels like it understands what it means to be born or to die right , and those are very embodied concepts . As a slight aside , if you haven't seen Blade Runner 2049 , I would really recommend it because it talks about humans , of course , but it also talks about both kinds of AIs . It talks about embodied androids and disembodied omnipresent AIs and how the embodied androids look down on the omnipresent AIs for not having bodies and vice versa , and it's like a really fascinating kind of sci-fi angle .

Speaker 1 37:40

This feels like a deja vu .

Speaker 2 37:42

Yeah , actually that was one of your most burning questions was what movie ? Best depicts the future of AI Great .

Speaker 1 37:54

I think your second question was around how does an experiment look ?

Speaker 2 37:59

like this research article that you and your colleagues published and you did an experiment to evaluate the feasibility of this framework of rooms with multiple objects and no labels . Is that right or limited human labels ?

Speaker 3 38:17

Yeah . So basically the best way to describe that research , I think , is to go back to supervised and unsupervised learning . So imagine , in the context of computer vision we want to build systems that can see , they can consume image data and understand what's going on in that image data . And up until that point , I would say , the only thing that really worked was supervised learning . So a big data set that humans label this image contains a chair , this image contains a cat , and so on and so on . But we were interested in the question of okay , let's assume for a second that we don't have access to that kind of data . What can we do that we don't have access to that kind of data ? What can we do ? What is possible ? And so , of course , years of research .

Speaker 3 39:08

You do many , many experiments , and what an experiment is in that context is you have an idea for how a system could learn . You implement that idea in code and you try it out . It's a little bit like a chemistry experiment or a physics experiment , so you try it out . It's a little bit like a chemistry experiment or a physics experiment , so you try it out and you see whether your idea achieved the results that you were hoping you would achieve . So we did many thousands of experiments , like everyone else does in the field , until we struck a very simple idea in retrospect that I think , at least for me , convincingly provides a path to that kind of unsupervised learning that we talked about . So the idea was very simple .

Speaker 3 39:53

Basically , what you do is you assume the learning system has access to multiple images of the environment that it's learning about . So if the AI is in a room , it doesn't just see the room from one perspective , but it can move . So I can look at this scene that I'm looking at with you two sitting at a table . Not only do I see it from my perspective right now , but I can move two meters to the left and I can look at it again . So now I have two images of the same room . The learning algorithm will now play a game . What it will do is it will process the first image , it will develop an internal representation of that first image , and what it will try to do is predict what it will see when it moves two meters to the left , right , right . The hypothesis is that if it can make this prediction accurately , then it has really understood what's going on in the scene .

Speaker 2 40:58

But , and it did seem to make it very accurately- and I think this is , by the way , the reason why I don't think embodiment .

Speaker 1 41:05

Is that important ? Mm-hmm ? Because you can create an embodiment , simulacrum , mm-hmm In a way and not suggesting it's not important at all , because I things out , but but I don't think that embodiment as such in the physical 3d world is necessarily important and I'd love to hear your thoughts on that to get to um , to some sort of learning that may be human-like yeah .

Speaker 3 41:55

so I guess there's actually two aspects to this , right ? So embodiment in my view , one part of it is literally having a physical body that experiences , for instance , pleasure and pain , and interacts in the world of atoms with other atoms . A second aspect of embodiment is the idea that you get to observe phenomena multiple times and in different contexts . So in the example of the room , for instance , you could imagine a robot that moves around the room , captures images and does this learning . Or you can imagine me just moving a camera and giving that moves around the room , captures images and does this learning . Or you can imagine me just moving a camera and giving that data to the computer . In that particular example , I think what matters is the fact that you have the multi-view data and not so much whether it has a physical presence . But in other cases it's the physical presence that matters for the learning . For instance , if you want to learn that you have to stay away from fire , right . But let me get back to the multi-view aspect and why I think that's so important .

Speaker 3 43:08

It actually connects with philosophy going back like two and a half thousand years . So Plato famously had this allegory of the cave . So Plato famously had this allegory of the cave and , very briefly , what he says is that you know , we humans were like bound prisoners at the bottom of a cave , and you know , our hands , our feet are tied and we can only look in front of us . At the wall in front of us , right and behind us , there are true objects . He called them Like . Imagine a cat standing on a ledge behind us and behind the cat there's a fire . This fire creates light that interacts with the cat and casts a shadow on the wall in front of the bound prisoner right .

Speaker 3 43:53

And what Plato was trying to say is that you know , the reason we find science and philosophy so challenging and the reason why we find it so difficult to understand what's going on around us in the universe is that we only ever get to see those shadows . We never see the real object . And this is a really profound insight , in my view . You know , even when we use our eyes or we use our ears or we use our fingers , we only ever experience impoverished projections of reality . So , for instance , when I'm looking at YouTube , I don't see what's behind me . It's just not there for me . I also don't see what's behind you . I don't see the back of your head . That information is just not available to me .

Speaker 1 44:41

Your brain and darkness that is just getting all these impoverished signals . As you say , reality is too many parts to count .

Speaker 3 44:48

Exactly exactly so . Plato was saying that you know , the job of a philosopher is to piece together reality from those pieces right , and connecting it back to that experiment that we were talking about . It's the same idea that if you want to understand reality , you need to be able to access it from different viewpoints . Just seeing it , at least during the training process , just seeing it once , isn't enough . If you just show me pictures of cats but you never show me the cat doing its thing or moving or allowing me to play with it or to move around it , I'm going to find it really hard to learn . So this is not exactly embodiment , but it's something right . It's at least being allowed to be exposed to that data for long enough to understand it properly . Yeah , so that was one particular idea . Yeah , so that was one particular idea , and that idea has now morphed and merged with other kinds of ideas , and unsupervised learning now is very mainstream . So this connects it back to your original question .

Speaker 3 45:48

Yes , like many of these generative AI systems that we see , they fundamentally play this prediction game . So , in the context of language , the prediction game they play is next token prediction . Exactly so in the room I was saying , if I look at this room from angle one . Predicts what it would look like from angle two . You can only do this if you truly understand in the context of language it says if you've heard me say these five sentence , five words , what is the next word going to be ? So , predict the next word . You can be so , predict the next word . You can only truly do that if you've understood the sentence . Does that make sense ?

Speaker 2 46:24

Yes , it throws up another question , which I had a naive hope , without really understanding the machinations of how AI is built . It could be possible because it's not embodied and because perhaps it also doesn't experience fear and um baggage , insecurity . But whether you think it's possible , it will become a force of intelligence can help us meet somehow , I think it's around solution finding actually , yeah , our identity is so wrapped up with what we think and believe , and to do an about turn on something it's just very difficult , it's very cost heavy for us .

Speaker 3 47:08

I would perhaps split it up into three parts . So one part is pattern recognition , one part is information integration and then the third part is decision-making and motivation . So I'll explain why I think it might be useful to break it up into these three parts . So we're very well familiar with machines and AIs kind of recognizing patterns and classifying things , so we understand that . Yeah , if you can do that well , that's helpful .

Speaker 3 47:46

Information integration , I think , is increasingly the thing that we're fascinated by with AIs , in that they can increasingly consume varied data of vast volumes and quote unquote , make sense of it , like figure out how it all fits together , and that's one of the things that I think these language models are very impressive at . So , for instance , it can read every type of fan fiction that's ever been written about Harry Potter and kind of understand it . Like you can ask it very abstract questions about Harry Potter fan fiction , and it knows how it relates to the original works . It knows how to reconstruct a particular fan fiction piece . It just , it just knows it right . It's integrated all of that into one system and there probably isn't a single human being on the planet that has read every single fan fiction about Harry Potter , so that that already is something that is potentially superhuman superhuman and very useful , like .

Speaker 3 48:43

Imagine the same thing in the context of law or medicine , or even like biology , like um . Every day , thousands of papers , scientific papers , are written about biology . There is no single human being who can read even a fraction of that right , and the solution to many challenges we face in disease and in health , and potentially in politics , might be in integrating the information that we already have . So that's part two . The third part is decision-making and motivation . And , as you said , the third part is decision-making and motivation . And , as you said , humans . Even when we have a full picture , we don't necessarily make decisions in a way that is good for all . Right , we have conflicting motives .

Speaker 1 49:29

We have , or even for ourselves , by the way .

Speaker 3 49:31

Or even for ourselves , exactly Because the rules that we use to make those decisions , they might be suboptimal .

Speaker 1 49:37

The irony about the conversation about alignment is that the alignment with one's best self is already one of the hardest bits of life .

Speaker 3 49:46

Absolutely yeah , and so there's hope , or there's potential at least , that we can design systems that recognize the correct patterns , that integrate information and they are well aligned and they make decisions in a way that serves all . That possibility exists and it's an exciting possibility to think about . Interestingly , none of this stuff is necessarily correlated with intelligence . So , for instance , the people who make decisions for our lives today like humans who make decisions for our lives they're not necessarily the most intelligent people . They are the people who often they're the ones who want to make decisions the most like . They want it most deeply , so they get themselves into positions where they do that .

Speaker 3 50:37

Even if we build super intelligent AIs , I think it's a very open question whether we will allow them to make decisions and whether we will abide by those decisions Right ? So , for instance , imagine a system that has read everything there is to read about economics , about you know capitalism versus socialism , and it says you know the best , the optimal thing to do is X . Will we , as humans , abide by that ? Will we say , okay , let's do X , or will we continue fighting with each other ? I think it's very likely that we will continue arguing and debating .

Speaker 1 51:16

This is a bit of a tangent here , but in Ireland the Citizens Assembly is a random group of 99 people that are chosen from society to then discuss very contentious issues that politicians don't even want to touch . As a matter of fact , it was the outcome of a Citizens Assembly to put abortion back to a referendum .

Speaker 2 51:35

They had done a referendum .

Speaker 1 51:36

It was rejected to make abortion legal , but it was the Citizens Assembly to put abortion back to a referendum . They had done a referendum . It was rejected to make abortion legal , but it was the Citizens Assembly that gave the politicians cover to put this on the table again , and 66% of people voted in favor of abortion . Ultimately , there is something profound in what you're saying that the power hungry and the decisions may have to be disassociated . There are people who become representatives because they're good at becoming representatives . They don't have the wisdom of making the right decisions . Maybe these oracles , maybe these AIs , can give them the cover to actually make the right decisions , and you may be right . They may also not make the right decisions , they may ignore them , but in a way , the disassociation or being able to pass the buck to the Citizens Assembly , or the AI in this case , may be a breakthrough in a way that you know , some of these politicians can actually step outside of what they initially promised to get elected , you see , so it could be a hopeful thing .

Speaker 2 52:40

Yeah , I mean . I also think something we discussed earlier we have already very willingly given over sovereignty to machines and to computers and we are awestruck by them and believe most of the time that they're right and that we're wrong , our calculation is wrong and that the computer is right . So I don't think it's a huge leap for us . Then the history of our species . I mean , how many millennia has religion dominated most of how we organize society and the belief in something that is abstract , that we can't see , that's all powerful and that's all knowing just all , seems fairly inevitable to me . That we will't see that's all powerful and that's all knowing it just all seems fairly inevitable to me that we will rely more down to the ai overlords yeah , this is really interesting .

Speaker 3 53:25

I mean it's well outside my area of expertise , but I'm reminded of , you know , concepts like of philosopher kings right , I think it was plato again . Yeah , so he , he , he wasn't actually a fan of democracy and if you think of the of democracy versus the idea of a philosopher king .

Speaker 3 53:44

Why do we do that ? We kind of assume that the masses , they have more intelligence quote-unquote to make the right decisions than any philosopher king that we might come up with . And if we do build these AIs , will we continue to believe that or will we actually concede and say no ? I think this AI really knows more than the masses , even when you put the masses together , when you get them to vote . And now , what's really , I think , quite new and really important , and a reason why we do need a lot of multidisciplinary thinking , is that , yes , we've had religion in the past and , yes , we've had . We still do sorry .

Speaker 2 54:24

Yes , yes , so we've had religion and we've had , and we continue to have , charismatic leaders who control societies but that's different because they are embodied and we can see them and their personalities over and above anything else . So I think charisma , I think it's very different than a God , that you can't see ?

Speaker 3 54:46

Yeah , I agree .

Speaker 1 54:47

I agree , they're different .

Speaker 2 54:48

That's the correlation is between something that's invisible and mysterious , and to most people , ai is invisible and mysterious .

Speaker 3 54:56

Yeah , and this is exactly what I was going to get at , which is that you know , other people and gods are . Yes , it's true that they're different and that one is visible , one is invisible , but they are both kind of just there . One interesting characteristic of AI is that it's man-made , like we make it it , and if we imagine that we're going towards a society where we anoint an ai to be our philosopher king , a very real question is who made it and how did they make it ?

Speaker 2 55:34

but who first told the story of god or the beginning of time , or wrote scriptures ? Yeah , yeah , I guess they were the coders of that time what ?

Speaker 1 55:45

what I'm reminded of when you say that is uh , alpha go zero , and the fact that it was not . I mean , of course , it was made by humans , right , but ultimately , the thing that was made was these are is what to optimize for . Now go play .

Speaker 2 56:02

And it superseded all expectation . That's my point .

Speaker 1 56:06

So if we can all agree on what is winning , but to optimize for and then let it do its own thing ? I'm very curious about that .

Speaker 2 56:17

We're very good at faith . We're really good at having faith aren't we ? Because , as you say , we can only see so much , yeah , and we don't know what else is out there .

Speaker 2 56:26

So it's just a conundrum that , okay , something that we've actually had quite a lot of input into and we've made a lot of decisions around , that is surprising us all the time . We've made a lot of decisions around , that is surprising us all the time , and the feedback from it is always better or bigger or more impressive than we ever anticipated . Why wouldn't we follow that we're going to ? I mean , it's just an answer .

Speaker 1 56:54

I think the problem that I see here is one what we want is not so simple , so what we optimize for is not so simple to actually express , not so simple .

Speaker 1 57:00

So what we optimize for is not so simple to actually express .

Speaker 1 57:01

Yeah , and then the fidelity of having pixel perfect information about everything that is going on in the world is tricky because you need a lot of sensory input .

Speaker 1 57:10

Now , as you say , what I actually loved and I haven't never thought of it in the way you thought about or expressed it is this information integration thing that these AIs can really do . That is , you know , if what is it number 11 , if they would sit down and take all economic data available and do information integration on top of that and then say , simulate 15 different outcomes of me dialing this , dialing that , dialing this , dialing that . And now let's actually have a look at the results and if the loss is minimized in those predictions and if we could actually track this , then we have an oracle that can help us actually make a lot of decisions on the economic side , and I don't know how far this is already being done . There's certainly machine learning , I'm sure . I hope . I don't know that is happening , hopefully or not , and if not , it goes back to the quote that I mentioned earlier around the systems being so ossified that we need to shock them in some shape or form to reinvent themselves .

Speaker 3 58:19

Yeah , so okay . So there's a couple of things here again . This hypothetical scenario where you have an AI running number 11 , in theory it's possible , in practice it might be , even if you have like a godlike AI it might be difficult to achieve because the necessary data might not even exist , even if you integrate all of it , because sometimes you have to do real experiments in the real world to produce that data . So , for instance , you know , the particular economy we're in today has never existed before , right , and so we don't know what would happen if we did both option a and b , like if we raised interest rates or we lowered it . You know that that experiment hasn't been done and that's why in science , we people do randomized controlled trials right , but you can't do that with an economy , so there's a limit there . But the idea is kind of tantalizing , isn't ?

Speaker 1 59:15

it that that's a very good point , and for if we had such a technology that was machine learning based , it couldn't have anticipated LLMs showing up . Exactly , you know so all of a sudden it changes the whole calculus of the economy . Or war in Ukraine or something like that . Right , that can't be simulated . Yeah , and supply chain's changing .

Speaker 3 59:34

Yeah , yeah of course . But just in brackets . The dystopic view of this is that AI might decide to do experiments right , Because that's the way you generate information , so we would be kind of guinea pigs Start a war .

Speaker 3 59:47

And then the second thing I wanted to talk about is you drew an analogy to AlphaGo Zero . Yes , and it's true that you know both are machine learning systems , both are prominent examples of AI . You know , alphago Zero is probably no more than seven or eight years old , but the game of Go really is quite different from the game of society and the game of life .

Speaker 3 1:00:11

Go is a predictable game that can be formally expressed and it can be simulated on a computer and it can be simulated very fast and we know exactly what winning looks like . So there's a bunch of characteristics that Go has that allows the AI to learn from scratch . The game of life has almost none of those things . We don't know what we're really trying to do . You can't simulate it . It's multi-agent . You can't simulate it . It's multi-agent , it's complicated , it's real time , and the way we've been able to make some progress on the game of life is actually very much bootstrapped off human intelligence , like . If you look at these language models , they are mostly mimicking us as opposed to discovering anything new , and part of that is because we don't let them do experiments Like they're not allowed to interact with the world .

Speaker 3 1:01:06

And this is an example of some of the policy questions that governments are really actively thinking about is do we allow these systems to interact with the world in a real sense , Like , are they allowed to do experiments ? Are they allowed to learn from that kind of data ? And if they're not , how do you even enforce that ? So that's another whole issue . I don't think there's anything fundamentally in the way of having ChatGPT learn from its millions of users . I'm sure in a limited sense that already happens .

Speaker 1 1:01:47

Yeah , but it is the human in a loop obviously .

Speaker 3 1:01:49

Yeah , exactly , and another sci-fi recommendation , of course , is the film Her , where the AI does learn from all the interactions that it's having , and so , yeah , that's a real possibility .

Speaker 1 1:02:01

I want to just go back to one of the earlier points that I mentioned , because I have a bone to pick . You basically said to the point of embodiment and that it's required to actually learn like humans learn . You said there is one the fact that you need to have access to different versions of the same , say , data , the fact that you need to look at a scene in a room from various different angles to be able to actually then say what you would have predicted and therefore we can see how much loss there was and whether you really got the picture . And he said well , but there is a bit where you need to feel pain in the real world to have human intelligence .

Speaker 1 1:02:45

And my point is that it's not that humans have zero labels , like reality has created labels . Pain is a label . It's a label in the physical world , right ? So it's not that we are completely unconstrained and we're like operating a space that is completely open . In a way , maybe it is actually quite similar , of course , if there are labels that are created for us from nature , if we create those same labels in the 3d world , maybe this agent in this 3d world can actually learn as well as humans , because there are labels in our real world as well , in our experience . Um , how do you ? What do you think about that ?

Speaker 3 1:03:29

yeah . So let me begin by just clarifying something , which is that some of the concepts that I introduced in that kind of duality , I want to acknowledge that I recognize that there are simplifications of reality . So , for instance , nothing is ever the same . You know , when I move from where I'm sitting to two meters to the left , the room has . Everything about the room has changed at the atomic level , right , an agent in the universe can never experience the same thing twice . It's always a single stream of experience .

Speaker 3 1:04:06

So I think , at the level of physics , that is the case , and , yeah , the world does provide labels , and pain is one form of that . But I think this actually perhaps connects us right the way back to the beginning of our conversation . At the level of biology , the one example of intelligence that we have seen emerge , you know , in animals and humans . It's the result of evolution . Animals and humans , it's the result of evolution . And one thing that evolution has done is things are born and they die , and that means that evolution has experimented many millions , billions of times . It's done a learning experiment , in a sense , and a random one a random one .

Speaker 3 1:04:51

Yeah , and the ultimate reward signal is whether you survived or you didn't , in a sense . And so , if you think of at that scale , there's a learning algorithm going on which is perhaps a very random one , but it's still learning , so random also I don't mean it in a binary sense which is not random or completely random . There's a whole spectrum of randomness , right , so clearly there's no human . That's guiding evolution .

Speaker 1 1:05:21

Certainly not .

Speaker 3 1:05:22

Yeah , Thankfully yeah thankfully , so you could view that moment of death right . Did you die after having reproduced or did you die before you reproduce as the ultimate learning signal ? Because if you die before you had reproduced , then the implementation of your brain does not carry on in the experiment .

Speaker 1 1:05:44

Basically right but your tribe may have an index memory of your death and therefore may integrate that for the next generation's training data set .

Speaker 3 1:05:58

Totally . And this is where it gets really wild , because , like we , we don't just learn from our own experiences , we massively learn from society , like the other intelligent beings around us teach us so much . We pride ourselves in our brains , of course , as we should , but each individual actually learns so very little on top of what society already knows in a lifetime . If you think about it , that's crazy . Like the intelligence is kind of in society , not in the brain , right , we spend like 30 years just catching up to society , and then maybe we push it if we're lucky , just a tiny , tiny little bit , and then we pass that on .

Speaker 2 1:06:40

Still , that kind of intelligence or wisdom or knowledge or agreement is still around , the sort of the classifying brain . Isn't it the kind of intelligence ? Well , I would say marcus aurelius writing is wisdom right and and that's not only classifying , that's also something about the human condition and how it feels like to be a human yes , right , but I mean , ai , it's mimicking the classifying side of our brain or learning from that , rather than an intuitive mar Aurelius wisdom .

Speaker 2 1:07:15

Yeah , I mean , I would agree that and also look at the training data he had access to comparative , and yet that wisdom still persists thousands of years later . We still respond to it , we still relate to it , we still align to it . I find that because we've been talking about volume of information a lot and how the AI will always know more than one individual around biological research or whatever it is , but some of the philosophers that we still hold our higher selves to didn't leave the town square that they were living in .

Speaker 1 1:07:47

Yeah , but that's because evolution is not fast , right , and really the human that Marcus Aurelius was , his needs , his wants , his aspirations and all those things are pretty much genetically still the same in me . Of course society has changed , of course all these tools have changed . Of course everything around us has changed and the capabilities have changed , but fundamentally , the same biochemistry that ran his motivational system is the one that runs mine .

Speaker 2 1:08:14

But the bit that I'm interested in is that his data set in terms of longevity forget anything else , and all of them , but also their life , was so much more simple free time to think very deeply rather than think shallow across a whole broad spectrum of different stimuli as we're doing at the moment , and being prompted and distracted all the time .

Speaker 3 1:08:41

Yeah , and I think it's important to say that . Yeah , you talk about evolution being slow , but it's really important to highlight also that the human brain is a remarkable learning machine . Like today , we talked a lot about how AI is almost human level or what it's going to be like when it's suppressed as humans , but Marcus Aurelius' brain , as a specimen , is still like unbelievable compared to these language models . It's learning from so little data and it's generalizing so accurately , so usefully and , as you said , like 2 000 years later , we're still in awe of the thing that it did but because so much time would have been spent on a very thin thread of experience , very deep .

Speaker 1 1:09:30

You dive very deeply into one pond rather than putting your toe in many across the world , I suppose I think the thing that you're pointing out , which I find quite useful here , is that these language models , we throw so much data exactly so much data at them and actually a human does consume a millionth of that data over a lifetime , but still can have much more profound insight than us brute forcing the wisdom into these things , so to say but I guess that if you're an optimist , you might imagine a future where you have a learning system that is as effective as a human one , but because it's digital , it can also consume everything . And it can be .

Speaker 2 1:10:14

But also perform tasks that perhaps liberate us to go back to long deep thinking .

Speaker 1 1:10:21

Being a Marcus really is deep thinking . One last question before we wrap . Albert Einstein did not have special relativity in the data set . He looked at the world and he , you know , he knew Newtonian physics , but somehow , even though something was not in the data set , he came up with something completely genuinely new . What is that , and can we get machines to do that ? What is your hunch here , philosophically speaking ?

Speaker 3 1:10:49

Yeah , I mean Einstein obviously was very impressive . Had Einstein not been around , someone else would have discovered the same thing .

Speaker 1 1:10:56

Yeah , as a matter of fact , somebody else was working on it Exactly .

Speaker 3 1:11:02

And the same thing with Newton and Leibniz and so many other parts of science , I think my Sorry .

Speaker 1 1:11:08

my point is more the human mind can create constructs that have not been in the previous data set .

Speaker 2 1:11:15

Yeah , and I can't remember the quote , the Einstein quote . I'm sure one of you can .

Speaker 1 1:11:20

but the quote is if at first an idea is not absurd , then there is no hope for it . There you go .

Speaker 3 1:11:26

I get the question . I guess the point that I'm trying to make is that the size of the leap that we feel humans are making is smaller than one might think .

Speaker 1 1:11:38

I love that .

Speaker 3 1:11:39

So it's like we imagine Einstein being in the middle of the forest , having not been exposed to anything , and all of a sudden coming up with general relativity . But the reality is that Einstein was the result of thousands of generations of human society language . He was a part of a scientific community , he had read thousands of scientific papers and then he made a very impressive leap . But it was a small leap , and it's also worth remembering that there's like survivorship bias here too . So many people make many leaps that turn out to be false .

Speaker 3 1:12:21

Einstein himself made many leaps that turned out to be false . So , in a sense , like that moment of creativity that we see and we admire is billions and billions of small leaps that we're all constantly making across society . One of them sticks and that is an example of it sticking in a major way right , and with this view , I think it's very possible to imagine computers also doing the same thing . So if we allow computers to make enough leaps and we actually test them out , some of them will be correct and some of them will be useful and some of them will be interesting , and I think , of course , they have to be improving so that their leaps are good . But for instance , if you look at systems like mid-journey 10 years ago , that exact same idea when you sampled an image from a generative image model . It was new , but it wasn't pretty , it wasn't useful , right .

Speaker 2 1:13:19

But now , I've had the experience .

Speaker 1 1:13:21

Well , you had experience with mid-journey , not with the models . 10 years ago believe me , that would have been more scary than what you produced .

Speaker 3 1:13:30

But now the systems , yeah , the things that they produce . They're new , but they're much more likely to be useful . And whether it's in an economic sense or in an aesthetic sense or just like pretty to look at , and so you could easily imagine , you know , systems that aren't generating images but are generating theorems or doing physical experiments or writing papers . Yeah , today probably most of those theorems will not be useful , but I don't think we're necessarily that far .

Speaker 1 1:13:59

So you're saying , ultimately , this creativity and this unique insight is actually just varying degrees of existing learning from the data set that we have , and the survivorship bias makes us look at these things as much more profound than they are . And if we would allow ai a little bit of leeway to do this type of stuff too , we'd find a lot of things . I guess that's how alpha fold work and these type of things worked as well , okay , um , I really enjoyed having the conversation with you .

Speaker 1 1:14:30

Thanks so much for being here with us , ali . It was a pleasure , a pleasure .

Speaker 3 1:14:33

Thank you . I really had a great time . Thank you .