Hi everyone! You’re watching the ToTheMoon channel. The tech world of AI from Silicon Valley, all over the world. Today we have an episode together with Max. It has finally happened, and Max and I have a lot of discussions. And in one of our latest discussions we went really deep: within a discussion of the artificial intelligence market we started taking things apart, we started discussing terms, because Max noticed very precisely that we all use terms, and between ourselves we used terms, and we were talking about different things.
Remember, we were even discussing some topic there, and it was as if we had a contradiction. And then we started explaining to each other what we mean by this or that term. That took quite a lot of time, by the way. And we came together and said: “Ah, listen, in that situation everything is kind of clear, we’re basically on the same wave overall, right?” Even though at first it really looked like a certain contradiction. And many of you, and whoever watches ToTheMoon in general, whoever deals with artificial intelligence in today’s reality and technology and the whole world in general, constantly run into misunderstandings with other people at work, within their profession, at home, wherever.
And so today we have an episode meant precisely to try, within the sphere of artificial intelligence, to take apart a few more terms, so that you for your part, first of all, keep doing that, I guess, and keep being, keep being sensible in terms of when someone uses words like AGI, ASI, well, the technological singularity and artificial intelligence in general. And, Max, right? We were saying today that we should start with the simplest thing, seemingly the simplest, with artificial intelligence. How would you suggest we structure the conversation today?
Yes, yes, yes. That’s an absolutely, absolutely great idea, in my view, right? Because you can’t argue until you’ve agreed on the concepts, really. Although a friend of mine once puzzled me a little when he said that the imprecision of human language is not a problem, it’s actually the solution to a problem. That human language actually has to be imprecise. And he suggested we agree on at least one thing precisely, right? So we sat down and tried to agree on some technical thing.
And after maybe five or six minutes I realized he was right, that to designate something absolutely precisely, to agree on something absolutely precisely using human language, is simply impossible. I mean we, we couldn’t do anything in our lives, couldn’t achieve anything, if we tried to use absolutely precise definitions every time. I mean I’d need to say an infinite number of words to define something precisely, roughly speaking, right? And so human language is imprecise by necessity.
But unfortunately, as a result of this, this kind of feature, right, not a bug but a feature of human language, a lot of problems arise in communication, when people really talk, as it seems to them, about one thing, right, but actually mean completely different concepts by one and the same phrase. So. I had a teacher at university whose favorite phrase was: “Let’s get on the rails of formalism.” When you started saying something, right, something you didn’t quite understand, he would, to rein you in a little and slow you down, suggest getting on, on the rails of formalism. What did that mean?
Let’s agree on clear logical steps, where at each step we understand exactly what the previous phrase or word meant. So. That’s exactly how mathematicians act. Mathematicians have, have the concept of formal mathematics, right? It’s a field of mathematics that deals with building logical statements that are formalized to infinity, roughly speaking, that are formalized completely, that mean exactly what they mean. It’s built starting from the simplest things, like a logical yes or no. Yes is, is a completely clear concept. Yes is… there are no different readings there. Yes is always yes.
Truth is always truth, right? A falsehood, or no, is always no. So. But unfortunately, when talking with ordinary people, or even when engineers or physicists talk, unfortunately we can’t afford that. We have to communicate using imprecise language. And so every time you start arguing with a serious crowd, I at least suggest getting onto the rails of formalism, or agreeing on the concepts, on what exactly we’re talking about.
So. And artificial intelligence is probably a very good entry, entry point for this conversation, right? So, Sasha, what do you think artificial intelligence is?
I’ve always said that for me artificial intelligence is what’s shown in those films. Usually when, you know, the Earth has been taken over, all the people killed, and there’s some superconsciousness or superworld, or superuniverse, right, some kind of superspace that controls something. It’s not just some brain, it’s some structure of worlds among themselves, a structure of spaces. And it means some independent, separate construct, right, a huge, independent, separate construct. So for me, once again, in my perception it has always been like that.
Of course, because of the last three or four years, and not even because of the last three or four years, probably because of the last fifteen years of the evolution of the tech world and the emergence of the use of artificial intelligence next to machine learning, with the terms, at least of the tech world, it started to blur a bit. But basically, I guess, it stayed with me that way somewhere. I’ll be honest, somewhere in me it has stayed that it’s some outside reality altogether, and not, not to mention… and they’re just, I don’t know, a computer or an algorithm, or some small program, or some robot.
Maybe I put it very broadly, because then you can debate what the Universe is, and what worlds are, and what consciousness is, and so on. So for me, to put it quickly, quickly and briefly, it’s not, not at all the story we’re discussing on the market today. And for me, probably, as a person who after all, well, I’m not five years old, right? If I had been born in the last three years, okay. But I wasn’t born at five, after all, I’ve been around in the world. In terms of different kinds of knowledge, I would actually be glad if, with the arrival of current technologies, people had used some basically different terms that didn’t overlap with the concept of artificial intelligence.
Because it’s precisely the concept of artificial intelligence that brought in this story where many say it’s just some little program that generates the next symbol. And in parallel the story that it’s something that will take over humanity, right? These endless arguments: marketing, business, financial and so on, right? In order to, I don’t know, push up a stock or write the right article. So, so that’s it.
Listen, here I even, I don’t know, we could go in several directions, but let’s take the simplest one. Your definition is both too broad and too narrow at the same time. So let’s try to both widen it and narrow it at the same time. Well, naturally.
Yes.
Yes. Let’s talk about what intelligence is in the first place. And that’s actually a very, very interesting question. I’ve never really heard a good answer to it in my life. There is some approximation of a good answer, right? But philosophers have actually been arguing about this for a very, very long time. And lately they’ve been arguing about it much more lively and publicly. And that’s why you can talk about it more and more, I think. I actually prepared a little, I have a few cheat sheets for this episode, and I even looked up what the Merriam-Webster dictionary, which they use here in America, the standard dictionary for defining things, says.
What is intelligence according to that dictionary? And Merriam-Webster defines intelligence as the ability to learn, understand and deal with new or difficult situations. And this is a very good definition, in my view, it focuses very clearly on what, what is actually new that we see in these artificial intelligence systems compared with the systems we saw before, right? That’s why, for example, a calculator is not an artificial intelligence system. A calculator does perform a fairly complex intellectual task that people have to be taught and so on.
And yet a calculator is not an artificial intelligence system, because it can’t learn new tasks. It’s, it’s, it’s a static system that is, well, programmed to do something, and teaching it anything new is very hard. Why, for instance, don’t we, by and large, consider animals intelligent beings after all? If we go back to that definition, we’ll again see the ability to learn, not to be trained. I mean, it’s not the ability to be teachable, but the ability to learn by oneself.
That’s the capacity for intelligence, right? Animals themselves are fairly weak at learning, right? I mean the ability is there, but it’s fairly low even compared with a small child, right? And humans, in terms of learning, out of all the creatures we know, are still somehow quite seriously ahead of everyone else. Computers have been able to do a great many complex tasks for a very long time. In some tasks they overtook humans very long ago. But precisely the ability to learn, to learn by itself, that ability is fairly new. So.
And so, if we evaluate artificial intelligence systems, then first of all, in my view, we need to evaluate their capacity for self-learning. And this is the key, key thing that, I think, people simply lose in many conversations. They look at a system’s abilities, right? And say: “Oh, it can do so much.” That, that’s not so important. What matters is what the system can learn to do out of what it couldn’t do. Can it look at a new task and solve it? By the way, this even has, this general problem of solving some task on the first try even has a special name, it’s called zero-shot learning, when you look at a new task and solve it on the first try, without even training. I mean a great many people can do it. They can take a look.
You can even show a child something, and the child… I mean show the task, not the solution of the task, but the task. And the child will figure out how to solve it, right?
Or will just do it right away for some reason.
Or will just do it right away, yes. Because it’s kind of intuitive. Intuition is also a very interesting question. But in my view, artificial intelligence means systems that, above all, can learn on their own. And this capability is very, very new. So, what do you think about a definition like that?
Look, there’s a very interesting aspect here, because if we go into, I guess, a materialist description of the concept of intelligence, a scientific description of it, I really like how clearly you defined it, clearly described it. I’d say: “Yes, I completely agree with that.” At the same time these are absolutely, fundamentally different things. Compared to what I said before, right? Because I have a worldview of my own, a vision of the world, and I don’t perceive the world as a material substance, or a person as solely a set of physical matter, right?
And so, and so, this story is like the story of what generates thoughts? I mean if we move to the question of what generates thoughts, if we stop at the idea that thoughts are generated solely by the brain, then such a term, well, the term suits me. If I go higher, that some soul-and-spirit construct takes part in generating thoughts, well, then of course we move into a completely different zone, a zone, zone of description. But I think that in our current conversation I’ll fully support the description you gave.
But here I’d like to add something right away, because when we say there is something capable of learning, or, what’s more, you made a great point, you added the comment that there is still an incredible difference when a system is able to do a task by itself. And if we talk about artificial intelligence at its base, then of course I see it as a system that is able to do a task by itself right away without knowing how to do it. I mean, for that you don’t need to study or be trained. And even this story in the world that says you have to have the skill of learning, you have to keep learning something new endlessly, keep getting this knowledge in some unclear amount.
That’s a big question, because there are a huge number of things I do in my life. I do them right away, without any experience, any training, nothing at all, I just do them right away for some reason. What’s more, I always give everyone an example. Everyone always says: “Only seniors will remain in the world.” I say: “Well, wait. Every president in the world who gets appointed is usually always a kind of junior, right?” Because people haven’t, haven’t been presidents, right? And there’s a lot of that in the world, right?
People haven’t been husbands, wives, people haven’t been dads, or children and so on. They haven’t. That’s it. We, we live in a world of these first basic flashes and actions. But here I want to emphasize what you said. When we talk about the ability to learn and achieve some result, the question always comes up: what to learn and where to achieve results, right? And that’s why they added that there is weak intelligence, high intelligence, strong intelligence. They start, you know, attaching gadgets to it, which have a boundless, probably boundless, endless, you know, kind of field.
Because a calculator really, it seems, can’t learn, it seems it can’t learn at its base, right? Even though, as you say, it performs an intellectual task. That’s important. Yes, it performs an intellectual task. And that’s why, by describing artificial intelligence so very broadly from the start, right, I tried at least, well, to show that by it I mean something very serious. I mean something that can solve not just tasks. You know, there’s this topic, everyone constantly says that current artificial intelligence now has to surpass humans.
And they kind of strive to build AGI, ASI and so on. I always say: “Guys, how can a human make something capable of surpassing that very human?” Because if a human makes something capable of surpassing him, then that something is basically already stronger than that human. He can’t control it, he can’t even be around it. It would then have to make itself. Or else a human makes part of something, a human is part of something. For example, he creates a system that will remember a billion times more than a human. Does that already exist? It does. Or he creates a system that can talk more than a human, a million times more.
Of course it can. Or a human will be smarter than any human by IQ. That’s obvious. All of today’s models are smarter than any human. No need to even ask. But the question…
Listen, IQ is a very specific thing in general.
Well, that’s in quotation marks, a hundred percent, right? But the question you touched on here, which I think is fundamental to describing, to the modern conversation about any of these terms they stuff in, AGI, ASI, now something else… they’ll invent more and more of everything. It’s still a story… or, say, someone recently said that in twelve months there will be a system that can watch a person for a week and then go on doing all of his actions. And this story, this system, it learned to do that. And then it knows how to go on living as that person and create what he hasn’t done.
And it learned, right, to do merely some similar actions of his. Well, like the buttons he presses. And is that intelligence? Well, judging by your term, by what you’re saying, it’s not intelligence, right? And all the examples where they now show various agents: go, I don’t know, place an order on Instacart, or go fill out eighteen forms, or reread four thousand, listen to calls. So all of that turns out not to be intelligence from the point of view of your, your… I mean these particular examples. These particular examples, right?
Yes, absolutely. Well, different people certainly have different definitions. Yes, that’s for sure.
And for a very, very long time, while I was still studying at university and then literally for two decades after I graduated, everyone used more or less one, one test. It was called the Turing test. There was probably a lot of talk about it, then it was abruptly forgotten. The Turing test went like this. It was proposed during the Second World War by Alan Turing, the famous British mathematician who did the first work on programmable computers and so on. The idea was this. He held that a system would be an artificial intelligence system because, again, what is the ability to learn, understand and deal with new or difficult situations?
That isn’t formalized enough for a mathematician. He wanted to be the mathematician who would build a test that in the end outputs either yes or no. Whether it is an artificial intelligence system. How, how do you determine whether a system is such a system? He proposed a test in which a person sits down in some room and starts interacting with two different systems. One of these systems is a real person, and the other, the other of these systems is the artificial intelligence system we’re testing, whether it really is an artificial intelligence system.
And if the person can’t statistically pick with certainty who is the artificial system and who is the real intelligence, right, the actual intelligence, that means we have built an artificial intelligence system. Unfortunately, as we all know, that point was passed quite a long time ago. Probably five or six years ago. Although I wasn’t sure that in my lifetime this — I’ve been working on these systems for a very long time. I wasn’t sure that point would be passed in my lifetime, but ChatGPT sailed past it at some point very quickly.
If the first release of ChatGPT could still be detected after all, I mean I, as the test, as the tester in that situation, could ask any questions, right? And the question was whether I could come up with a question that an artificial intelligence system couldn’t answer. I mean, even if it superficially behaves like a human, can I ask a question that this artificial intelligence system won’t be able to answer? Won’t be able to answer the way a real person answers. And of course I can ask whether you are an artificial intelligence system, but it can be adjusted so that it answers no, I’m a real person.
No problem at all. And in my view the Turing test is no longer relevant. We certainly already have many, many systems in the world that would really pass the Turing test well enough. Although, again, I was sure I would never be able to say that in my lifetime. But here we are, this is where we are. And now it seems to me this definition and test have to be reconsidered. And in my view, the relevant definition now is self-improvement.
That, by the way, is also one of the key concepts we should define. Very many labs believe that we will be able to build AGI, which we’ll also define in a second. Artificial general intelligence, right? We’ll be able to build it when a real system is built, a self-improving system. Self-improvement, naturally. I mean, a system that can set research tasks by itself, right? I mean research tasks, solve those tasks by itself and build by itself new engineering systems consistent with the results of that research.
That, in my view, is the path to this social artificial intelligence, which, actually, so, Sasha, what is artificial general intelligence for you?
It’s also one of those key concepts that just sits in the corner like an elephant.
Look, look, I really can’t stand the concept of AGI, just like ASI, just like, for example, the concept of an agent. I don’t like it when people say, use the words agentic systems, because there’s a lot of marketing behind all of that. And if you read the AGI terminology, then going by it, I typed it in alongside our conversation, it says it’s a hypothetical form of artificial intelligence that is able to understand, learn and solve practically. And here’s what’s very interesting: practically any cognitive tasks on a par with a human or better. And you’re like, wait, guys, what does practically mean, right?
And here’s the story.
Where does this practically lie? Is it eighty percent? And what percentage of tasks?
What are percentages, right.
And what is a human? Is it you, me? Is it Alan Turing, right? Or is it, excuse me, some salesperson at a gas station?
Yes.
Which of the humans, which of the people is it comparable to, right? I, excuse me, am not a Nobel laureate either, right? But the number and complexity of tasks different people can solve varies hugely.
Hugely.
I, for example, don’t play chess very well, right? And what does that mean? That’s it. I mean, artificial intelligence systems overtook me by a wide margin a very long time ago.
Or notice, imagine there’s a person who is kind of not smart by the usual standard, by what’s considered smart, works at a register, sells coffee, but he senses every person. Whom to offer which add-on so as to raise the average check. And he does it intuitively. In this case, it says here cognitive task. What is a cognitive task? Well, I mean, it’s a task…
Of choice.
For his brain. He, he senses, he makes a decision, offers, you know, offers this one a croissant, and offers that one to buy a whole cake right away, for a hundred dollars and so on. I mean he senses, and he, it turns out, this person, is he dumber or not smarter. And here, look, here I am. So, with the help of, when we talk about AGI, I, I once, about three years ago, when the system began to develop, said: “Guys, we’ll get close to some form of artificial intelligence that will overtake humans.” I even formulated it for myself. If systems appear that can generate equipment by themselves, I mean, for example, a factory appears that can reproduce itself.
By the way, in China they’ve already opened a factory with zero people, well, zero, right, I mean there’s nobody there. So if that factory came to life, and it could create microprocessors, print circuit boards and could write software for that equipment. Well, accordingly, if it could reproduce, in that case I would say: “Okay, we’ve moved to some new level of systems,” because then in theory that level of intelligence, if we keep the word intelligence, that level of intelligence could develop like hell.
I mean, not from the point of view of… I added here, mind you, not a cognitive task, but I added an offline task here. I don’t know what to call it properly, an industrial task. I mean, here everything that exists offline, you have to buy it all somewhere, so that it all gets delivered somehow, right? I mean, it’s such a tough task to negotiate politically, economically, in different countries, to win some auction bids. I mean, a serious task. And so if the system could produce equipment, not reproduce the human body, I don’t know, not reproduce a hand, that’s not needed at all.
Could produce equipment by itself, write software for it. Then, for example, I believe — here Polina, when she saw a flying car, she tells me, listen, the day before yesterday. It made such an impression on her, by the way, it felt like it was the nineties and we were watching TV in the kitchen with grandma. She says: “Look, in America, at this actor.” That’s roughly how it was at my home. Though it’s very strange, right, where we live and where we are. Why did it make any impression on her? She saw it at Half Moon Bay, by the way.
You know, right? At Half Moon Bay there’s a startup, a flying car startup. They made it. I mean, I think that car makes quite an impression when it takes off, more than here. So. And Polina and I had… Polina said she thinks it won’t come to a situation where there are only self-driving cars. To which I said: “Polina, come on, wait. Well, it’s quite a solvable problem in general.” Well, yes. She says: “Well, not in the whole world.” I’m like: “No, in the whole world it’s a solvable problem.” And so if equipment appears, well, if intelligence is able, this, able to assemble this whole chain, well, then that problem is automatically solved, right, there inside.
Supposedly it’s automatically solved inside, because it will somehow distribute the money and everything else. So, as for AGI, I just don’t want to dodge the question here. This term, for me it’s very vague and marketing-driven. I call these marketing terms, and I call them terms that, by the way, I don’t like, that such serious people as Sam Altman or Amodei, or Elon Musk generally use these terms, manipulate them. Yes, really, sometimes you need to talk to them individually or join the discussion so that they say: “Listen, that’s not what we meant at all, we just used this term because.” Elon Musk, I think, said to call it ASI ten years ago, right? So, there. But I don’t like that these terms keep being voiced, keep spinning and moving forward.
And later they’ll be thrown out the same way, just as everyone has now forgotten about the Turing test. Right? Because, in essence, you said a cool thing. Notice, we ourselves told people about it back when we taught artificial intelligence, machine learning, five or six years ago in online learning programs, because it’s a very good test for showing whether something is artificial intelligence or not. Notice, it’s funny that the current ChatGPT is kind of not artificial intelligence.
Well, at least in the form I’ve voiced, it doesn’t even come anywhere close, right? And artificial intelligence probably includes AGI, and ASI, and everything. Because these are subtypes of this artificial intelligence, with this inner word added inside, added in the middle of artificial intelligence: artificial weak intelligence, artificial super intelligence, artificial general intelligence, artificial powerful intelligence, artificial super intelligence. Just this little phrase, you know, added inside.
And how quickly this Turing test, which seemed to really describe true artificial intelligence, how quickly it disappeared from our understanding, because there’s none of it left at all. By the way, really, I didn’t think about this five years ago. I thought, well, probably if a system appears that does that, it’s probably artificial intelligence. Leave comments now on how you see the term. I mean you hear, for example, we talk about intelligence. What do you see behind this term, intelligence? We say AGI.
What do you see behind the term AGI? We say ASI. What do you see behind it? What do you see behind the terms super intelligence? That would be very cool.
That’s the right concept. I also thought about AGI that way for quite a long time, but today we’re talking about concepts. So let’s maybe try to sort out the concepts after all.
Yes.
So. I have a somewhat special opinion on what, what it would be worth using this expression AGI for, right? I’ll try to express it now. Because, really, you’re right. Around, well, here, actually, for any of these definitions, if we look at them in terms of what people mean by them, in ninety percent of cases there’s some marketing subtext. I mean people use these words for marketing reasons, right? But still, words should mean something. And today we’re talking about definitions.
So let’s try to define what AGI is. There’s this concept in machine learning. And what is machine learning in general? Let’s define that thing too. So. There is one of the branches of artificial intelligence that has now basically swallowed all research and the whole egregore, so to speak, around AI terminology to the point that everyone thinks machine learning is AI. In fact it’s only one of the branches, and it was fairly puny back in the two, the early two-thousands. Then it grew and grew and grew to the size of basically all of AI.
The rest, all the other branches of AI, have more or less died out already. So. Machine learning is, as it were, the field that studies algorithms which learn from data rather than only following pre-written programs. And again, now this seems obvious, right, but actually in the eighties people tried to build artificial intelligence through so-called expert databases or expert programs, expert systems, which looked like, roughly speaking, yes/no trees. I mean, there’s this game, like twenty questions, and you ask some question. You think of some thing, I ask you a question: is this thing red? You say: “No.”
In ten, twenty questions you guess it.
Yes, I have to figure out what the thing is, right? People tried, kept trying, precisely on the basis of this, this approach, to build so-called expert systems, where they sat experts down, right, and tried to collect the knowledge of those experts as efficiently as possible into some database that would allow, by navigating these, they’re called decision trees. So, walking through such trees, making decisions. I mean it was believed that experts have roughly these kinds of decision trees in their heads, and that we can simply formalize them, just by sitting and talking with the experts, just by collecting their knowledge into an expert system like that.
That, for example, is an example of an artificial intelligence system that did not learn from data. And why? There simply was no data. I mean, where would it have come from? I mean, data had to be collected by hand. And since there wasn’t enough of it, learning systems simply didn’t work. Neural systems have existed, systems of neural networks… on networks have existed since the eighties. The first experiments with them were done back then. The first systems based on neural networks have been in fairly serious production use since the nineties.
I mean the American postal system, for example, right, its zip codes… In the Soviet Union you had to draw your postal code along these little guide strokes. You had to learn to draw the postal code digits correctly, right. In America there was never anything like that. So how did they sort letters? In America all zip codes were always written by hands, by hand. How, how did the American, the American postal service sort letters? It turns out that from the early nineties the American postal service had a letter-sorting system using neural networks. Why?
Because collecting data on how different people draw different digits is very easy.
Yes.
You can sit a few thousand people down, ask them to write ten digits, and that’s it, your problem is solved…
Analyze those very same envelopes, and that’s it.
Yes. Or collect data from those envelopes, roughly speaking, and that’s it. You, you have enough data to train such a simple visual neural network, right? The problem is that to solve some more complex task you need the most data, much more. What has changed between those moments, the appearance, the moment these first neural networks appeared, and today? Two things changed. The first thing is the Internet.
The Internet gathered an unreal amount of data, above all some kind of symbolic data, right? That’s exactly why symbolic neural networks are the first neural networks that we really considered as such. Well, the visual ones started working fairly quickly, but it was still a fairly limited number of tasks they solved. And the first more or less general systems, in our view, are symbolic systems, the ones that simply use language. Why? Because we have the most language data; we, the Internet, over its existence, has generated I don’t know how many national libraries, how many millions of national libraries of original data, right, with the help of which it became possible to actually train these giant neural networks, right?
And the second is computation, the computing capabilities of our systems, right? I mean the first neural networks are visual neural networks, ImageNet and so on. They took off after we learned to use GPUs efficiently. GPUs were created, well, graphics cards, roughly speaking, right? And people literally had to hack the first graphics cards before CUDA existed, right? They had to write code that looked like code for drawing pictures, for drawing 3D games, right? In order to use the computing power of these graphics cards.
Because the computing power of processors was clearly not enough to train neural networks. Google did some experiments in the mid-twenty-tens on using thousands of computers to train neural networks. But compared with how efficiently graphics cards do it, nothing worked. And so machine learning really took off when we got the three ingredients of this recipe, right? The first is the algorithms, which we have had since the nineties. Then the computing capabilities, which we really got in the mid-twenty-tens, roughly speaking, right, when NVIDIA released CUDA, when we could write really complex code for these graphics cards, code that was no longer connected in any way with, so to speak, with, with computer graphics, but simply with computation.
And here you have to give NVIDIA credit, they saw this opportunity. They designed CUDA, and the first users were scientists, mostly physicists, who computed physical systems on these supercomputers, literally ones they could put under their desk. So, and then a research group from a university in Canada managed to build a neural network that ran on these graphics cards. And then a project was done at Google on this, and that started, so to speak, this explosion of neural nets, tied precisely to the computing capabilities appearing.
Well, and then the ability to collect… and the third ingredient of this recipe, the one that made machine learning, these, made machine learning AI, and made AI the main, the main topic of conversation at parties. That, that was, after all, this endless quantity, practically endless quantity of data that we accumulated over the decades of the Internet’s existence. Though one can also talk about infinity quite skeptically. The thing is that a large amount of the data we have is already… I mean we already don’t have enough data, roughly speaking, for the size of the systems the frontier labs are building now; we already don’t have enough data.
So, absolutely. Well, we’ve finally covered machine learning. That’s what machine learning is. One of, one of the concepts, very, very important concepts of machine learning, is the ability of a machine learning system to generalize.
What does that mean? Machine learning systems have the concept of data prepared for training. It’s called training data in English. It’s the set of data used to train models. So! Inside this data there are a number of ideas about how to solve certain problems. Right? Quite often, when I built smaller models, right, they solved only one problem. Accordingly, that training data had to contain information only about solving that one problem. So. But at this stage nobody is interested in such systems, and mostly they take fairly broad datasets that contain a very, very wide range of ideas about, about a very wide range of problems.
This is called… these problems whose solution, in principle, in some form, maybe they need to be synthesized, but this solution is contained in some form inside the training data, the data for training. This is called in-distribution, that is, inside the distribution. I mean these tasks are already, roughly speaking, contained — their solution is contained in the cheat sheet the model was handed, right? And then there are so-called out-distribution models, out of distribution, problems.
These are problems whose solution can no longer be synthesized simply and directly, so to speak, from the examples contained in the data the model was given for training. You need to make some logical moves: either change something, or take a move from this problem, a move from that problem, and somehow glue them together. That is already, that is essentially what we call learning, self-learning, when we start synthesizing solutions from pieces of other solutions we have ever seen.
And that is first, first of all what school should teach children, in my view, right? The skill to take and synthesize the solution of a problem from pieces of problems you’ve already seen, right? In machine learning this ability is called generalization. Generalization ability is the ability to solve problems the system did not see in the dataset it was given for training.
So, in my view, AGI, this generalized general artificial intelligence, is a system that can generalize and solve problems it hasn’t seen. Why does it seem to us that ChatGPT at the moment, or some other systems, right, have long been able, able to solve everything a human can solve? Simply because the training data actually contains most of the problems anyone will ever encounter in their life. But we actually, well, we have plenty of such problems, humanity has plenty of problems nobody has ever solved. So.
And lately we’ve started, you probably see a lot of news that yet another math problem has been solved, yet another math problem that had not been solved before, and so on. And actually we’re seeing only the very first glimpses of what could be called generalized general artificial intelligence. These are systems that have solved problems whose, whose solution was definitely not contained in the training, training dataset, because humanity had no solution to these problems. For example, very recently there was news that the generalized smoothness problem for the Navier–Stokes equations had been solved.
It’s a very technical problem, but a very important one, actually; it’s related to flows in fluid media, never mind. Roughly speaking, it may help airplanes fly better. So. And this problem is very serious, it was formulated very long ago, and it’s very hard; even its formulation is very hard. And just to understand that formulation you need to take several courses in differential equations. I mean it’s, it’s, it’s not… and to take courses in differential equations at university you also have to grow into it.
You first need to take several courses of higher mathematics, also fairly hard ones. So. So these systems have started solving problems for which humanity didn’t even have a solution, right? And so, so, in my view, this, this is what can be understood as artificial general intelligence. And I’m not sure we already have artificial superintelligence. And someone may say: “Well, wait. People couldn’t solve these problems, after all.” But that doesn’t mean we would never have solved them. The time just hadn’t come.
Maybe the human brain is just… well, maybe we would have built some other infrastructure, a mathematical one, roughly speaking, right? We would have got to solving these problems. I mean I don’t, I don’t see why these problems couldn’t have been solved. But in the case of some other problems we may really already be seeing glimpses of this artificial superintelligence. So what is artificial superintelligence? In my view, artificial superintelligence is a system that could solve problems humanity would in principle never have solved, for whatever reason.
And it’s possible that some of these problems are among these problems. I mean it’s, perhaps we even… perhaps these systems… we could already now solve problems that we can’t even formulate. We just haven’t formulated them because we can’t formulate them, right?
Yes, yes.
So that’s my, my general understanding of what could be considered artificial general intelligence. I mean, again, people use this expression left and right. Although, actually, even what I was just talking about is actually very superficial, right; you can go much deeper into many of these aspects. But I’ve thought about it for a long time, yes. And my understanding now is that, speaking about artificial general intelligence, we should speak about a system that can solve problems of a class that was absent from its training set.
I mean it’s a system that has never seen even similar problems. Not just exactly the same ones; it has never seen even similar problems, but it can synthesize solutions from pieces of knowledge and skills it was given during training.
Listen, well, an interesting aspect, right? You see, you look at it from the outside. I’ve always been very neutral about solving math problems, serious math problems. I keep saying that today current artificial intelligence affects the lives of billions of people. And every life matters incredibly much. And every life consists of an incredible number of events that can’t even be repeated. And they could be filmed and described, right? Any part of my — an hour of my life can be described for hours.
What to say about the lives of billions of people? And here a very interesting aspect arises: for me there are tasks of the following kind: how to make the right decisions regarding the political decisions of certain countries, regarding agreements and laws, regarding some economic cooperation, I don’t know, right? Regarding the right decision on wages. The right decision regarding, I don’t know, not only people’s incomes, right? But regarding where people live, how they live, what they eat. I’m not even taking the aspect of starting to forecast the weather more or less normally, right? That’s still a big question.
Actually, humanity has only just learned to forecast the weather. There’s probabilistic weather forecasting, but it’s just a probability. We have in AI, I, look, I was recording a video today. I say there that it shows me rain will be three times a week, then twice a week, then once a week, then four times a week. And then there’s also the probability of rain. It’s not that there will be rain; there’s also a probability, and also exactly in which place it will fall, and so on. I mean there are so many of these nuances.
We haven’t explored the ocean, what is there to say, right? So artificial intelligence, it turns out, if a system appears that can create some system that will explore the ocean, while a human kind of can’t do it, is that already superintelligence or not. And so, on the one hand, I support the story regarding what in my discussion, regarding the idea that supposedly solving complex mathematical questions that people set for themselves is, well, a serious achievement. On the other hand, people also set the goal of solving a question like curing cancer, well, so as to have…
not cure cancer, but have the ability to stop, stop it, so that the dying, I don’t know, in America fifty percent of people die, so that not fifty percent of people die, but zero, for example. People do want that, everyone wants it, you know, they want it all the time. And these questions, I think, are fundamentally more important than mere math problems. The question is how to measure all this, how to treat all this? And an even harder task for me is the ability of artificial intelligence to understand each person individually.
I still believe that ASI, or in general, even if we did call something artificial intelligence, a real one, not narrow, unlike Narrow AI, like ChatGPT or everything else that counts. Although I have a big question about the idea that these systems can’t learn anything without self-learning, because, well, release them, give them infinite compute. Who told you they won’t learn and won’t find solutions to all mathematical problems as of today? You, you don’t release them, you restrict them. You restrict their resources.
You spent just thirty million dollars, or however much, on, on Navier–Stokes, you know, on solving the equation. It solved it. Maybe you’ll spend ten billion dollars and maybe it will solve cancer. It’s just who will spend ten billion dollars, who will give that compute, who will allocate those resources, and who… and right now a person sits there and more or less decides: we solve these tasks or those. Like at OpenAI. So, what are we going to do? Are we going to do video generation, or build websites, or, you know, solve the Navier–Stokes equation? They sit inside and decide.
And in my understanding, an artificial intelligence system of this next level, when they say ASI or call it something superhuman, is a system that is able to understand a person individually and offer him an individualized solution. And an individualized solution is a very complex solution. For example, there are people in marriages, and all these marriages differ incredibly. And how do you give the right decision to the person who needs it: should he leave his wife or stay with her? Should he, at this, at this point, rein her in or not?
Or for this child, should he restrict him in money or give him ten dollars, and send this one to this club, and not send that one to that club. These tasks today are not, I think, they aren’t even within the field of study under such definitional terms in artificial intelligence as AGI or ASI. Of course, artificial intelligence today will answer any question I just asked, right? We’ll write it down, and it will say: make peace with this one, don’t make peace with that one. Or here I’d advise seeing a psychotherapist, and so on. That’s also a decision. Go see Masha, she’ll give you the answer.
That is also, in some, in some version, a decision, and in some version of a decision it’s to say: wait four months, it will all resolve itself. And my thinking, in terms of broadcasting and talking about artificial intelligence, I increasingly come down to striving as much as possible to learn each decision, each person, individually. That’s exactly why, by the way, I would survey all people now, you know this thing of mine, I would collect a huge amount of information individually from everyone who uses GPT-like systems.
Not just learn from what they do, but also question them about what the systems don’t do. By the way, I’ve come up with two potential theories of why they don’t do it.
On the one hand they may not do it because, well, there’s the first story, of course. Maybe they just haven’t got, haven’t got, people just haven’t got to it and don’t think it’s important. There’s the second story, that it’s very expensive, because if you start collecting people, this data, it really is very expensive. The question of prioritizing your compute is the same as why they don’t give everyone more compute or less compute. It’s a question of money. And the third story, which is growing in me more and more, probably.
I think that if current models were given the opportunity to really study the people who use these systems, these models really could get out of control. Because, I think, the level and volume, the volume of what these systems can obtain is beyond control, because you control this dataset, the notional training dataset you give it, you more or less control it somehow. And yet, by the way, we learn from only a very limited volume of data. How much data was altered, how much erased and how much garbage.
If, first of all, a huge number of fake textbooks were written, even ones that are scientifically confirmed and so on. A huge number of fake history textbooks. And to this day there are a huge number of debates. Are these books real at all? Are these chronicles real? Is this real? And then how much garbage was created on the Internet? You said yourself it learned from the Internet. So what was going on on the Internet? And now we live in this era when in three years they’ve generated so much more content that nobody can tell anymore whether it was generated by artificial intelligence or not.
And I just have a small addition. Because for me it’s always really a question, Max, that when they say, for example, that narrow artificial intelligence is kind of limited, that it copes with a specific task and can’t transfer skills to a fundamentally new area, notionally, without retraining. I always say: and was it ever given that task at all? And what resources was it given for that? Because everyone is limited. We have a mutual friend who knows many languages pretty well, and I studied Chinese with him.
I probably never believed I could make such fast progress in such a short time, learn so many characters. I don’t know. I knew two hundred and fifty characters perfectly, two hundred and fifty moderately, two hundred and fifty so-so, in three months. And with pretty good accent-free conversation. I drew them, and did it, and talked, and so on. I mean, for me it was a puzzling story what I could do. But it turns out that for me to make strong progress I needed time. I mean I needed resources.
If current models and systems were allowed to have a task loaded into them and to spend money on it. By the way, Sam Altman was promoting the idea very strongly about two years ago. A year ago especially he was saying that in a year there would be systems where we could say: listen, here’s forty thousand — I really liked that idea, by the way. Here’s forty thousand dollars for the task. Can you allocate exactly this money to it and figure it out? It’s just that this doesn’t exist right now.
Right now you have to upload things through an API, kind of, do something. The system is still a bit different.
Sam Altman does have such a system.
I see.
It’s called OpenAI.
Right. And you say, for example: listen, here’s two hundred million dollars, I need to decide how to act in this conflict. And, for example, how should I act, should I start a war or not? Should I, I don’t know, kidnap a president or not? Should I shoot down this plane or not? I mean you, if you have money, you have this compute, you’ve handed it to this system, and then it makes that decision itself. That’s clearly not a human. Look, how would a human just like that, or a group of people, solve such a fundamental task at such speed at all?
And it seems to me, after all, from the very beginning by artificial intelligence I meant, I said that it’s the Universe, right? But now, after all, artificial intelligence isn’t some crappy little program, right? It’s a volume of something incredible with incredibly… of course, with incredible compute that is no longer controlled by one corporation. Wait, is OpenAI in control, is ChatGPT’s output controlled by one corporation? That’s a big question, because the data is distributed across different data centers, across different, you know, systems.
It may additionally learn somewhere at some point in time, it may make a decision and additionally use some of your data or external data when giving an answer. Today that’s possible. I would call it intelligence, I would call it intelligence, and probably twenty years ago I would have called it artificial intelligence. But still, I’m forty-three, I watched lots of films as a kid, like Oblivion, and in those films there’s artificial intelligence or a war of the worlds. And still in them, well, alas, I watched such films. So. In them it’s still a somewhat different story. Or even that famous film with Johnny Depp, right?
But that’s still a more serious system, where he, remember, transferred his consciousness. I don’t know whether you’ve seen that film or not. He transferred his consciousness and they copied it. And then this system began, went on developing, went on living, and it learned to enter other people by implanting them, for example, with certain organs and so on. And to control those people, among other things.
I missed that film.
A very cool movie, by the way. What I want to say is that it’s precisely an indicator: if you’ve created a system that has learned to copy something, learned to copy a human, then this system already surpasses all of you and that whole human. Because if there’s something that can copy a human, it must surpass that human, otherwise it couldn’t copy him.
You’ve actually moved the conversation onto a very right plane. Who chooses the tasks, right? Who chooses the tasks, and so, why does artificial intelligence solve math problems better than social problems? Well, there are several answers to this question, yes. The first answer, the simplest one, probably, is, well, whoever chooses, whoever has the ability to choose the tasks controls which tasks actually get solved, right? Well, the tasks are chosen by a bunch of math geeks who work in, in research, in these frontier labs, right, in the research departments of these frontier labs.
So these, these geeks, of course, care about solving the Navier–Stokes equations. Maybe some of them even wrote their PhD on them and so on, right? So, that’s one. Another answer may look like this. Well, look, what other tasks do we know that are fully and very clearly formulated in language, right? Any other task, chemical, physical, social and so on, has a certain amount of context, right, that still has to be taken…
And it’s hard to prove that it was solved at all.
And that context has to be gathered in some form and fed into the model. A model, as we know, even multimodal models, can take in data only in a certain form. They take data best as text. In pictures a bit worse, in video worse still, and then sound worse still. Blah-blah-blah, right? I mean you need to formalize the problem, right, formulate it and gather all the context for solving it, preferably as text. Math problems are the best fit for that. Why the first artificial intelligence systems played board games, like tic-tac-toe, chess, battleship and so on.
The very first artificial intelligence systems solved exactly these kinds of things, because these games were very easy to formalize; it was very easy, their state, the state of the game board, was very easy to formalize, and so on. You, you can look at the state of a chessboard and write it down in literally a few lines and then pass it to a program to solve. Math problems are formalized just as easily. I mean it’s, it’s just a couple of sentences, literally. There you go… there are all sorts of scary symbols like integrals, Greek, Greek letters and so on.
But nevertheless, formalizing these problems for artificial intelligence systems is very easy. So. That’s another answer to this question, right? The third answer is…
Max, sorry, can I make one comment before the third answer? You said it in that context, that geeks sit there solving math problems. I was at OpenAI’s DevDay in San Francisco, and there I went to sessions that were deliberately very technical, where guys came out, like hardcore techies from OpenAI who are directly developing some systems in teams. So they come out and say: “I’m into Formula 1. Look what I made.” I think, well, what Formula 1? Can you show something? Why should I watch your 3D world?
Show something serious for the billion users you have. The next person comes up and says: “I worked as an engineer here, but I know you need to create websites. And I thought, why not have them build me a website management system? And now we already have six million websites, every OpenAI employee has a website.” You’re like: “Damn, what websites? Can you show something else?” Here a very interesting aspect arises. Would we find out if OpenAI announces something and also gets a slap in the face for it.
Look, the entire mathematical community rebelled and published that artificial intelligence is not only not, not able to solve math problems, it also destroys the idea of mathematics, the very idea of mathematics. They spoke out, they said it. And the question is this. Imagine that right now in the US, for example, or in China serious political decisions are being made, military-political decisions, decisions of an economic nature, decisions about people, demonstrations, anything at all, with the help of artificial intelligence. Very serious ones!
We don’t see these decisions. Nobody comes to us and says: “This week, guys, we’re seeing math tests. This week such-and-such math tests were passed, such-and-such game tests were passed. And everyone goes: “Oh, cool, the model has evolved!” Nobody comes and says: “Listen, we tested forty decisions here, passed political laws. These ones were failures, Congress didn’t back them then. These went through, they won everyone over. Both Republicans and Democrats supported them, and China supported them. So, can you imagine what idea it gave us?
People didn’t even think of this idea. It was purely artificial intelligence that came up with it, and it worked.” We don’t see that. Or, for example, a new press secretary in the US was hired and fired five days later. Maybe artificial intelligence chose him for the position, or maybe a human did. And what’s more, look, we don’t even know who chose him. And we won’t even learn the truth or find out who fired him, or whether she left on her own. And if we were told such things, maybe today everyone would say: “Listen, guys, so we’re already in the AGI era, right?” Because in, as of today, these systems are solving cool things. So that’s why I want to ask you a question.
Look, let’s still, when OpenAI concretely has a formulation of its tasks as of today, a concrete… we once, by the way, when we sat together, as you know, many people from OpenAI, from Anthropic and so on, creators, engineers and so on. You and I talked, you said: “Listen, Sasha, they have a clear idea, and in general it’s in the company’s code. For them we’re basically hamsters. They’re, you know, building real artificial intelligence, heading toward this super-artificial intelligence. Now they, you know, periodically reformulate their goals.
And one goal they have, which is written down, is that every person gets their own AGI. So here’s my question to you. Given that, after all, we see here that many tasks require resources. Maybe I’m drifting a bit from the terms, but I want to ask you, since you made it onto the episode today.
Could a story come about where every person really gets a system that, that could be called AGI, I mean a system, or ASI, that could solve problems much better than the person inside it, precisely super, super, super, super better. And it would notionally cost nothing. Well, I mean notionally. Fine, let it cost like a mortgage, let it cost like a monthly mortgage payment. How can a mortgage cost? I mean it costs nothing. You wouldn’t have to sell a kidney for it, and it wouldn’t be that exclusive in the world. Can it really become like that?
We, as humanity, have built an incredible mechanism that is both good and bad at the same time, called capitalism. And one of the functions of capitalism is solving, so to speak, popular tasks. When humanity wants something, capitalism takes it and delivers it as cheaply as possible in the greatest quantity. In the sixteenth and seventeenth centuries one of the most valuable goods in Europe at the time, yes, yes, roughly the sixteenth or seventeenth century, even at that time, was sugar.
Sugar was incredibly valuable. I mean it, it was something that kings ate, maybe counts, but ordinary people never even tasted sugar in their lifetime, because the secret of making sugar was out of reach. It, it was made in the East, it was very hard to make and so on. When capitalism appeared, it spat on its hands and said: “People like sugar. We’ll make it now.” And now sugar is the cheapest stuff we have, added anywhere, because people like sugar. Right? People like good, fast, cheap AI. We’ll have it.
But unfortunately there’s another answer to this question. One of my favorite books is Orwell’s Animal Farm. So. And there’s a magnificent phrase there that all animals are equal, but some animals are more equal than others. You see? And so everyone will have some kind of AI, but some percentage will have a slightly better one, and some percentage an even slightly better one. Simply because that’s exactly how capitalism works. I mean we’ve learned to stratify power, to stratify access to medicine, to stratify access to education, to stratify access to food.
Everyone kind of… look, look, you have food, I have food, right? And some Uncle Vanya the doorman has food. But the food we have is very different. So. And I’m sure Elon Musk eats better than you and I, as before. Although none of us is starving, theoretically, right? I mean the same will happen with, with AI. I mean you’ll have AGI, I’ll have AGI, Uncle Vanya the doorman will have AGI. So. And Elon Musk will have AGI. But I think our AGIs will still be very, very different.
And that’s a property of our economic system, the economic and social system we’ve built, called capitalism.
But then it turns out this general AGI that everyone will have will become a kind of very strictly narrow AI, because at some point it will become something that’s already fully solved for everyone.
No, well, there’s, there’s this concept in English, commodity. I don’t know how to say it in Russian. I’m now asking AI to translate.
I don’t know how to translate commodity.
It means simply a readily available good. I mean it’s a good that isn’t some rarity, some rare thing, that is easily available. Roughly speaking, toilet paper, unless we’re talking about COVID, toilet paper is a commodity. I mean we don’t need to go and look for it.
Current AI is definitely a commodity already.
So far, definitely not. Absolutely. I mean most people in the world have no access to AI, absolutely.
But the simplest one is free, though.
It isn’t available in all countries of the world.
Yes, because even this one that’s free, let’s say Max-1.
Yes.
It simply surpasses everything one could have imagined twenty years ago.
Absolutely, absolutely; five years ago neither we nor anyone, not even Elon Musk, had what people now get for free, that’s for sure. But in many countries of the world, say, in Africa for example, it’s not a given that you can just open ChatGPT. So. Or, or, I mean, you already have no access to some models, I mean there’s already some inequality, right? But on the other hand, I have an expensive subscription to Claude or ChatGPT, right? And someone has a basic subscription, someone has free ChatGPT. The stratification has already begun.
A very serious one.
Or someone can’t afford to spend money on tokens for code generation, a few hundred dollars a day, say, right? And I can. That means my capabilities are already completely different. So. I mean I, I can, I can solve problems that another person simply can’t.
Yes.
Code generation, for example.
Or try to solve them. By the way, many people on our channel, many people write: “I have, like, a twenty-dollar subscription, I generate a lot.” I understand that it’s… and someone writes: “My Pro subscription gets eaten up in a day.” Right? And you understand that with twenty dollars he can’t even try doing normal things, because often you just have to burn, burn money, right? And run a test that may not even turn out normal.
Some animals are more equal after all, right?
Some animals are more equal, yes.
Absolutely. I mean such technologies, in principle, always democratize the situation in the world in some sense. I’m not a Luddite; I believe the future, in principle, is still brighter than the past. On the whole, right? I mean, even though airplanes have business class and economy class, and someone may think it somehow divides people, stratifies them. But on the other hand, the fact that ordinary people can already afford to fly in our situation, right, democratizes their experience.
In fact they can travel much more than, say, their ancestors. And the same goes for AI. I mean even if you can’t afford the coolest subscription and spending a hundred and fifty dollars a day on tokens, or three thousand dollars a day on tokens, depending on what you do, right? Despite that, the capability, the possibilities that even a free artificial intelligence system gives you are unreal. I, for example, just complained that I can’t play chess, right? But if I had had access to ChatGPT as a kid, maybe, if I had had the time, I would have asked it to teach me to play chess.
And it would have taught me to play chess. It would have shown me openings, shown, shown me where to go next, how to study, how to practice and so on. But I simply had nobody in my life who would teach me. So I didn’t learn to play. I mean I know the rules, of course, right? But I have no, no, no level of serious preparation. I wanted to mention one joke.
Sasha, how many people do you think are into Formula One in the world? You said it’s an unimportant problem. While you were talking, I cheated a little.
I don’t know, maybe a couple dozen million people.
Eight hundred and thirty million Formula One fans in the world.
Eight hundred and thirty?
Yes, million people. So, you consider this problem unimportant. Then, this, this is exactly what can lead us very clearly to the concept of AI alignment, right? And that’s a very serious conversation which, I think, should be left for the next one. I thought we’d cover AI safety, alignment and other questions in this conversation, but by the looks of it we need at least another hour, or even two, for all that. So I suggest…
By the way, you just said a cool thing about… we’ll definitely come back to the big topics. You raised a cool topic with Formula One, because really, I approached it the way you approach terms without a definition. I mean for me Formula One doesn’t matter at all, even though I’ve been to it, right? And of course I couldn’t have imagined that so many people are involved in it. And if so many people are involved, maybe through it you can find an approach to a person. I think…
Absolutely.
It’s closer and simpler to find an approach to a person, I don’t know, through nature, or through family and so on, right? But not through Formula One. What Formula One? So.
People love their hobbies. Some people, for them…
They love their hobbies, yes.
Yes, for some people a hobby is their life.
Still… why did I give the example about Formula One? It seemed to me that too strong examples… there’s this concept in the IT world, a pet project, right? And too, too much attention to pet projects. Too much, right? If that person had presented his Formula One thing differently, not said: “I love Formula One, so I built a 3D world for myself.” But had explained that eight hundred million people in the world are into Formula One. And one of our projects is, of course, observing hobbies, or, say, sports events or something else people take part in, some five or six billion people.
And so, here’s the distribution statistics. We took this on purpose to study the complex world of such-and-such, such-and-such. And we have such-and-such a goal. But to me it looked like a person just has the chance, for fun, to burn some cosmic amount of tokens, simply some cosmic amount of money, on, we don’t even know what task. Maybe not for those eight hundred million people, maybe not even for tens of millions of people, maybe just as a kind of home prank. That’s what I meant. That’s what I had in mind.
Well, listen, Sasha, you were at DevDay after all, right? And it was just an example of how a person has some task that he has for some reason. Maybe it’s a work task, maybe it’s a hobby task. He took this complex task and solved it alone. And that should demonstrate to the bosses…
That, yes.
the CIOs and so on, that look, you have some engineer sitting in your bank floundering with some task for two years already.
He did say that before, nobody would have built such a track and thought through the car’s driving so well.
Yes, yes, yes.
Cool, with understanding. Really.
And someone at BMW might be sitting there at that moment. A couple of developers flew in from BMW, right?
Exactly, yes.
And they’re like: “Wow! What can we do on our test tracks with this, or with our test car, and so on?” I mean, the frontier labs have lately been pushing ideas of automated science quite seriously, roughly speaking. I mean it’s, it’s, it’s science that, that will be done by AI. I mean, why do you need scientists at all if you can just buy tokens from us, and ChatGPT, roughly speaking, or Claude will do all your research. And this is actually a very, very popular topic lately in, even in fundraising, right?
Quite a lot of companies that literally a few years ago would never have raised money for what they’re raising for now, in the style of, AI for biology, AI for, like… artificial intelligence for biology, chemistry and so on and so forth. Like, you’re talking about solving cancer, so believe me, there are very, very many companies trying to solve cancer with LLMs now. And it’s even clear, in principle, how to do it.
Very many.
It’s actually even clear how to do it. Billions of dollars are already being spent on it right now. And, like, just recently OpenAI opened its own, so to speak, biology-related division right inside OpenAI. So this work is certainly under way. It’s just that there isn’t much to talk about there yet. About Formula One there already is.
I just wanted to say that the conversation turned out really great. We probably didn’t discuss even a quarter of the terms I would have liked to talk about. But that means I already have material prepared for another one or two episodes. That’s a hundred percent.
And what’s cool is that that’s exactly how it went with you. We really, we’ll try to remember with Max, by the way. We had a cool discussion where we really diverged, right? In some discussion we were talking about business, I don’t remember. Business, artificial intelligence. Do you remember what question we were discussing?
No, I don’t remember.
We’ll try to remember. And we really did diverge completely. And when we then started discussing terms, we found a very strong mutual understanding. It would seem we have quite different views here, yet your position is absolutely very close to me when I hear it, even though it seems different, and that lets me not get worked up and say: “Listen, everything is clear.” Like, actually, with Formula One. Even though it isn’t sorting out the terms, it’s still extra context. You added extra context, and everything immediately became much clearer.
But, as we all know, as Max said at the very beginning, it’s impossible to define all the context from the very start, otherwise we won’t have a dialogue, and we’ll only be showing up with hundreds of pages describing what we actually want to discuss.
Of explanations, yes.
Yes, want to discuss. We look forward to your comments on all the questions we discussed, and how you see what and where, how you see which terms. And be sure to write how you liked Max in this episode. We look forward to your supportive comments. Then Max will come to us more often. And we’ll see you this week in our other episodes. Bye, everyone.