Stephen Wolfram: ChatGPT and the Nature of Truth, Reality & Computation | Lex Fridman Podcast #376

Stephen Wolfram 0:00 You know, I can tell ChatGPT, "Create a piece of code," and then just run it on my computer. And I'm like, you know, that- that sort of personalizes for me the what could, what could possibly go wrong, so to speak.

Lex Fridman 0:11 Was that exciting or scary, that possibility?

Stephen Wolfram 0:14 It was a little bit scary actually, because it's kind of like, if you do that, right?

Lex Fridman 0:18 Yeah.

Stephen Wolfram 0:18 What is the sandboxing that you should have? And that's sort of a, that's a- a version of- of that question for the world. That is, as soon as you put the AIs in charge of things, you know, how much, how many constraints should there be on these systems before you put the AIs in charge of all the weapons and all these, you know, all these different kinds of systems?

Lex Fridman 0:36 Well, here's the fun part about sandboxes, is, uh, the AI knows about them and has the tools to, uh, crack them. The following is a conversation with Stephen Wolfram, his fourth time on this podcast. He's a computer scientist, mathematician, theoretical physicist, and the founder of Wolfram Research, a company behind Mathematica, Wolfram Alpha, Wolfram Language, and the Wolfram Physics & Metamathematics projects. He has been a pioneer in exploring the computational nature of reality. And so, he's the perfect person to explore with together the new quickly evolving landscape of large language models as human civilization journeys towards building super intelligent AGI. This is the Lex Fridman Podcast. To support it, please check out our sponsors in the description. And now, dear friends, here's Stephen Wolfram.

Lex Fridman 1:33 You announced the integration of ChatGPT and Wolfram Alpha and Wolfram Language. So, let's talk about that integration. What are the key differences, from the high philosophical level, maybe the technical level, between the capabilities of, broadly speaking, the two kinds of systems, large language models and this computational, gigantic computational system infrastructure that is Wolfram Alpha?

Stephen Wolfram 1:57 Yeah. So what does something like ChatGPT do? It's- it's mostly focused on make language like the language that humans have made and put on the web and so on.

Lex Fridman 2:08 Yeah.

Stephen Wolfram 2:09 So, you know, its- its primary sort of underlying technical thing is you've given a prompt, it's trying to continue that prompt in a way that's somehow typical of what it's seen based on a trillion words of text that humans have written on the web. And the way it's doing that is with something which is probably quite similar to the way we humans do the first stages of that, using a neural net and so on, and just saying, "Given these, given this piece of text, let's ripple through the neural net one wo- and- and get one word at a time of output." And, uh, it's kind of a- a shallow computation on a large amount of kind of training data that is what we humans have put on the web. That's a different thing from sort of the computational stack that I spent the last, I don't know, 40 years or so building, which has to do with what can you compute many steps, potentially a very deep computation? It's not sort of taking the statistics of what we humans have produced and trying to continue things based on that statistics. Instead, it's trying to take kind of the formal structure that we've created in our civilization, whether it's from mathematics or whether it's from kind of systematic knowledge of all kinds, and use that to do arbitrarily deep computations to figure out things that- that aren't just, "Let's match what's already been kind of said on the web," but, "Let's potentially be able to compute something new and different that's never been computed before." So as a, as a practical matter, you know, the- the, um, what we're, you know, the- Our goal is to have made as much as possible of the world computable in the sense that if there's a question that in principle is answerable from some sort of expert knowledge that's been accumulated, we can, uh, compute the answer to that question. And we can do it in a sort of reliable way that's- that's the best one can do given what the expertise that our civilization has accumulated. It's a very, it's a- it's a much more sort of labor-intensive on the side of kind of being, creating kind of the- the computational system to do that. Um, obviously the, in- in the, the kind of the ChatGPT world, it's like take things which were produced for quite other purposes, namely the, all the things we've written out on the web and so on, and sort of forage from that things which were, are like what's been written on the web. So I think as, you know, as a practical point of view, I- I view sort of the ChatGPT thing as being wide and shallow and what we're trying to do with sort of building out computation as being this sort of deep, also broad, but- but, uh, most importantly, kind of deep type of thing. I think another way to think about this is if you go back in human history, you know, I don't know, 1,000 years or something, and you say, "What- what can the typical person, what's the typical person going to figure out?" Well, the answer is there are certain kinds of things that we humans can quickly figure out. That's sort of what- what our, uh, you know, at our neural architecture and the kinds of things we learn in our lives let us do. But then there's this whole layer of kind of formalization that got developed in which is, you know, the kind of whole sort of story of intellectual history and whole kind of depth of learning. That formalization turned into things like logic, mathematics, science, and so on. And that's the kind of thing that allows one to kind of build these towers of- of, uh, uh, of- of, uh, sort of towers of things you work out.

Lex Fridman 5:45 Mm-hmm.

Stephen Wolfram 5:45 It's not just, "I can immediately figure this out." It's, "No, I can use this kind of formalism to go step by step and work out something which was not immediately obvious to me." And that's kind of the story of what- what we're trying to do computationally, is to be able to build those kind of tall towers of what implies what implies what and so on. Um, and, uh, as opposed to kind of the, "Yes, I can immediately figure it out. It's just like what I saw somewhere else in something that I heard or remembered," or something like this.

Lex Fridman 6:14 What can you say about the kind of formal structure or the kind of formal foundation you can build such a formal structure on, about the kinds of things you would start on in order to build this kind of, uh, deep computable knowledge?

Stephen Wolfram 6:29 ... trees. So the question is sort of how do you, how do you think about computation? And there's, there's a couple of points here. One is, what computation intrinsically is like, and the other is, what aspects of computation we humans with our minds and with the kinds of things we've learnt can sort of relate to in that computational universe. So if we start on the kind of what can computation be like, it's something I've spent some big chunk of my life studying, is imagine that you're, you know, we, we, usually write programs where we kind of know what we want the program to do and we carefully write many lines of code and we hope that the program does what we, what we intended it to do. But the thing I've been interested in is, if you just look at the kind of natural science of programs. So you just say, "I'm gonna make this program. It's a really tiny program. Maybe I even pick the pieces of the program at random." But it's really tiny and by really tiny, I mean, you know, less than a line of code type thing. You say, "What does this program do?" And you run it. And big discovery that I made in the early '80s is that even extremely simple programs when you run them can do really complicated things. Really surprised me. It took me several years to kind of realize that that was a thing, so to speak. But that, that realization that even very simple programs can do incredibly complicated things that we very much don't expect, that discovery, I mean, I realized that that's very much I think how nature works. That is, nature has simple rules, but yet does all sorts of complicated things that we might not expect. You know, as a big thing of the last few years has been understanding that that's how the whole universe and physics works. But that's a, a quite separate topic. But so there's this whole world of programs and what they do, and very rich sophisticated things that these programs can do. But when we look at many of these programs, we look at them and say, "Well, that's kind of... I don't really know what that's doing." It's not a very human kind of thing. So on the one hand, we have sort of what's possible in the computational universe. On the other hand, we have the kinds of things that we humans think about, the kinds of things that are developed in kind of our intellectual history. And that's, uh... And the, the... Really, the challenge to sort of making things computational is to connect what's computationally possible out in the computational universe with the things that we humans sort of typically think about with our minds. Now, that's a complicated kind of moving target because the things that we think about change over time. We've learnt more stuff. We've invented mathematics. We've invented various kinds of ideas and structures and so on. So it's, it's gradually expanding. We're kind of gradually colonizing more and more of this kind of intellectual space of possibilities. But the, the real thing, the real challenge is, how do you take what is computationally possible? How do you take... How do you encapsulate the kinds of things that we think about in a way that kind of plugs into what's computationally possible? And, and actually the, the, uh, the big sort of idea there is this idea of kind of symbolic programming, symbolic representations of things. And so the, the question is, when you look at sort of everything in the world and you kind of, you know, you take some visual scene or something you're looking at and you say, "Well, how do I turn that into something that I can kind of stuff into my mind?" You know, there are lots of pixels in my visual scene, but the things that I remembered from that visual scene are, you know, there's a, there's a chair in this place. It's a kind of a symbolic representation of the visual scene. There are two chairs and a table or something, rather than there are all these pixels arranged in all these detailed ways. And so the question then is how do you take sort of all the, all the things in the world and make some kind of representation that corresponds to the types of ways that we think about things? And, and human language is, is sort of one form of representation that we have. We talk about chairs. That's a word in human language and so on. How do we, how do we take... But human language is not in and of itself something from... that plugs in very well to sort of computation. It's not something from which you can immediately compute consequences and so on. And so you have to kind of find a way to take sort of the, the, the stuff we understand from human language and make it more precise. And that's really the s- story of, of symbolic programming. And, you know, what, what that turns into is something which I didn't know at the time it was going to work as well as it has, but back in the 1979 or so, I was trying to build my first big computer system and trying to figure out, you know, how should I represent computations at a high level? And I kind of invented this idea of using kind of symbolic expressions, you know, structured as, it's kind of like a, a function and a bunch of arguments, but that function doesn't necessarily evaluate to anything. It's just a, a thing that sits there representing a structure. And so building up that structure, and it's turned out that structure has been extremely... It's a, it's a good match for the way that we humans... It seems to be a good match for the way that we humans kind of conceptualize higher level things. And it's been for the last, I don't know, 45 years or something, it's served me remarkably well. So building up that structure using this kind of symbolic representation. But what can you say about abstractions here? Because you could just start with your physics project. You could start at a hypergraph at a very, very low level and build up everything from there. But you don't. Right. You take shortcuts. Right. Uh, you, you take the highest level of abstraction, convert that, uh, the kind of abstraction that's convertible to something computable using symbolic representation. And then th- that's your new foundation for that little piece of knowledge. Yeah. Somehow all of that is integrated. Right. So the, the sort of... A very important phenomenon that, that is kind of a thing that I've sort of realized is just, it's one of these things that sort of in the, in the future of kind of everything is going to become more and more important is this phenomenon of computational irreducibility. And the, the question is, if you know the rules for something, you have a program, you're gonna run it, you might say, "I know the rules. Great. I know everything about what's gonna happen."Well, in principle you do, because you can just run those rules out and just see what they do. You might run them a million steps, you see what happens, et cetera. The question is, can you, like immediately jump ahead and say, "I know what's gonna happen after a million steps-"

Narrator 12:58 Yeah.

Stephen Wolfram 12:58 ... and the answer is 13 or something.

Narrator 13:00 Yes.

Stephen Wolfram 13:00 And the- the- one of the very critical things to realize is, if you could reduce that computation, there is, in a sense, no point in doing the computation.

Narrator 13:10 Yeah.

Stephen Wolfram 13:10 The place where you really get value out of doing a computation is when you had to do the computation to find out the answer. But this phenomenon that you have to do the computation to find out the answer, this phenomenon of computational irreducibility seems to be tremendously important for thinking about lots of kinds of things. So one of the things that happens is, okay, you've got a model of the universe, at the low level, in terms of atoms of space, and hypergraphs, and rewriting of hypergraphs, and so on. And it's happening, you know, 10 to the 100 times every second, let's say. Well, you say, "Great, then we- we've nailed it. We- we know how the universe works." Well, the problem is, the universe can figure out what it's gonna do. It does those 10 to the 100, you know, steps. But for us to work out what it's gonna do, we have no way to reduce that computation. The only way to do the computation, to see the result of the computation is to do it. And if we're operating within the universe, we're kind of... There's no, there's no opportunity to do that, 'cause the universe is doing it as fast as the universe can do it.

Narrator 14:07 Mm-hmm.

Stephen Wolfram 14:08 And that's, you know, that's what's happening. So what we're trying to do, and a lot of the story of science and a lot of other kinds of things, is finding pockets of reducibility.

Narrator 14:17 Mm-hmm.

Stephen Wolfram 14:17 That is, you could have a situation where everything in the world is full of computational irreducibility. We never know what's gonna happen next. The only way we can figure out what's gonna happen next is just let the system run and see what happens. So in a sense, the story of, of most kinds of science, inventions, a lot of kinds of things, is the story of finding these places where we can locally jump ahead.

Narrator 14:39 Mm-hmm.

Stephen Wolfram 14:40 And one of the features of computational irreducibility is there are always pockets of reducibility. There are always places, there are always an infinite number of places where you can jump ahead. There's no way where you can jump completely ahead, but there are little, little patches, little places where you can jump ahead a bit.

Narrator 14:56 Mm-hmm.

Stephen Wolfram 14:56 And I think, you know, we can talk about physics project and so on, but I think the thing we realize is we kind of exist in a slice of all the possible computational irreducibility in the universe. We exist in a slice where there's a reasonable amount of predictability. And in a sense, as we try and construct these kind of higher levels of, of abstraction, symbolic representations and so on, what we're doing is, we're finding these lumps of reducibility that we can kind of attach ourselves to, and about which we can kind of have fairly simple narrative things to say. Because in principle, you know, I say, "What's gonna happen in the next few seconds?" You know, oh, there are these molecules moving around in the air in this room, and oh gosh, it's an incredibly complicated story, um, and that's a whole computationally irreducible thing, most of which I don't care about.

Narrator 15:44 Mm-hmm.

Stephen Wolfram 15:45 And most of it is, well, you know, the air's still gonna be here and nothing much is going to be different about it. And that's a kind of reducible fact about what is ultimately a- at an underlying level, a computationally irreducible process.

Narrator 15:59 (sighs) And, uh, life would not be possible if we didn't have a, uh, large number of such reducible pockets. Uh...

Stephen Wolfram 16:07 Yes.

Narrator 16:07 ... pockets amenable to, uh, reduction into something symbolic.

Stephen Wolfram 16:11 Yes, I think so.

Narrator 16:12 Okay.

Stephen Wolfram 16:12 I mean, life in, in the way that we experience it, that, I mean, you know, one might, you know, depending on what we mean by life, so to speak, the- the- the experience that we have of sort of consistent things happening in the world, the idea of space, for example, where there's, you know, we can just say, "You're here, you move there," it's kind of the same thing. It's still you in that different place, even though you're made of different atoms of space and so on. This is, this idea that it's, uh, that there's sort of this level of predictability of what's going on, that's us finding a slice of reducibility in what is underneath this computationally irreducible kind of system. And I think that's- that's sort of the- the thing which is actually my favorite discovery of the last few years, is the realization that it is sort of the interaction between the sort of underlying computational irreducibility and our nature as kind of observers who sort of have to key into computational reducibility.

Narrator 17:16 Mm-hmm.

Stephen Wolfram 17:16 That fact leads to the main laws of physics that we discovered in the 20th century. So this is, we talk about this in- in- in more detail, but this is a, uh... To me, it's kind of our nature as observers, the fact that we are computationally bounded observers, we don't get to follow all those little pieces of computational irreducibility. To stuff what is ab- out there in the world into our minds requires that we are looking at things that are reducible, we are compressing, kind of we- we're extracting just some essence, some kind of symbolic essence of what's the detail of what's going on in the world. That, together with one other condition that at first seems sort of trivial but isn't, which is that we believe we are persistent in time. That is-

Narrator 18:03 (laughs) Yes.

Stephen Wolfram 18:04 ... um, you know-

Narrator 18:05 So some sense of causality.

Stephen Wolfram 18:07 Here's the thing. At every moment, according to our theory, we're made of different atoms of space. At every moment, sort of the microscopic detail of what- what the universe is made of is being rewritten.

Narrator 18:20 Mm-hmm.

Stephen Wolfram 18:20 And that's... And in fact, the very fact that there's coherence between different parts of space is a consequence of the fact that there are all these little processes going on that kind of knit together the structure of space. It's kind of like if you wanted to have a fluid with a bunch of molecules in it. If those molecules weren't interacting, you wouldn't have this fluid that would pour and do all these kinds of things. It would just be sort of a free-floating collection of molecules.

Narrator 18:40 Mm-hmm.

Stephen Wolfram 18:40 So similarly, it is with space, that the fact that space is kind of knitted together as a consequence of all this activity in space, and the fact that kind of what- what we consist of, sort of this- this series of- of- of, uh... You know, we're- we're continually being rewritten.And the question is, why is it the case that we think of ourselves as being the same us through time? That's kind of a, a key assumption. I think it's a key aspect of what we see as sort of our consciousness, so to speak.

Lex Fridman 19:08 Mm-hmm.

Stephen Wolfram 19:09 Is that we have this kind of consistent thread of experience.

Lex Fridman 19:12 Well, isn't that just another limitation, uh, of our mind that we wanna reduce-

Stephen Wolfram 19:20 Yeah.

Lex Fridman 19:20 ...reality into some... Th- that kind of temporal-

Stephen Wolfram 19:24 Yes.

Lex Fridman 19:24 ...consistency is just a nice narrative-

Stephen Wolfram 19:27 Right.

Lex Fridman 19:27 ...to tell ourselves?

Stephen Wolfram 19:28 Well, the fact is, I think it's critical to the way we humans typically operate is that we have this single thread of experience. You know, if you, if you imagine sort of a, a mind where you have, you know, maybe that's what's happening in various kinds of minds that aren't working the same way other minds work, is that you're splitting into multiple threads of experience. It's also, it's also something where, you know, when you look at, I don't know, quantum mechanics for example, in the insides of quantum mechanics, it's splitting into many threads of experience. But in order for us humans to interact with it, you kinda have to, you have to knit all those different threads together so that we say, "Oh yeah, a definite thing happened," and now the next definite thing happens, and so on. And I think, you know, sort of inside, uh, it- it's, it's sort of interesting to try and imagine, what's it like to have kind of these, uh, fundamentally multiple threads of experience going on? I mean, right now, different human minds have different threads of experience. We just have a bunch of minds that are interacting with each other. But we don't have a, a... You know, within each mind, there's a single thread, and that's a, that is indeed a simplification. I think it's a, it's a thing, you know, the general computational system does not have that simplification.

Lex Fridman 20:41 Mm-hmm.

Stephen Wolfram 20:41 And, um, it's one of the things, you know, I, I, people often seem to think that, you know, consciousness is the highest level of kind of things that can happen in the universe, so to speak, but I think that's not true. I think it's actually a, a, a specialization in which, among other things, you have this idea of a single thread of experience, which is not a general feature of anything that could kind of computationally happen in the universe.

Lex Fridman 21:04 So it's a feature of a computationally limited system that's only able to, uh, observe, uh, reducible pockets. So-

Stephen Wolfram 21:13 Yeah.

Lex Fridman 21:14 ...so, I mean, this word observer, it, it means something in quantum mechanics, it means something in a lot of places. It means something to us humans-

Stephen Wolfram 21:24 Right.

Lex Fridman 21:24 ...as conscious beings, so what, what's the importance of the observer? What is the observer and what's the importance of the observer in the computational universe?

Stephen Wolfram 21:33 So this question of, what is an observer, what's the general idea of an observer, is actually one of my next projects, which got somewhat derailed by the, the current sort of AI mania. But, um-

Lex Fridman 21:43 Is there a connection there or is that, uh, do you, do you think the observer is primarily a physics phenomena?

Stephen Wolfram 21:48 Is it related to the whole AI thing?

Lex Fridman 21:50 Yes.

Stephen Wolfram 21:50 Yes, it is related. So one question is, what is a general observer?

Lex Fridman 21:53 Mm-hmm.

Stephen Wolfram 21:54 So, you know, we know, we have an idea of what is a general computational system. We think about Turing machines, we think about other models of computation. There's a question, what is a general model of an observer? And the- there's kind of observers like us, which is kind of the observers we're interested in.

Lex Fridman 22:10 Mm.

Stephen Wolfram 22:10 You know, we could imagine an alien observer that deals with computational irreducibility and it has a mind that's utterly different from ours and, and completely incoherent with what, what we're like. But the fact is that, that, you know, if we are talking about observers like us, the, one of the key things is this idea of kind of taking all the detail of the world and being able to stuff it into a mind.

Lex Fridman 22:32 Mm.

Stephen Wolfram 22:32 Being able to take all the detail and kind of, you know, extract out of it a smaller set of, of kind of degrees of freedom, a smaller number of, of elements that will sort of fit in our minds. And I think this, this question, so I've been interested in trying to characterize, what is the general observer? And the general observer is, I think, in part, there are many... L- let me give an example of a-

Lex Fridman 22:56 Mm-hmm.

Stephen Wolfram 22:57 ...you know, you have a gas, it's got a bunch of molecules bouncing around, and the thing you're measuring about the gas is its pressure.

Lex Fridman 23:03 Mm-hmm.

Stephen Wolfram 23:04 And the only thing you as an observer care about is pressure. And that means you have a piston on the side of this box and the piston is being pushed by the gas, and there are many, many different ways that molecules can hit that piston.

Lex Fridman 23:15 Mm.

Stephen Wolfram 23:16 But all that matters is the kind of aggregate of all those molecular impacts, 'cause that's what determines pressure.

Lex Fridman 23:22 Mm-hmm.

Stephen Wolfram 23:22 So there's a huge number of different configurations of the gas which are all equivalent. So I think one key aspect of observers is this equivalencing of many different configurations of a system, saying, "All I care about is this aggregate feature, all I care about is this, this overall thing." And that's, that's sort of one, one aspect. And w- we see that l- in lots of different... Uh, again, it's the same story over and over again, that there's, there's a lot of detail in the world, but what we are extracting from it is something, a sort of a thin, uh, a thin summary of that, of that detail.

Lex Fridman 23:53 I- is that thin summary nevertheless true? Is, can it be, uh, a crappy approximation-

Stephen Wolfram 24:01 Sure.

Lex Fridman 24:01 ...that on average is, is correct? I mean, if we look at the observer that's the human mind, it seems like there's a lot of very, um, as represented by natural language for example, there's a lot of really crappy approximation.

Stephen Wolfram 24:12 Sure.

Lex Fridman 24:13 And that could be maybe a feature of it.

Stephen Wolfram 24:15 Well, yes, but-

Lex Fridman 24:16 That there's ambiguity.

Stephen Wolfram 24:17 Right, right. You don't know, you know, it could be the case. You- you're just measuring the aggregate impacts of these molecules. But there is some tiny, tiny probability that the molecules will arrange themselves in some really funky way, and that just measuring that average isn't going to be the main point.

Lex Fridman 24:33 Yeah.

Stephen Wolfram 24:33 By the way, an awful lot of science is very confused about this, because, you know, you look at, you look at papers, and people are really keen, they draw this curve and they have these, you know, these bars on the curve and things.

Lex Fridman 24:44 Yeah.

Stephen Wolfram 24:45 And it's just this curve....and it's this one thing, and it's supposed to represent some system that has all kinds of details in it. And this is a way that lots of science has gotten wrong, because people say... I remember years ago, I was studying snowflake growth. The-

Lex Fridman 24:59 (laughs)

Stephen Wolfram 24:59 You know, you have a snowflake and it's growing. It has all these arms, it's doing complicated things. But there was a literature on this stuff, and it talked about, you know, what's the rate of snowflake growth? And, you know, it- it got pretty good answers for the rate of the growth of the snowflake. And then I looked at it more carefully, and- and they had these nice curves of, you know, snowflake growth rates and so on. I looked at it more carefully and I realized, according to their models, the snowflake would be spherical.

Lex Fridman 25:22 Mm-hmm.

Stephen Wolfram 25:23 And so, they got the growth rate right, but the detail was just utterly wrong. And, you know, the... not only the detail, the- the whole thing was- was not capturing, you know, it was capturing this aspect of the system that was, in a sense, missing the main point of what was going on, and-

Lex Fridman 25:39 What is the geometric, uh, shape of a snowflake?

Stephen Wolfram 25:43 Snowflakes start in, in the phase of water that's relevant-

Lex Fridman 25:47 Yeah.

Stephen Wolfram 25:47 ...to formation of snowflakes. It's a phase of ice which starts with a hexagonal arrangement of- of water molecules.

Lex Fridman 25:53 Mm-hmm.

Stephen Wolfram 25:54 And so it starts off growing as a hexagonal plate, and then what happens is, it's-

Lex Fridman 25:58 Just a plate. Oh, oh, versus sphere. Sphere versus plate.

Stephen Wolfram 26:00 Well, no, no, but it's- it's much more than that.

Lex Fridman 26:02 Yeah.

Stephen Wolfram 26:02 I mean, snowflakes are fluffy, you know? Typical snowflakes have little- little dendritic arms.

Lex Fridman 26:06 Okay. Yeah, yeah, yeah.

Stephen Wolfram 26:07 And- and what actually happens is, it's kind of- kind of cool because you can make these very simple discrete models with cellular automata and things that- that figure this out. You start off with this, you know, hexagonal thing, and then the places... It- it starts to grow little arms, and every time a little piece of ice a- adds itself to the snowflake-

Lex Fridman 26:27 Mm-hmm.

Stephen Wolfram 26:27 ...the fact that that ice condensed from the water vapor m- heats the snowflake up locally, and so it makes it less likely for, uh, for another piece of ice to accumulate right nearby. So this leads to a kind of growth inhibition. So you grow an arm and y- it- it is a- a separated arm because right around the arm, it got a little bit hot and it didn't add more ice there. So what happens as it grows, you have a hexagon, it grows out arms. The arms grow arms, and then the arms grow arms grow arms, and eventually... Actually, it's kind of cool because it- it actually fills in another hexagon, a bigger hexagon. And when I first looked at this, you know, I had a very simple model for this. I realized, you know, when it fills in that hexagon, it actually leaves some holes behind. So I thought, "Well, you know, that... Is that really right?" So I look at these pictures of snowflakes, and sure enough, they have these little holes in them that are kind of scars of the way that these arms grew out. Um, so that-

Lex Fridman 27:19 Mm-hmm. So you can't fill in, backfill holes. So it just keeps going up.

Stephen Wolfram 27:24 They don't backfill. Yeah, they don't backfill.

Lex Fridman 27:25 A- and presumably, there's a limitation on how big. Like, you can't arbitrarily grow.

Stephen Wolfram 27:31 I'm not sure. I mean, the thing falls through the- the... I mean, I think it does... you know, it hits-

Lex Fridman 27:35 That's-

Stephen Wolfram 27:35 ...the ground at some point.

Lex Fridman 27:36 Oh, I see.

Stephen Wolfram 27:37 I think you can grow. I th- I think you can grow in the lab. I think you can grow pretty big ones. I think you can grow many- many iterations of this kind of... It goes from hexagon, it grows out arms, it turns back, it fills back into a hexagon. It grows more arms again.

Lex Fridman 27:49 In 3D?

Stephen Wolfram 27:50 No, it's flat usually.

Lex Fridman 27:52 S- well, why is it flat? Why doesn't it, uh, span out? Okay, okay, okay, wait a minute. You said it's fluffy, and fluffy is a three-dimensional property, no? Or...

Stephen Wolfram 28:00 No. It's- it's fluffy... Snow is...

Lex Fridman 28:02 (laughs)

Stephen Wolfram 28:02 Okay, so, you know, what makes... We're really, uh, we're really in a- in a detail here-

Lex Fridman 28:06 I like this. Let's go there.

Stephen Wolfram 28:07 ...but- but, uh-

Lex Fridman 28:08 Does multiple snowflakes become fluffy? A single snowflake is not fluffy, is it?

Stephen Wolfram 28:12 No, no. A single snowflake is fluffy.

Lex Fridman 28:14 Mm-hmm.

Stephen Wolfram 28:14 And what happens is, you know, if- if you have snow that is just pure hexagons-

Lex Fridman 28:19 Mm-hmm.

Stephen Wolfram 28:20 ...they- they can, you know, they- they fit together pretty well. It's not... It doesn't- it doesn't make... It doesn't have a lot of air in it.

Lex Fridman 28:25 Mm-hmm.

Stephen Wolfram 28:26 And they can also slide against each other pretty easily, and so the snow can be pretty... you know, it can... I think avalanches happen sometimes when- when the things tend to be these, you know, hexagonal plates and it kind of slides. But then when the thing has all these arms that have grown out, it's not... they don't fit together very well, and that's why the snow has lots of air in it. And if you look at one of these snowflakes, you know, if you catch one, you'll see it has these little arms. And people... Actually, people often say, "You know, no two snowflakes are alike." Um, that's mostly because as a snowflake grows, they do grow pretty consistently with these different arms and so on, but you capture them at different times as they... you know, they fell through, th- through the air in a different way. You'll catch this one at this stage, and as it goes through different stages, they look really different. And so that's why, you know, kind of looks like no two snowflakes are alike because you caught them at different- at different times.

Lex Fridman 29:18 So the rules under which they grow are the same.

Stephen Wolfram 29:20 Yes.

Lex Fridman 29:20 It's just the timing is-

Stephen Wolfram 29:22 Yes.

Lex Fridman 29:22 Okay, so the point is science is not able to, uh, describe the full complexity of snowflake growth.

Stephen Wolfram 29:29 Well, science... If you- if you do what people might often do, which is say, "Okay, m- let's make it scientific. Let's turn it into one number."

Lex Fridman 29:37 Mm-hmm.

Stephen Wolfram 29:38 And that one number is kind of the growth rate of the arms or some such other thing, that fails to capture sort of the detail of what's going on inside the system. And that's, in a sense, a big challenge for science, is how do you extract from the natural world, for example, those aspects of it that you are interested in talking about?

Lex Fridman 29:57 Yeah.

Stephen Wolfram 29:57 Now, you might just say, "I don't really care about the fluffiness of the snowflakes. All I care about is the growth rate of the arms," in which case, you know, you have... you can have a good model without knowing anything about the fluffiness. Um, but the fact is, as a practical... you know, when... if you- if you say, "What's the- what is the most obvious feature of a snowflake? Oh, that it has this complicated shape," well, then you've got a different story about what you model. I mean, this- this is one of the features of- of sort of modeling in science, that, you know, what is a model? A model is some way...... of reducing the actuality of the world to something where you can readily sort of give a narrative for what's happening, where you can basically make some kind of abstraction of what's happening and answer questions that you care about answering. If you wanted to answer all possible questions about the system, you'd have to have the whole system-

Lex Fridman 30:44 Mm-hmm.

Stephen Wolfram 30:44 ... because you might care about this particular molecule, where did it go? And, you know, your model, which is some big abstraction of that, has nothing to say about that. So, you know, one of the things that's- that's often confusing in science is people will have, "I've got a model," somebody says. Somebody else will say, "I don't believe in your model because it doesn't capture the feature of the system that I care about." You know, there's always this controversy about, you know, is the- is it a correct model? Well, no model is a, except for the actual system itself, is a correct model in the sense that it captures everything. The question is, does it capture what you care about capturing? Sometimes that's ultimately defined by what you're going to build technology out of, things like this. The one counter-example to this is, if you think you're modeling the whole universe all the way down, then there is a notion of a correct model. But even that is more complicated because it depends on kind of how observers sample things and so on. That's a, that's a separate story. But at least at the first level to say, you know, this thing about, "Oh, it's an approximation, you're capturing one aspect, you're not capturing other aspects," when you really think you have a complete model for the whole universe, you better be capturing ultimately everything, even though to actually run that model is impossible because of computational irreducibility. The only- the only thing that successfully runs that model is the actual running of the universe.

Lex Fridman 32:01 Is the universe itself, but... Okay, so what you care about is an interesting concept. So that's a- that's a human concept, so that's what you're doing with, uh, Wolfram|Alpha and Wolfram Language, is you're trying to come up with symbolic representations-

Stephen Wolfram 32:17 Yes.

Lex Fridman 32:17 ... as simple as possible. Uh, so a model that's as simple as possible that fully captures stuff we care about.

Stephen Wolfram 32:26 Yes. So I mean, for example, you know, we could... We'll have a- a thing about, you know, data about movies, let's say. We could be describing every individual pixel in every movie and so on, but that's not the level that people care about, and it's... Yes, this is a- I mean... And- and that level that people care about is somewhat related to what's described in natural language.

Lex Fridman 32:48 Mm-hmm.

Stephen Wolfram 32:49 But what- what we're trying to do is to find a way to sort of represent precisely so you can compute things. See- see, one thing when you say- you give a piece of natural language. The question is, you feed it to a computer, you say, "Does the computer understand this natural language?" Well, you know, the computer processes it in some way, it does this, maybe it can make a continuation of the natural language, you know, maybe it can go on from the prompt and say what it's gonna say. You say, "Does it really understand it?" Hard to know. But for... In this kind of computational world, there is a very definite definition of, "Does it understand?" Which is, could it be turned into this symbolic computational thing from which you can compute all kinds of consequences?

Lex Fridman 33:31 Mm-hmm.

Stephen Wolfram 33:31 And that's the- that's the sense in which one has sort of a target for the understanding of natural language, and that's kind of our goal, is to have as much as possible about the world that can be computed in a- in a reasonable way, so to speak, be able to be sort of captured by this kind of computational language, that's- that's kind of the goal. And- and I think for us humans, the- the main thing that's important is, as we formalize what we're talking about, it gives us a way of kind of building a structure where we can sort of build this tower of consequences of things. So if we're just saying, "Well, let's talk about it in natural language," it doesn't really give us some hard foundation that lets us, you know, build step by step to work something out. I mean, it's kind of like what happens in- in math. If we were just sort of vaguely talking about math, but didn't have the kind of full structure of math and all that kind of thing, we wouldn't be able to build this kind of big tower of consequences. And so, you know, in a- in a sense, what we're trying to do with the whole computational language effort is to make a formalism for describing the world that makes it possible to kind of build this- this tower of consequences.

Lex Fridman 34:38 Well, can you talk about this dance between natural language and Wolfram Language? So there's this gigantic thing we call the internet where people post memes and diary-type thoughts and very important-sounding articles and all of that, that makes up the training data set for GPT, and then there's, uh, Wolfram Language. How can you map from the natural language of the internet to the Wolfram Language? Is there a manual? Is there an automated way of doing that as-

Stephen Wolfram 35:13 Right.

Lex Fridman 35:13 ... we look into the future?

Stephen Wolfram 35:15 Well, so Wolfram|Alpha, what it does, its kind of front end is turning natural language into computational language.

Lex Fridman 35:24 Right. What you mean by that is, there's a prompt, you ask a question, "What is the capital of some country?"

Stephen Wolfram 35:29 Yeah, right. And it- and it turns into, you know, what's the distance between, you know, Chicago and London, or something?

Lex Fridman 35:35 Yeah.

Stephen Wolfram 35:35 And that will turn into, you know, geo-distance of entity, city, you know, et cetera, et cetera, et cetera. Each one of those things is very- is very well-defined. We know, you know, given that it's the entity, city, Chicago, et cetera, et cetera, et cetera, you know, Illinois, United States, you know, we know the geolocation of that, we know its population, we know all kinds of things about it-

Lex Fridman 35:55 Mm-hmm.

Stephen Wolfram 35:56 ... which we have, you know, curated that data to be able to- to know that with some degree of certainty, so to speak.

Lex Fridman 36:02 Mm-hmm.

Stephen Wolfram 36:03 And then- then we can compute things from this, and that's- that's kind of the, um... Yeah, that- that's- that's- that's the idea.

Lex Fridman 36:11 But then something like GPT, large language models, do they allow you to, uh, make that conversion much more powerful?

Stephen Wolfram 36:19 Okay, so it's an interesting thing, which we still don't know everything about, okay?

Lex Fridman 36:24 Mm-hmm.

Stephen Wolfram 36:24 The, um... I mean, this question of going from natural language to computational language-

Lex Fridman 36:28 Yes.

Stephen Wolfram 36:29 ... in Wolfram|Alpha, we've now... You know, Wolfram|Alpha's been out and about for, what, 13 and a half years now and, you know, we've achieved......I don't know what it is, 98%, 99% success on queries that get put into it. Now obviously, there's a sort of feedback loop 'cause the things that work are things people go on putting into it. So-

Lex Fridman 36:48 Yeah.

Stephen Wolfram 36:48 ...that, that, uh... But, you know, we've, we've got to a very high success rate of the, the little fragments of natural language that put, people put in. You know, questions, math calculations, chemistry calculations, whatever it is. You know, we can, we can... We, we do very well at that, turning those things into computational language. Now, I... From the very beginning of Wolfram|Alpha, I thought about, for example, uh, writing code with natural language. Um, in fact I had, um, I, I was just looking at this recently, I had a post that I wrote in 2010, 2011 called something like, "Programming with natural language is actually going to work."

Lex Fridman 37:23 Mm-hmm.

Stephen Wolfram 37:23 Okay? And so, you know, we had done a bunch of experiments using methods that were, yeah, a little bit... Some of them a little bit machine learning-like, but certainly not the same, you know, the same kind of idea of vast training data and so on-

Lex Fridman 37:37 Mm-hmm.

Stephen Wolfram 37:37 ...that, uh, is the story of large language models. Actually, I know that, that post, a piece of utter trivia, but that, that post, um... Uh, Steve Jobs forwarded that post around to all kinds of people at Apple-

Lex Fridman 37:48 (laughs)

Stephen Wolfram 37:48 ... and he, you know, that was... 'Cause he never really liked programming languages, so he was very happy to see the idea that, that, that, that you could get rid of this kind of layer of kind of engineering-like structure. Uh, he would have liked, you know, I think what's happening now, because it really is the case that you can... You know, this idea that you have to kind of learn how the computer works to use a programming language is something that is, I think, a, a thing that... You know, just like you had to learn the details of the op codes to know how assembly language worked, and so on. It's kind of a thing that's, that's, that's a limited time horizon. But-

Lex Fridman 38:22 Yeah.

Stephen Wolfram 38:23 ...but kind of the, the... You know, so this idea of how elaborate can you make kind of the prompt, how elaborate can you make the natural language and abstract from it computational language? It's a very interesting question. And, you know, what ChatGPT, you know, GPT-4 and so on can do is pretty good. Um, it isn't... It's a very interesting process. I mean, I'm still trying to understand this workflow. We've been working out a lot of tooling around this workflow, but it's pretty interesting.

Lex Fridman 38:53 So natural language to computational language.

Stephen Wolfram 38:56 Right. And that-

Lex Fridman 38:56 A process.

Stephen Wolfram 38:57 Right.

Lex Fridman 38:57 Especially if it's conversation, like dialogue. It's like multiple queries kinda thing.

Stephen Wolfram 39:02 Yeah, right. There so many things that are really interesting-

Lex Fridman 39:05 Yeah.

Stephen Wolfram 39:05 ...that, that work and so on. So, first thing is, can you just walk up to the computer and expect to sort of specify a computation? What one realizes is humans have to have some idea of kind of this way of thinking about things computationally.

Lex Fridman 39:19 Mm-hmm.

Stephen Wolfram 39:19 Without that, you're kind of out of luck, because you just have no idea what you're going to walk up to a computer... I remember when I... I should tell a silly story about myself. The very first computer I saw-

Lex Fridman 39:28 Mm-hmm.

Stephen Wolfram 39:28 ...which was when I was 10 years old, and it was a big mainframe computer and so on, and I didn't really understand what computers did. And it's like, somebody's showing me this computer and it's like, uh, you know, "Can the computer work out the weight of a dinosaur?"

Lex Fridman 39:41 Mm-hmm.

Stephen Wolfram 39:41 It's like, that isn't a sensible thing to ask.

Lex Fridman 39:43 Mm-hmm.

Stephen Wolfram 39:44 That's kind of... You know, you have to give it... That's not what computers do. I mean, in Wolfram|Alpha for example, you could say, "What's the typical weight of a stegosaurus?" And it will give you some answer-

Lex Fridman 39:52 Mm-hmm.

Stephen Wolfram 39:52 ... but that's a very different kind of thing from what one thinks of computers as doing. And so the, the kind of the, the question of... Uh, you know, first thing is people have to have an idea of what, what computation is about. Um, you know, I think it's a very... You know, for education, that is the key thing, is kind of this, this, uh, this notion... Not computer science, not sort of the details of programming, but just this idea of how do you think about the world computationally? Computation... Thinking about the world computationally is kind of this formal way of thinking about the world. We've had other ones, like logic was a formal way, you know, as a way of sort of abstracting and formalizing some aspects of the world. Mathematics is another one. Computation is this very broad way of sort of formalizing the way we think about the world. And the thing that's, that's cool about computation is, if we can successfully formalize things in terms of computation, computers can help us figure out what the consequences are. It's not like you formalized it with math, well, that's nice, but now you have to, if you're, you know, not using a computer to do the math, you have to go work out a bunch of stuff yourself. So, I think... But this idea... Uh, let's see. I mean, the, the... You know, we're trying to take kind of the... Uh, we're talking about sort of natural language and its relationship to computational language. The, uh, the thing, the sort of the typical workflow I think is, first, human has to have some kind of idea of what they're trying to do. That if, if it's something that they want to sort of build a tower of, of capabilities on, something that they want to sort of formalize and make computational. So then, human can type something into, you know, some LLM system and, uh, uh, sort of say vaguely what they want, in sort of computational terms. Then, it does pretty well at synthesizing Wolfram Language code, and it'll probably do better in the future, 'cause we've got a huge number of examples of, of natural language input, together with the Wolfram Language translation of that.

Lex Fridman 41:50 Mm-hmm.

Stephen Wolfram 41:51 So it's kind of a, a, um, uh... You know, that, that's a thing where you can... Kind of extrapolating from all those examples, uh, makes it easier to do that, that task.

Lex Fridman 42:02 Is the prompter test but also kind of debugging the Wolfram Language code? Or, is your hope to not do that debugging?

Stephen Wolfram 42:08 Oh, no, no, no. I mean, so, so there are many steps here.

Lex Fridman 42:11 Yeah.

Stephen Wolfram 42:11 Okay? So first, the first thing is you type natural language. It generates Wolfram Language code.

Lex Fridman 42:16 Do you have examples, by the way? Do you, do you have an example that... Is it, is it the- the dinosaur example? Do you have an example that jumps to mind, that we should be thinking about? Some dumb example?

Stephen Wolfram 42:24 It's like, take my heart rate data-

Lex Fridman 42:27 Mm-hmm.

Stephen Wolfram 42:27 ...and, uh, you know, figure out whether I, uh, you know, make a moving average every seven days or something, and work out what the, um... And, and make a plot of the results.

Lex Fridman 42:39 Mm-hmm.

Stephen Wolfram 42:39 Okay? So that's a thing which is, you know-... about two thirds of a line of Wolfram Language code.

Lex Fridman 42:45 Mm-hmm.

Stephen Wolfram 42:45 I mean it's, you know, list plot of moving average of some data bin or something of the, of the data, and then you'll get the result. Um, and you know, the vague thing that I was just saying in natural language could, would almost certainly correctly turn into that very simple piece of Wolfram Language code.

Lex Fridman 43:03 Mm-hmm. So you start mumbling about heart rate-

Stephen Wolfram 43:06 Yeah, and, and-

Lex Fridman 43:07 ... and then kinda, you know, you arrive at the moving average kind of idea.

Stephen Wolfram 43:12 Right. You say average over seven days, maybe it'll figure out that that's a moving, you know, that-

Lex Fridman 43:16 Yeah.

Stephen Wolfram 43:16 ... that can be encapsulated as this moving average idea, I'm not sure. But then the typical workflow that I'm seeing is you generate this piece of Wolfram Language code. It's pretty small usually. Um, it's, uh... And if it isn't small, it probably isn't right. But, um, (laughs) you know, if it's, it's pretty small. And, you know, Wolfram Language is... One of the ideas of Wolfram Language is it's a language that humans can read. It's not a language which... You know, programming languages tend to be this one-way story of humans write 'em and computers execute from them. Wolfram Language is intended to be something which is sort of like math notation, something where, you know, humans write it and humans are supposed to read it as well. And so kind of the workflow that's emerging is kind of this, this human mumbles some things, you know, large language model produces a fragment of Wolfram Language code. Then you look at that, you say, "Yeah, that looks..." Well, typically you just run it first and you see does it produce the right thing? You look at what it produces, you might say, "That's obviously crazy." You look at the code, you see, "I see why it's crazy." You fix it. If you really care about the result and you really wanna make sure it's right, you better look at that code and understand it because that's the way you have the sort of checkpoint of, "Did it really do what I expected it to do?" Now, you go beyond that, I mean it's, it's, it's... You know, what we find is, for example, let's say the code does the wrong thing.

Lex Fridman 44:39 Mm-hmm.

Stephen Wolfram 44:39 Then you can often say to the large language model, "Can you adjust this to do this?"

Lex Fridman 44:44 (laughs)

Stephen Wolfram 44:44 And it's pretty good at doing that.

Lex Fridman 44:46 Interesting. So you're using the output of the code to give you hints about the, the, the function of the code. So you're debugging-

Stephen Wolfram 44:58 Yes.

Lex Fridman 44:59 ... on, based on the output of the code-

Stephen Wolfram 45:00 And, and by the way-

Lex Fridman 45:00 ... not the code itself.

Stephen Wolfram 45:01 Right. The plug-in that we have, uh, the, the, you know, for ChatGPT, it does that routinely. You know, it will send the thing in, it will get a result, it will discover, the LLM will discover itself that the result is not plausible.

Lex Fridman 45:14 Mm-hmm.

Stephen Wolfram 45:14 And it will go back and say, "Oh, I'm sorry." It's very polite and it, it, you know, (laughs) it goes back and says, "I'll rewrite that piece of code." And then it will try it again and get the result. The other thing that's pretty interesting is when you're just running... So one of the new concepts that we have, we invented this whole idea of notebooks back 36 years ago now.

Lex Fridman 45:32 Mm-hmm.

Stephen Wolfram 45:33 And so now there's the question of sort of how do you combine this idea of notebooks where you have, you know, text and code and output, how do you combine that with the notion of, of chat and so on? And there's some really interesting things there. Like, for example, a very typical thing now is we have these, these notebooks where as soon as the... If, if the thing produces errors, if the... You know, run this code and it produces messages and so on, the, the, the LLM automatically not only looks at those messages, it can also see all kinds of internal information about stack traces and things like this. And it can then... It does a remarkably good job of guessing what's wrong and telling you. So in other words, it's, it's looking at things. It's sort of interesting. It's kind of a typical sort of AI-ish thing that it's able to have more sensory data than we humans are able to have 'cause it's able to look at a bunch of stuff that we humans would kind of glaze over looking at. And it's able to then come up with, "Oh, this is the explanation of what's happening."

Lex Fridman 46:32 And, and what is the data? The stack trace, the, the code you've written previously, the natural language you've written previously?

Stephen Wolfram 46:36 Yeah, it's also what's happening is one of the things that's, uh, is, is for example, when there's these messages, there's documentation about these messages, there's examples of where the messages have occurred otherwise.

Lex Fridman 46:46 Oh, nice.

Stephen Wolfram 46:46 All these kinds of things. The other thing that's really amusing-

Lex Fridman 46:49 Yeah.

Stephen Wolfram 46:49 ... with this is when it makes a mistake, one of the things that's in our prompt when the code doesn't work is read the documentation.

Lex Fridman 46:56 (laughs)

Stephen Wolfram 46:57 And we have a, you know, another piece of the plug-in (laughs) -

Lex Fridman 47:00 Yeah.

Stephen Wolfram 47:00 ... that lets it read documentation. And that again, is very, very useful 'cause it, it will, you know, it will figure out, sometimes it'll get... It'll make up the name of some option for some function that doesn't really exist, read the documentation. It'll have, you know, some wrong structure for the function and so on. It's, um... That's a powerful thing. I mean, the thing that, you know, I've realized is we built this language over the course of all these years to be nice and coherent and consistent and so on, so it's easy for humans to understand. Turns out there was a side effect that I didn't anticipate, which is it makes it easier for AIs to understand.

Lex Fridman 47:34 So it's almost like another natural language, but-

Stephen Wolfram 47:37 Yeah.

Lex Fridman 47:37 So, so Wolfram-

Stephen Wolfram 47:38 Yeah.

Lex Fridman 47:38 ... Language is a kind of foreign language.

Stephen Wolfram 47:40 Yes. Yes.

Lex Fridman 47:41 (laughs) You have a lineup, English, French, Japanese, Wolfram Language.

Stephen Wolfram 47:47 Yeah.

Lex Fridman 47:47 And then, uh, I don't know, Spanish, and then-

Stephen Wolfram 47:50 And-

Lex Fridman 47:50 ... the system is not gonna notice.

Stephen Wolfram 47:52 Well-

Lex Fridman 47:53 Hopefully.

Stephen Wolfram 47:53 Yes. I mean, maybe. You know, that's an interesting question 'cause, uh, it really depends on what I see as being a, a, an important piece of fundamental science that basically just jumped out at us with ChatGPT.

Stephen Wolfram 48:06 Um, 'cause I think, you know, the, the real question is why does ChatGPT work? How is it possible to encapsulate, you know, to successfully reproduce all these kinds of things in natural language, um, you know, with a, you know, a comparatively small, he says, you know, couple hundred billion, you know, weights of neural net and so on. And I think that, uh, you know, that, that relates to kind of a fundamental fact about language, which, uh, you know, the, the main, the main thing is that I think there's a structure to language that we haven't kind of really explored very well as kind of this semantic grammar I- I'm talking about, about, um, about language. I mean, we, we kind of know that when we, when we set up human language-... we know that it has certain regularities. We know that it has a certain grammatical structure, you know, noun followed by verb followed by noun, adjectives, et cetera, et cetera, et cetera. That's its kind of grammatical structure. But I think the thing that ChatGPT is showing us is that there's an additional kind of regularity to language, which has to do with the meaning of the language beyond just this pure, you know, part of speech combination type of thing. And I, I think the, uh, the, the kind of, the, the one example of that, that we've had in the past is logic. And, you know, I, I think my, my sort of, uh, uh, uh, kind of picture of how was logic invented, how was logic discovered, uh, it really was the thing that was discovered in its original conception. It was discovered, presumably by Aristotle, who kind of listened to a bunch of people, orators, you know, giving speeches. And this one made sense, that one doesn't make sense, this one... And, you know, you see these patterns of, you know, if the, uh, you know, I don't know what, you know, if the, uh, if the Persians do this, then this does that-

Lex Fridman 49:56 Mm-hmm.

Stephen Wolfram 49:57 ... et cetera, et cetera, et cetera. And what, what Aristotle realized is there's a structure to those sentences, there's a structure to that rhetoric that doesn't matter whether it's the Persians and the Greeks, or whether it's the cats and the dogs.

Lex Fridman 50:09 Mm-hmm.

Stephen Wolfram 50:10 It's just, you know, P and Q. You can abstract from this, the, the details of these particular sentences. You can lift out this kind of formal structure, and that's what logic is.

Lex Fridman 50:21 Th- that's a heck of a discovery, by the way, logic. You're making me realize now.

Stephen Wolfram 50:25 Yeah.

Lex Fridman 50:25 It's not obvious.

Stephen Wolfram 50:27 The fact that there is an abstraction from natural language that has where you can fill in any word you want-

Lex Fridman 50:34 Yeah.

Stephen Wolfram 50:35 ... is a very interesting discovery. Now, it took a long time to mature. I mean, Aristotle had this idea of syllogistic logic, where there were these particular patterns of how you could argue things, so to speak.

Lex Fridman 50:46 Mm-hmm.

Stephen Wolfram 50:47 And, you know, in the Middle Ages, part of education was you memorized the syllogisms. I forget how many there were, uh, 15 of them or something. And they all had names, they all had mnemonics. Like, I think Barbara and Celarent were two of the, the mnemonics for the, the syllogisms. And people would kind of, "This is a valid argument 'cause it follows the Barbara syllogism," so to speak. And, and it took until 1830, um, you know, with, uh, George Boole to kind of get beyond that and kind of see that there was a, a level of abstraction that was beyond the, this particular template of a sentence, so to speak. Um, and that's... You know, and what, what's interesting there is, in a sense, you know, G- y- you know, ChatGPT is operating at the Aristotelian level. It's essentially dealing with templates of sentences. By the time you get to Boole and Boolean algebra and this idea of, you know, you can have arbitrary depth-nested collections of ands and ors and nots, and you can resolve what they mean, um, that's the kind of thing that's a computation story. That's, you know, you've gone beyond the pure sort of templates of natural language to something which is an arbitrarily deep computation. But the thing that I think we realized from, from ChatGPT is, you know, Aristotle stopped too quickly.

Lex Fridman 51:59 Mm-hmm.

Stephen Wolfram 51:59 And there was more that you could have lifted out of language as formal structures. And I think there's, you know, in a sense, we've captured some of that in, in, you know, some of what, what is in language that there's, there's a, there's a lot of kind of little calculi, little algebras of, of what you can say, what language talks about. I mean, whether it's, I don't know, if you say, "Uh, I go from place A to place B, place B to place C," then I know I've gone from place A to place C. If A is a friend of B and B is a friend of C, it doesn't necessarily follow that A is a friend of C. These are things that are, uh, you know, there, there are... If, if you go from, from place A to place B, place B to place C, it doesn't matter how you went. Like logic, it doesn't matter whether you flew there, walked there, swam there, whatever, you still... This transitivity of, of where you go is still valid. And there are, there are many kinds of, kind of features, I think, of the way the world works, uh, that are captured in these aspects of, of language, so to speak. And I think what, what ChatGPT effectively has found, just like it discovered logic, you know, people are really surprised it can do these, these logical inferences, it discovered logic the same way Aristotle discovered logic, by looking at a lot of sentences effectively and noticing the patterns in those sentences.

Lex Fridman 53:20 But it feels like it's discovering something much more complicated than logic. So this kind of semantic grammar, I think you wrote about this. Uh, maybe we can call it the laws of language, I believe you call, or which I like, the laws of thought.

Stephen Wolfram 53:36 Yes. That was the title that George Boole had for his, his Boolean algebra back in 1830. But yes, uh-

Lex Fridman 53:41 Laws of thought?

Stephen Wolfram 53:42 Yes. That was what he said.

Lex Fridman 53:43 Whew.

Stephen Wolfram 53:44 (laughs)

Lex Fridman 53:45 All right.

Stephen Wolfram 53:45 So he thought, he thought he nailed it with Boolean algebra.

Lex Fridman 53:48 Yeah.

Stephen Wolfram 53:49 There's more to it.

Lex Fridman 53:50 I think it's a good question, how much more-

Stephen Wolfram 53:53 Right.

Lex Fridman 53:53 ... is there to it? And it seems like one of the reasons, as you, uh, imply that the reason GPT works, ChatGPT works is that, uh, there's a finite num- ber of things to it.

Stephen Wolfram 54:06 Yeah. I mean, it's-

Lex Fridman 54:07 Like, it's discovering the laws... In some sense, GPT is discovering these laws of semantic grammar that underlies language.

Stephen Wolfram 54:15 Yes. And what, what's sort of interesting is, in the computational universe, there's a lot of other kinds of computation that you could do.

Lex Fridman 54:22 Mm-hmm.

Stephen Wolfram 54:22 They're just not ones that we humans have cared about and, and operate with. And that's probably because our brains are built in a certain way. And, you know, the neural nets of our brains are not that different in some sense from the neural nets of, of, uh, of a large language model. And that's kind of... And, and so when we think about, and, you know, maybe we can talk about this some more, but when we think about sort of what will AIs ultimately do, the answer is i- insofar as AIs are just doing computation, they can run off and do all these kinds of crazy computations. But the ones that we sort of have, have decided we care about are the, is this kind of very limited set.

Lex Fridman 55:01 Mm-hmm.That's where the reinforcement learning with human feedback seems to come in. The more the AI say the stuff that kinda interests us, the more we're impressed by it. So it can do a lot of interesting intelligent things, but we're only interested in the AI systems when they communicate human, in a humanlike way-

Stephen Wolfram 55:22 Yes.

Lex Fridman 55:23 ... about humanlike topics.

Stephen Wolfram 55:25 Yes. Well, it's, it's like technology. I mean, in a sense, the physical world provides all kinds of things, you know. There are all kinds of processes going on in physics. Only a limited set of those are ones that we capture and use for technology, because there are only a limited set where we say, you know, "This is a thing that we can sort of apply to the human purposes we currently care about." I mean, you might have said, okay, you pick up a piece of, of rock. You say, "Okay, this is a nice silicate. It contains all kinds of silicon. I don't care." Then you realize, "Oh, we could actually turn this into a, a, you know, semiconductor wafer and make a microprocessor out of it," and then we care a lot about it.

Lex Fridman 56:03 Yes.

Stephen Wolfram 56:03 Um, and it's, it's, you know, it's this thing about what do we... You know, uh, in the evolution of our civilization, what things do we identify as being things we care about? I mean, it's, it's like, you know, when, when there was a little announcement recently of the possibility of a high temperature superconductor-

Lex Fridman 56:18 Mm-hmm.

Stephen Wolfram 56:18 ... that involved, you know, the element lutetium, which, you know, generally nobody has cared about.

Lex Fridman 56:23 (laughs) Yes.

Stephen Wolfram 56:24 And, uh, you know, it, it's kind of, um... But suddenly, if there's this application that relates to kind of human purposes, we start to care a lot.

Lex Fridman 56:32 So given your thinking that GPT may have discovered inklings of laws of thought, do you think such laws exist? Can we linger on that? Do you-

Stephen Wolfram 56:43 Yeah.

Lex Fridman 56:44 What's your intuition here?

Stephen Wolfram 56:45 Oh, definitely. I mean, the fact is, look, the, the... Logic is but the first step. There are many other kinds of calculi about things that, uh, we consider, you know, the, the... A- about sort of things that happen in the world or things that are meaningful.

Lex Fridman 57:02 Well, how do you know logic's not the last step? You know what I mean? So like, wh-

Stephen Wolfram 57:06 Well, because we can plainly see that the thing... I mean, if, if you say, "Here's a sentence that is syntactically correct," okay? You look at it and you're like, you know, "The happy electron," you know, "ate..." I don't know, what? Some- something that it just, it... You look at it and it's like, "This is meaningless." It's just a bunch of words. It's syntactically correct. The nouns and the verbs are in the right place, but it just doesn't mean anything. Um, and so there clearly, uh, is some rule that... There are rules that determine when a sentence is... has the potential to be meaningful that go beyond the pure parts of speech syntax. And so the question is, what are those rules and are there fi- fairly finite set of those rules? My guess is that there's a fairly finite set of those rules. And they... You know, once you have those rules, you have a kind of a construction kit. Just like the s- the rules of sy- syntactic grammar give you a construction kit for making syntactically correct sentences. So, you can also have a construction kit for making semantically correct sentences. Those sentences may not be realized in the world. I mean, I think, you know, "The elephant flew to the moon."

Lex Fridman 58:16 Yeah.

Stephen Wolfram 58:17 A, a, a syntactic, uh, uh, a semantically... You know, we know, we have an idea. If I say that to you, you kind of know what that means. But the fact is, it hasn't been realized in the world, so to speak.

Lex Fridman 58:28 So semantically correct, perhaps is things that can be imagined with a human mind. No. Uh, things that are consistent with both our imagination and our understanding of physical reality. I don't... W- w- what, w- what... (laughs)

Stephen Wolfram 58:43 Yeah, it's a good question. I mean, it's a good question.

Lex Fridman 58:46 semantically correct means.

Stephen Wolfram 58:46 It's, it's, it's a good question. I mean, I think it is, it is... Given the way we have constructed language, it is things which, which fit with the things we're describing in language. It's a bit circular in the end, because, you know, you can... And, and the, and the, the sort of boundaries of what is physically realizable. Uh, okay. Let- let's take the example of motion, okay? Motion is a complicated concept. It might seem like it's a concept that should have been figured out by the Greeks, you know, long ago. But it's actually a really pretty complicated concept, because what is motion? Motion is... You can go from place A to place B and it's still you when you get to the other end, right? You, you take an object, you move it, and it's still the same object, but it's in a different place. Now, even in ordinary physics, that doesn't always work that way. If you're near a space time singularity in a black hole, for example, and you take your teapot or something, you don't have much of a teapot by the time it's near the space time singularity. It's been completely, you know, deformed beyond recognition. But... So that's a case where pure motion doesn't really work. You can't have a, a thing stay the same. But so this idea of motion is, is something that sort of is a, is a slightly complicated idea. But once you have the idea of motion, you can start... Once you have the idea that you're gonna describe things as being the same thing but in a different place, that sort of abstracted idea then has, you know... That has all sorts of consequences, like this transitivity of motion. Go from A to B, B to C, you've gone from A to C. Um, and that's... So at that level of description, you can have what are sort of, uh, inevitable consequences. They're inevitable features of the way you've sort of set things up. And that's, I think what this sort of semantic grammar is capturing, is things, things like that. And I, you know, I think that it's a question of what does the word mean when you say, "I go from..." "I move from here to there." Well, it's complicated to say what that means. This is this whole issue of, you know, is pure motion possible, et cetera, et cetera, et cetera. But once you have kind of got an idea of what that means, then there are inevitable consequences of that idea.

Lex Fridman 1:00:55 But the, the very idea of meaning, it seems like there's some words that become, um... It's like there's a latent ambiguity to them. I mean, it's the word like emotionally loaded words like hate and love.

Stephen Wolfram 1:01:12 Right.

Lex Fridman 1:01:12 It's like, what- what do they- what do they-

Stephen Wolfram 1:01:14 Well-

Lex Fridman 1:01:14 ... mean exactly? What- what, um, so especially when you have relationships between complicated objects, we seem to take this kind of shortcut, descriptive shortcut of, to describe like-

Stephen Wolfram 1:01:25 Right.

Lex Fridman 1:01:26 ... Object A hates Object B. What's- what's that really mean?

Stephen Wolfram 1:01:29 Right. Well, words are defined by kind of our social use of them.

Lex Fridman 1:01:35 Sure.

Stephen Wolfram 1:01:35 I mean, it's not, you know... A word, in computational language for example, when we say we have a- a construct there, we expect that that construct is a building block from which we can construct an arbitrarily tall tower.

Lex Fridman 1:01:50 Mm-hmm.

Stephen Wolfram 1:01:51 So we have to have a very solid building block. And, you know, we have to... It turns into a piece of code, it has documentation, it's, you know, it's a whole, it's a whole thing.

Lex Fridman 1:01:59 Mm-hmm.

Stephen Wolfram 1:01:59 But the word hate, you know, the documentation for that word, well, there isn't a standard documentation for that word, so to speak. It's a complicated thing, defined by kind of how we use it. When, you know, if it wasn't for the fact that we were using language, I mean... So- so what is language at some level? Language is a way of packaging thoughts so that we can communicate them to another mind.

Lex Fridman 1:02:23 Can these complicated words be converted into something, uh, that a computation engine can use?

Stephen Wolfram 1:02:30 Right. So- so I think the answer to that is that- that what one can do in computational language is define, make a def- make a specific definition. And if you have a complicated word, like let's say the word eat, okay? You'd think, "Oh, it's a simple word." It's, you know, animals eat things, whatever else. But, you know, you do programming, you say, "This function eats arguments."

Lex Fridman 1:02:53 Mm-hmm.

Stephen Wolfram 1:02:53 Which is sort of poetically similar-

Lex Fridman 1:02:56 Mm-hmm.

Stephen Wolfram 1:02:56 ... to the animal eating things. But if you start to say, "Well, what are the implications of, you know, uh, the function eating something?" You know, does it, can- can the function be poisoned? Well, maybe it can actually. But, um, uh, you know, if there's a type mismatch or something (laughs) in the, in some language. But- but, you know, in what... how far does that analogy go? And it- it's just an analogy.

Lex Fridman 1:03:17 Yeah.

Stephen Wolfram 1:03:17 Whereas, uh, if you use the word eat, in a computational language level, you would define there isn't a thing which you anchor to the kind of natural language concept eat.

Lex Fridman 1:03:29 Yeah.

Stephen Wolfram 1:03:29 But it is now some precise definition of that, that then you can compute things from.

Lex Fridman 1:03:33 But don't you think the analogy is also precise, software eats the world?

Stephen Wolfram 1:03:36 (laughs)

Lex Fridman 1:03:37 Don't- don't you think there's, um, there is something concrete in terms of meaning about analogies?

Stephen Wolfram 1:03:45 Sure. But the thing that sort of is the first target for computational language is to take sort of the ordinary meaning of things and try and make it precise, make it sufficiently precise you can build these towers of computation on top of it.

Lex Fridman 1:03:59 Yeah.

Stephen Wolfram 1:03:59 So it's kinda like, if you start with a piece of poetry and you say, "I'm going to define my program with this piece of poetry."

Lex Fridman 1:04:06 Mm-hmm.

Stephen Wolfram 1:04:07 It's kinda like that's- that's a difficult thing. It's better to say, "I'm gonna just have this boring piece of prose and it's using words in the ordinary way."

Lex Fridman 1:04:16 Yeah.

Stephen Wolfram 1:04:16 "And that's how I'm communicating with my computer, and that's how I'm going to build this solid building block from which I can construct this whole kind of computational tower."

Lex Fridman 1:04:25 So there is some sense where if you take a poem and reduce it to something computable, you're gonna have very few things left. So maybe there's a bunch of human interaction that's just poetic, aimless nonsense.

Stephen Wolfram 1:04:39 Well-

Lex Fridman 1:04:39 That's just-

Stephen Wolfram 1:04:40 ... it- it-

Lex Fridman 1:04:40 ... like recreational, like hamster in a wheel. It's not actually producing anything. (laughs)

Stephen Wolfram 1:04:45 Well, I- I- I- I think that that's a complicated thing, because in a sense, human linguistic communication is, there's one mind, it's producing language. That language is having an effect on another mind.

Lex Fridman 1:04:58 Yeah.

Stephen Wolfram 1:04:58 And the question of, there's sort of a- a type of effect that is well-defined, let's say, where, where for example, it's very independent of the two minds, that the, it doesn't, you know, there- there, there's communication where it can matter a lot, sort of what the experience of- of, um, uh, of one mind is versus another one and so on.

Lex Fridman 1:05:20 Yeah, but, uh, what is the purpose of natural language communication? Can we-

Stephen Wolfram 1:05:27 Well, I think, I think the-

Lex Fridman 1:05:28 Versus... So computation, computational language somehow feels more amenable to the definition of purpose. It's like, yeah, you're given two clean representations of a concept and you can build a tower based on that.

Stephen Wolfram 1:05:43 Right.

Lex Fridman 1:05:43 Is- is natural language the same thing but more fuzzy? What-

Stephen Wolfram 1:05:46 Well, I think the- the story of natural language, right, and the- the... That's the great invention of our species. We don't know whether it exists in other species, but we know it exists in our species. It's the thing that allows you to sort of communicate abstractly from like one generation of the species to another. You can, you know, there is an abstract version of knowledge that can be passed down.

Lex Fridman 1:06:08 Mm-hmm.

Stephen Wolfram 1:06:08 It doesn't have to be, you know, genetics. It doesn't have to be, you know, you don't have to apprentice the next specie- you know, the next generation of birds to the previous one to show them how something works.

Lex Fridman 1:06:18 Yeah.

Stephen Wolfram 1:06:18 There is this abstracted version of knowledge that can be kind of passed down. Now, that, you know, it relies on, it still tends to rely, because language is fuzzy, it does tend to rely on the fact that, you know, if we look at the, you know, some ancient language that, where we don't have a chain of translations from it until what we have today, we may not understand that ancient language. Um, and we may not understand, you know, its concepts may be different from the ones that we have today. We still have to have something of a chain, but it is something where we can realistically expect to communicate abstract ideas. And that's, you know, that's one of the big- big roles of- of language. I think, you know, in- in, um, uh, it's, you know, and that- that's been, this- this ability to sort of concretify abstract things is what- what language has provided.

Lex Fridman 1:07:11 Do you see natural language and thought as the same, the stuff that's going inside your mind?

Stephen Wolfram 1:07:17 Well, that's been a long debate in philosophy.

Lex Fridman 1:07:20 It seems to be, become more important now when you- we think about... how intelligent GPT is.

Stephen Wolfram 1:07:27 Whatever that means.

Lex Fridman 1:07:28 Whatever that means. But it seems like the stuff that's going on in the human mind seems something like intelligence.

Stephen Wolfram 1:07:34 Yes.

Lex Fridman 1:07:34 And its language.

Stephen Wolfram 1:07:35 Well, we call it intelligence.

Lex Fridman 1:07:36 Yeah. We call... Well, yes.

Stephen Wolfram 1:07:38 (laughs)

Lex Fridman 1:07:38 And then so you- you start to think of, okay, what's the relationship between thought, the language of thought, the laws of thought, the laws of-

Stephen Wolfram 1:07:46 Right.

Lex Fridman 1:07:47 ...and words like reasoning and, uh, the laws of language, and how that has to do with computation.

Stephen Wolfram 1:07:53 Right.

Lex Fridman 1:07:53 Which seems like more rigorous, precise ways of reasoning.

Stephen Wolfram 1:07:57 Right. Which are beyond human.

Lex Fridman 1:07:59 Yeah.

Stephen Wolfram 1:07:59 I mean, much of what computers do, hu- humans do not do. I mean, you- you might say-

Lex Fridman 1:08:04 Humans are a subset.

Stephen Wolfram 1:08:06 Yes.

Lex Fridman 1:08:06 Presumably.

Stephen Wolfram 1:08:07 Yes.

Lex Fridman 1:08:07 Hopefully.

Stephen Wolfram 1:08:08 Yes.

Lex Fridman 1:08:09 (laughs)

Stephen Wolfram 1:08:09 The- the... Yes, right. You know, you might say, who needs computation when we have large- large language models? Large language models can just, you know... Eventually, you'll have a big enough neural net it can do anything.

Lex Fridman 1:08:19 Mm-hmm.

Stephen Wolfram 1:08:19 But they're really doing the kinds of things that humans quickly do. And there are plenty of sort of formal things that humans never quickly do. For example, I don't know... I, you know... You can sum... People can do mental arithmetic. They can do a certain amount of math in their, in their minds. I don't think many people can run a program in their minds, of any sophistication.

Lex Fridman 1:08:39 Mm-hmm.

Stephen Wolfram 1:08:39 It's just not something people do. It's not something people have even thought of doing, 'cause it just... It's kind of a... It's kind of not, you know... You can easily run it on a computer.

Lex Fridman 1:08:48 Well, an- an arbitrary program.

Stephen Wolfram 1:08:50 Yeah.

Lex Fridman 1:08:50 Aren't we running specialized programs?

Stephen Wolfram 1:08:52 Yeah, yeah, yeah. But if I say to you-

Lex Fridman 1:08:54 Run this program.

Stephen Wolfram 1:08:54 ...here's a Turing machine-

Lex Fridman 1:08:56 Yeah.

Stephen Wolfram 1:08:56 ...you know, tell me what it does after 50 steps. And you're, like, trying to think about that in your mind. That's really hard to do. It's not what people do. I mean it-

Lex Fridman 1:09:04 Well, in some sense, people program... They build a computer, they program it, just to answer your question about what the system does after 50 steps. I mean, humans build computers.

Stephen Wolfram 1:09:16 Yes.

Lex Fridman 1:09:17 (laughs)

Stephen Wolfram 1:09:17 Yes, yes, that's right.

Lex Fridman 1:09:18 So... (laughs)

Stephen Wolfram 1:09:18 But they've created something which is then, you know... Then when they run it, it's doing something different than what's happening in their minds. I mean, they've outsourced that- that piece of computation from something that is in- internally happening in their minds to something that is now a tool that's external to their mind.

Lex Fridman 1:09:37 So by the way, humans, to you, didn't invent computers, they discovered them.

Stephen Wolfram 1:09:43 They discovered computation.

Lex Fridman 1:09:44 Which-

Stephen Wolfram 1:09:45 They invented the technology of computers.

Lex Fridman 1:09:51 This... The computer is just a... kind of way to plug into this whole, this stream of computation.

Stephen Wolfram 1:09:58 Yes.

Lex Fridman 1:09:58 There's probably other- other ways. There's probably a lot of ways.

Stephen Wolfram 1:10:01 Well, for sure. I mean, the- the, you know... The particular ways that we make computers out of semiconductors and electronics and so on, that's the particular technology stack we've built. I mean, the story of a lot of what people try to do with quantum computing is finding different sort of underlying physical s- uh, you know, infrastructure for doing computation. You know, biology does lots of computation. It does it using an infrastructure that's different from semiconductors and electronics.

Lex Fridman 1:10:26 Mm-hmm.

Stephen Wolfram 1:10:26 It's a, you know, it's a molecular scale, uh, sort of computational process that hopefully we'll understand more about. I have some ideas about understanding more about that. But, uh, you know, that's a, that's another in- you know, it's another representation of computation. Things that happen in the physical universe at the level of, you know, these evolving hypergraphs and so on, that's another sort of implementation layer for this abstract idea of computation.

Lex Fridman 1:10:52 So if GPT or large language models are starting to form, starting to develop, or, uh, implicitly understand the laws of language and thought, do you think they can be made explicit?

Stephen Wolfram 1:11:04 Yes.

Lex Fridman 1:11:05 How? (laughs)

Stephen Wolfram 1:11:07 With a bunch of effort. I mean-

Lex Fridman 1:11:09 So do- do they have-

Stephen Wolfram 1:11:10 ...it's- it's like doing natural science. I mean, what is happening in natural science? You have the world that's doing all these complicated things, and then you discover, you know, Newton's laws, for example. This is how motion works. This is the way that this particular sort of idealization of the world, this is how we describe it in a simple, computationally reducible way. And I think it's the same thing here. It's, there are sort of computationally reducible aspects of what's happening that you can get a kind of narrative theory for, just as we've got narrative theories in physics and so on.

Lex Fridman 1:11:43 (sighs) Do you think it will be depressing or exciting when all the laws of thought are made explicit, human thought are made explicit?

Stephen Wolfram 1:11:54 I think that once you understand computational reducibility, it is, uh, it's neither of those things. Because the fact is, people say... For example, they, people will say, "Oh, but, you know, I have free will. I, I kind of, um, you know, I operate in a way that is, uh, uh, um..." You know, you- you, uh, th- the, they- they have the idea that they're doing something that is sort of, uh, of... e- internal to them that they're figuring out what's- what's happening. But in fact, we think there are laws of physics that ultimately determine, you know, every, uh, every nerve, uh, you know, every electrical impulse in a nerve and things like this. So you might say, isn't it depressing that we are ultimately just determined by the rules of physics, so to speak? Um, it's the same thing. It's at a higher level. It's like it's- it's- it's a shorter distance to get from kind of semantic grammar to the way that we might construct a piece of text than it is to get from individual nerve firings to how we construct a piece of text. But it's not fundamentally different. And by the way, as soon as we have this kind of level of, you know, this other level of description, it's kind of... It helps us to go even further. So we'll end up being able to produce more and more complicated kinds of, kinds of things that just like when we... You know, if- if we didn't have a computer and we knew certain rules, we could write them down, we could go a certain distance. But once we have a computer, we can go vastly further, and this is the same kind of thing.

Lex Fridman 1:13:22 You wrote a blog post titled, "What is ChatGPT Doing and Why Does It Work?" We've been talking about this, but can we just step back and linger on this question? What, what's it, what's ChatGPT doing? What, what are these, um, bunch of billion parameters trained on a large-

Stephen Wolfram 1:13:40 Right.

Lex Fridman 1:13:40 ... number of words? Like, (sighs) wh- why does it seem to work, again? Is it be- is it because to the point you made that there's laws of language that can be discovered by such a process? Is there something more to it?

Stephen Wolfram 1:13:53 Well, let- let's talk about sort of the low level of what ChatGPT is doing.

Lex Fridman 1:13:57 Yeah.

Stephen Wolfram 1:13:57 I mean, ultimately, you give it a prompt, it's trying to work out, you know, what should the next word be.

Lex Fridman 1:14:04 Mm-hmm.

Stephen Wolfram 1:14:04 Right?

Lex Fridman 1:14:05 Which is wild. Isn't that, isn't that surprising to you that this kinda low-level, dumb training procedure can create something syntactically correct, first, and then semantically correct-

Stephen Wolfram 1:14:19 You know-

Lex Fridman 1:14:19 ... second?

Stephen Wolfram 1:14:20 ... the thing that has been sort of the story of my life-

Lex Fridman 1:14:22 (laughs)

Stephen Wolfram 1:14:22 ... is realizing-

Lex Fridman 1:14:23 Yeah.

Stephen Wolfram 1:14:23 ... that simple rules can do much more complicated things than you imagine, that something that starts simple and starts simple to describe can grow a thing that is, you know, vastly more complicated than you can imagine.

Lex Fridman 1:14:37 Yeah.

Stephen Wolfram 1:14:37 And, and honestly, i- it's taken me ... I mean, I don't know. I've, I've sort of been thinking about this now 40 years or so, and it always surprises me. I mean, even, for example, in our physics project, sort of thinking about the whole universe growing from these simple rules, I still resist because I keep on thinking, you know, how can something really complicated arise from something that simple? It just seems, you know, it seems wrong, but yet, you know, the majority of my life, I've kind of known from, from things I've studied that this is the way things work. So yes, I- it is wild that it's possible to write a word at a time and produce a coherent essay, for example. But it's worth understanding kinda how that's working. I mean, it's kinda like if, if it was going to say, you know, "The cat sat on the," what's the next word? Okay, so how does it figure out the next word? Well, it's seen a trillion words written on the internet, and it's seen, "The cat sat on the floor, the cat sat on the sofa, the cat sat on the" whatever.

Lex Fridman 1:15:34 Mm-hmm.

Stephen Wolfram 1:15:35 So its minimal thing to do is just say, "Let's look at what we saw on the internet." We saw, you know, 10,000 examples of, "The cat sat on the."

Lex Fridman 1:15:44 Mm-hmm.

Stephen Wolfram 1:15:45 What was the most probable next word? Let's just pick that out and say that's the next word. And that's, that's kind of what it is at some level it's trying to do. Now, the problem is there isn't enough text on the internet to, uh, for if you have a reasonable length of prompt to ... That, that pro- that specific prompt will never have occurred on the internet. And as you, as you kind of go further, there just won't be a place where you could've trained, you know, where you could just, just worked out probabilities from what was already there.

Lex Fridman 1:16:15 Mm-hmm.

Stephen Wolfram 1:16:16 Um, you know, like, if you say two plus two, there'll be a zillion examples of two plus two-

Lex Fridman 1:16:20 Mm-hmm.

Stephen Wolfram 1:16:20 ... equaling four, and a very small number of examples of two plus two equals five, and so on.

Lex Fridman 1:16:25 Mm-hmm.

Stephen Wolfram 1:16:25 And you can pretty much know what's going to happen. So then the question is, well, if you can't just work out from examples what's gonna happen, just n- with no probabilistic for different examples what's gonna happen, you have to have a model. And this kind of an idea, this idea of making models of things, is an idea that really, I, I don't know, I think Galileo probably was