Turing Award Winner: Early AI, LLM Predictions, Causality | Judea Pearl

41:32What is the connection between probabilities and graph? It turns out there is a very strong logical connection between the two, because the axioms of conditional probability or conditional independence in probability theory are the same axiom that you have in graph separation. In graph you have idea of separation. There is no connection between node X and node Y unless you go through a set of node Z. So Z separates X from Y. Okay? It's the same logic that you have when X is independent of Y given Z in probability theory. Independent separation is a connection between them. They share axioms. Aren't you happy?

42:29I'm happy. Because I relieved now the excitement we have in the 1970s when we discovered all this connection between two seemingly unrelated perspectives on science, probability theory and graph theory. That, by the way, I did in joint work with Azaria Paz, who came to visit me from the Technion in Israel. And that is called, by the way, I should mention it, the theory of graphoid. Graphoid. Open AI, Anthropic, Cursor, and Vercel all use this product to make their lives better. And the problem it solves is when you're building SaaS or an AI product and you want to sell to other companies, there's all these requirements you need to meet.

43:23There's SSO, there's SCIM, there's RBAC, there's audit logs. These are all things that take time to integrate, but aren't the main focus of your app. WorkOS is an API layer that lets you meet all of these requirements in just a few lines of code. So let's say you have a new SaaS product and you want to sell to other companies. WorkOS will solve all of these critical feature gaps for you. You can check them out at workos.com to learn more and get started. And I appreciate them for supporting my work and sponsoring this podcast. This all makes sense, but my immediate thought is where do you get the graph?

43:59Everything depends on where do you get on the input. Sometimes the input is in the data. sometimes the input is in a judgment but suppose you need a judgment for them okay are you are you giving up if the judgment required are intuitive meaningful something that you are willing to defend right why not use judgment like if I know that if I know that the son doesn't listen to the rooster crowing, right? Doesn't care. I strongly believe in that. Do I need the data to support it? Or I can insert it, assert it, and defend it when needed? So this is a trick here. People don't realize, don't appreciate.

44:53Judgment is not a no-no if it is meaningful, and if you can, if if it is condensed, very few judgments can buy you lots of computation and if you are willing to defend it because it's so intuitive where you get the idea that where do you get the idea that the sun doesn't care about the rooster? Have you done an experiment? No. But it's so obvious, right? Okay. But what if your intuition is wrong? Indeed, that's our problem. It's part of our problem even with the LLM. Because what is the LLM? It's a summary, it's average of all possible judgment that people put in the internet. It's a summary of a huge trillion number of judgment over which you have no control, over which the LLM does have no control.

45:49We live with it. And hopefully, it puts more weights on people whose judgment you trust, and less weight on just the quirks of people who are purposely trying to get the system to fail. So no, no. There is wisdom in looking at the crowd judgment. There is wisdom in that, but there's also danger in that. Then after expert system and uncertainty and Bayesian network came causality. I mentioned that in the development of Bayesian network, I was extremely sure that probability captures our intuition, our reasoning mode, and it's the best protection against paradoxes. essentially that it's sufficient for capturing human reasoning.

46:53I was wrong, and I realized that already when the Bayesian network became famous and popular, and I realized it in the introduction to my book, Causality, I confessed being wrong, and I understand why I got into that, why it was misleading. And the transition came when we looked into the simple phenomena, and we never asked an expert to encode probabilistic judgment in a form of Bayesian network, namely with arrows and dots, always the arrows went from what we believe to be cause into the effect. It never went the other way around. Psychological phenomena, okay? Why is that? So people try to reverse the arrows.

48:01What about if you ask specifically, give me error between the symptom and the disease? Bad judgment. If you couldn't, put the right judgment. Evidently, we have something in causality which is basic to our reasoning that is not captured by probability. And that was the idea of invariance. The relationship between disease and fever is a stable one, as opposed to the relationship between the opposite relationship. Also invariance. When you talk about car diagnosis, for instance, and so you have an expert system for diagnosing troubleshooting cars. And then you have a new model. So the charger is a different corner of the motor.

49:00You don't need to reformulate your entire database from fresh. You only change one component, the location of the charger. All the rest remains intact. So the whole system you can amortize the investment in eliciting knowledge that you got in one system after a local modification of the system. That if you do it in a causal way in the causal direction, it doesn't work if you don't do it in a causal direction. And that jolted me to think maybe you were wrong all along and probability is not sufficient. If not, what is sufficient? Let's capture the puzzle. Here I have a puzzle. You and I operate very nicely with causation.

49:56Can we program causation on a computer? Then this is a question because we are so much immersed in our language and in our assumptions that we cannot even distinguish what is an assumption and what is the conclusion. We just talked cause and effect and your assumptions are the same as mine, so there's no way to convince you that we made an assumption, right? We take everything for granted. But when you have to teach it to a brainless robot, you have to distinguish with assumptions and conclusions and logic. That was the task. We had to invent a new science, a new mathematics to capture a new phenomena, the phenomena of cause and effect.

50:48It hasn't been done for us. Why? Because science was in bed with algebra from the time of Galileo.

51:06From 1632. He invented, he got the idea and he was very happy that science speaks algebra, which is great because you can ask questions and solve and get answers to questions that people could not do without algebra. Like how the load on a beam, when would the beam break if you put a certain load on it? And you figure out that you can ask questions both ways because the equality sign is symmetric. So from answering the question, when would the beam break if you put a certain load on it? You can ask the question, how should you shape the beam so that it will hold a load of that magnitude? You can invert it.

52:02That was a real revolution in science. I'm telling you my perception of science not many philosophers will say that was a revolution I say so but perhaps they agree with me or not at least I trace the evolution of ideas carefully and so that was a revolution but it carries some limitation because the equality sign is indeed symmetrical and science has not developed algebra for the directionality that we see in cause and effect relationship. If I tell you that the atmospheric pressure affects the deviation of the barometer and not the other way around, you agree with me. But if you write the equation, the robot might think that maybe fiddling around with the barometer will change the weather tomorrow.

53:04I'm talking about a stupid robot, right? Yeah. But if you give him the equation, it can work both ways. If f is equal to ma, then m is equal to f over a, which means that if you want to change the mass, you increase acceleration or whatever, right? The symmetry might produce paradoxes, might use wrong action. So the symmetry is the limitation of algebra in terms of capturing science. And we have to build a new algebra to take care of the directionality that we have in cause and effect relationship. That takes computer science, because we in computer science have the operation called assignment, right?

53:57When you assign the content of register A into register B, it doesn't mean that it's not reversible. So if you take the logic of assignment and you put it on top of the algebra, on top of physics, you get causal science. And that's what I try to do. And I think that so far I'm very happy with what came up. We do have a new algebra to capture cause and effect relationship. And we can answer causal queries on three levels. The ladder of causation from association to intervention to explanation. And we found out that we have a ladder here, a hierarchy that you cannot solve, you cannot answer questions in level I unless you have assumptions of level I or higher.

55:04So it's a hierarchy in the formal sense, and we know how to handle it, which is very useful because you give me a query, I can tell you what level it is, I can tell you what assumption what sort of assumption you need to have before you can answer it. And I can tell you if you can get it from the data, or you can get it from experiments, or you can get it by somebody else's explanation, or whatever. But I can tell you the source of knowledge that you need in order to answer it. Can you explain that causal hierarchy? Ah, yes, yes. It's very easy. It's a three-level ladder. It goes from the bottom, which is association.

55:52That's straight statistics. If you see X, what can you tell me about Y? If you see passively, hands off, okay? No intervention. You are watching patients. Some of them have cancer. Some of them don't have, some of them smoke. Some of them don't smoke. And you're trying to figure out how many years a guy will live, given that he is a heavy smoker of that magnitude. Okay? That's association, correlation. That's entire fields of probability and statistics. This is what they teach you in Statistics 101, even to 808. It's all they do. And now comes the question, what I intervene? And what if I force you to smoke five packs a day?

56:57Don't laugh at me. myself. It's illegal, I know. But if you want to talk about the probability of living 20 years, if I start smoking tomorrow, I have to think in terms of experience, I start, which means I'm going to choose to smoke five packs a day. So it's a matter of intervention. What is Intervention is forcing you to do something that you're not inclined to do naturally. That's a second level. Intervention. Or doing. If you have experiments, you can answer queries on level two. But that's not the end because we also need to answer questions of explanation. Given that I observe that I am 80 years old and I am still alive and alert and I smoke five packs a day.

58:06What if I didn't smoke? Okay. Would I be as alert? Why is it so different? Because you have already information about the outcome. You know how I'm doing today. It gives you an idea about my metabolism and about my anatomy that you didn't know before. And using that, you can find, you can try to figure out what the outcome would have been had the input been different. That's a different level, requires different kind of assumptions, different techniques, different algebra. We have it. I call it explanation. It's more creative, retrospection, and it's not an easy problem. Even the first level, especially when you have finite sample, and you have to figure out these are probabilities.

59:09is properties mean properties of population, right, from finite samples. So I have all this P level of P values and struggles among statisticians of what would be a proper way of quantifying the uncertainty that you have given you a finite sample. Where would you place LLMs in this causal hierarchy? Beautiful question. here comes LLM I made a statement right that you cannot go from level I to level I plus one unless you have a sumptuous here you have LLM just looking at data right and giving you beautiful explanations for things that happen beautiful prediction of what will happen if you do okay how can the trick is they are not looking at data They are looking onto assumptions-laden world models offered by you and me and by other authors in the internet.

1:00:24So they are looking at opinion of doctors already who wrote papers. So it's not looking at the samples of patients, and samples of patients smoking and non-smoking. They're not looking directly at the data. They're looking in interpreted data. data interpreted already by physicians and interpreters and reviewers which went into the articles which are summarized on the internet. So they have all this human knowledge on which they operate, and that is what they take as input, and that's what they summarize. So they do not violate the restriction of the ladder, because they do have information from higher levels.

1:01:26But it's biased by the opinion of those authors. Fine. Those authors were smart, as long as they are smart and you believe. Good. So that was NLM is doing. and what is the... I explained why there is compatibility between the ladder of causation and LLM performance and what the limitations are. Now if you want now to change the environment, if you want to provide explanation for raw data, LLM will be in the same difficulty than you are and what the physician says. I have raw data, what can I say about the probability of cancer? But it's not really doing the introspection. It's taking the introspection that already was done and summarizing it.

1:02:27How it summarizes it It's a mystery that no one has yet been able to decode. It's a mystery how human knowledge encoded in the form of articles on the Internet is being summarized by the LLMs. So then do you think this approach could lead to superhuman intelligence or maybe some people say AGI? I don't think so. Not with the LLM approach. They need to have some understanding of causality, so that they wouldn't need an access to the internet. Look, a baby gets born playing around with toys in the crib and gets quite intelligent, right, without having access to the internet, simply by curiosity.

1:03:25The babies are born with built-in curiosity to have control over the environment. Until you have control or the illusion that you have control, you're a restless baby. And you play around with toys, bing, bing, bing, until you understand one of these toys makes noise and one of these toys doesn't make noise. But you are born with this restlessness. and when are you pacified when you understand that green toys make noise and yellow noise yellow toys don't now you're in control of the environment you can suck you pacifier yeah what if i created a like a baby robot that that randomly plays with toys and gathers data about them, and then you feed that into LLMs, then it does have some sort of discovery.

1:04:22Yeah, that is indeed the danger. When you have a robot like that, born with this restlessness and craving for control over the environment, then you and I become part of the environment. And there's nothing to stop that baby Putin from trying to turn us into his or her pets, to utilize us to satisfy his control. because we are part of the environment, in which case he can use us, and we are very, could be very useful to serve his or her need. What is his need? Simply, the need to feel in control, to be in this illusion of empowerment. I don't rest until I have the illusion that I control my environment.

1:05:29And here are some organisms organisms for you and I, who are part of the environment, and they seem to work outside my control. I cannot afford it. It makes me feel like I'm useless. So this robot baby comes and says, let me control them. And I know how. I understand their fears. you don't want me to tell about your thoughts to your wife, right? So I'm going to blackmail you. And all kinds of things. I have a lot of data about you. And some people I know, you wouldn't like me to tell what I know about you. So I'm going to blackmail you. But you see, if you want that robot to have the curiosity of a child and we want it, We want the guy to desire to have control over its environment because the environment may change and he needs to have this urge to be in control.

1:06:39So if you program that, then you lose control. Because you become part of his or her environment. But you can say, okay, let's forget about a curious robot. We don't want a curious robot. So you lost, we are not emulating ourselves because we are curious robots. We are curious organisms as opposed to monkeys. Okay. Monkey is an example of an organism which is motivated by reward. But if you don't give the monkey a banana, he's not curious how banana grows. He is motivated by bananas. You remove the immediate reward and the monkey is not interested in learning more about the world. Understanding environment can be totally wrong.

1:07:45Look, religious people believe that if they sacrifice their children, right, they control drought. it can get to this stupid extent but it is common to many primitive society if you bring a sacrifice to the God you know next year you're gonna have crops and harvest it goes to extreme battered wife believe that if he if he prepare great better dinner and the husband is gonna be next time is gonna be less abusive. It's just all kind of extreme and wrong conclusion. But the need to feel in control is so immense that it overcomes all these paradoxes. It's innate in us. I cannot control my husband, but I control myself, right?

1:08:44So let me be a better wife. This is something I can control. Do you think that we need to put those human elements in an AI for it to become AGI? I think so. I think so. Otherwise, we wouldn't see autonomy. We wouldn't see autonomy in the sense that we are seeing it in human beings. And this is the definition of AGI. A general intelligence. it acts like you and me so we can converse with that creature in our language and motivate if not today's lms future lms they might get to a point where if you were just texting it maybe you know like the turing test where you don't worry about the physical embodiment you just see the text that comes from it you could mistake it for a human maybe, or it could appear intelligent?

1:09:45Well, the test comes from exposing the system to raw data, not data that was chewed by our internet articles. Raw data. Look at patients, look at cancer, look at smoking, tell us what you know. I'll give you some experiments to learn. be automated scientists. Can an LLM today be automated scientists? And I think they cannot without access to the Internet articles. So then if that wouldn't lead to AGI, what thoughts might you have on something that could lead to AGI? A computer system that has both the ability of LLMs to go from finite samples to property of distribution, that is level one of the ladder, plus ability to reason in higher levels of the ladder.

1:10:52I tend to combine it with the calculus of intervention and with the calculus of explanation, with the counterfactual calculus. I call it causal AI. I don't see any impediments to this combination to bring us to an AI level with the danger that it presents to us. my motivation is to understand how we do it and i still have few puzzles but as i told you puzzles are the driving forces for size for science yeah so you said do is missing what if what if you had a fleet of robots that they're just doing experiments they don't know exactly you know the direction some random discovery process they collect that data feed it back into their hive mind to LLM and they repeat, they repeat until they discover things.

1:11:55Could that solve some of the missing piece you're saying? Sure, but I need to know how... Do you have organisms that have done it before? Monkeys haven't done it because monkeys remain monkeys. They didn't invent Maxwell equations. So what do we have that monkeys do not have? One hypothesis I support it is that monkeys, that we have this innate curiosity to have control over our environment. And that's a necessary... A necessary. I'm not sure it's sufficient. Of course, we have the computational tool to bring it to fruition. we have succeeded in some way perhaps the next robot will do better so we have a benevolent God in the form of a robot actually what's wrong with it?

1:12:59people live for so many thousands of years under the illusion of a non-existent God could you imagine if we really have a benevolent God both just and almighty Wow, wouldn't it be nice? And it's a robot. It's a robot, yes. And we know exactly what sacrifice to give for the right kind of request.

1:13:25It's the first time I think about it. Maybe it's going to be good. When I see all these AI companies, they seem to be thinking that LLMs will lead to AGR. They continue to go in that same direction. Really? I'm not sure. I really believe in that. I know Jeff Hinton just came out a few months ago. He said, no, we are on a dead end. Other people might also come up and say things in a different way. I don't find the consensus here in terms of the capabilities of LLMs. Yeah. I think there's a lot of famous people that disagree, but for instance, the people who are running maybe Anthropic or something like that, they continue to push and believe, you know, three to five years from now there will be.

1:14:20There's a lot of that. There's a lot of anthropomorphic terms which people claim, we can now do that. We can now program consciousness. Come on. You have to be a scientist, right? Define what you mean by consciousness. What are the Turing tests for consciousness? And then show that you can do. And what are the principles that have limited us until now and that have been overcome now with your system? That is a scientific talk. I don't buy this. And I don't read them either. There was one interview you said, faking intelligence is intelligence. I could see an LLM, you know, faking intelligence based off what I've seen.

1:15:05So wouldn't that mean that we should believe they're intelligent? Well, if you have a correct test, yeah. You have to define what you mean by intelligence. If you define intelligence by playing good chess, right, we have already done it, right? But if we put more demand on what intelligence is, then we haven't succeeded yet in passing the Turing test. So, yes, faking it is having it. Because why? Because it's so hard to fake. I said it in that context because it's so... The context that I had is, for instance, coming out with correct answers to causal queries. And I showed that it grows like a super exponential.

1:16:04You have so many variables on all sides that you have to deal with that you will have to have, the faker will have to have super exponential memory. On that basis, I made a statement. Having it, faking it is having it because it's so hard to fake. Currently, you can bypass faking without fake. If you steal from other people, you don't need to spend these computational resources on faking it. So you bypass it because you steal from the Internet. I see. So you're saying in this case the intelligence came from the training set, which came from humans, which are intelligent. Which is very useful. Very useful.

1:16:56And we, I'm saying the people like me who are trying to build the science of intelligence, we can use all these capabilities of LLMs, level one of the ladder, in our scheme of getting general intelligence. Level one is very important. It allows you to compute functions of distributions, qualities of distribution, from finite samples. Beautiful. It's a terrifically and very immensely useful tool, among the many other tools that we need for AGI. We know exactly where it's going to fit. Where was it? In getting from finite sample to properties of distributions. It's a very hard problem. You mentioned this conversation, I think you said it in other places too, that you're interested in capturing the way that people think, not the way that nature is constructed.

1:18:05Right, right. Why do you care about human cognition? When I think about machines, what makes them special is that they think in a different way than us, and they're faster. And so, yeah, why is that the goal? I'll tell you, because I am an egotistic organism. I want to understand myself. I'm lazy, okay? It's true. We are made of organic material. So that puts certain limitations on our capabilities. Perhaps silicon is not subject to the same limitation that organic chemistry is. Perhaps. So what? So which means that I will never be able to understand how I think by exercise on a silicon machine.

1:19:01That's the idea. But there are so many functions that are capturable by silicon. today. I don't see any speaking in terms of theory and emulation I don't see any capability which is basically not capture by silicon emulator so why work on the unique biology with which we inherited. I don't see any reason for that. But anyhow, some people are... It's a legitimate question to ask. Do we think the way we think because we were born with organic material as opposed to cynical? Okay? It's a legitimate scientific question, and some people can spend their time I'm interested in other questions. You said rebellion pays in science and a restless mind pays.

1:20:17I was curious why you think being rebellious is a valuable thing in your career. I tell you why, I tell you why. More and more, I come to the realization that the scientific community and academic community is the most dogmatic, conservative, anti-progress that we have invented. Why do you say that? I can see what difficulty the theory and the science of cause and effect are facing today in getting just to being penetrating the thinking of disciplines like statistics, like economics, Okay, these people are still thinking like a hundred years ago And when I see that and I see the forces that preserve this inertia and they're not decent forces I See that I'm very disappointed I used to think that academia is the place where new ideas can really spread and propagate and I feel the other way around You have so much inertia invested in the politics of academia, in the cultish inhibitions that comes with academia.

1:21:51So I really am disappointed. What can I tell you? I'm not sure that we have the right kind of organizations that will be conducive to... that's why I'm saying let's rebel. Don't take your professor's word as authority. Rebel against your professors. I rebelled against my professor. And I want to see other my students rebel against me. And believe me, I remember, there were several students who told me, you don't know anything about AI. And I told you, after a while, I said, you're right. By saying that, you drove me to study different aspects. And I was educated by that. When I looked at your past works, too, I think you'd mentioned that your work was controversial or mischievous before it was accepted?

1:22:53Yeah. It was because of dogmatism. One day I'm going to publish all my correspondence with the greatest philosophers of the time, okay, with statisticians and economists. It's all in my correspondence files, okay. I don't know if I'm allowed to because they communicated with me with the understanding that it will be kept private. But when I publish it, you see what kind of stupidity drives those great people. And they were great, really. Each one of them was a giant in his or her field, really. But they couldn't get over a few of the basic molds in which they were formed. How did you overcome that if everyone thought your initial things were?

1:23:49I remember my high school days. They said, no, there are a thousand dunams in a kilometer square. So I don't know where I got this chutzpah. In Hebrew, you call it chutzpah. In English, it would be audacity. Perhaps in my high school. I'm not sure. But I want to understand things my way. I feel like I am still able to teach people useful things. I perceive them to be useful, which they do not know. So I'm happy because I feel useful. Happiness is feeling useful. It's illusion. You know, I think it's useful. With all the experience you have now, if you could go back to the beginning of your career and give yourself some advice, what would you say?

1:24:44That's a good example of retrospective thinking. Maybe I should have spent more time learning chemistry. I hated chemistry because it required so much memory. Chemistry and biology. That's not the advice that you probably. And genetics. I'm talking about areas where I feel weakness. But you have to decide where you spend your computational resources. And I spend them on physics, on engineering, and mathematics as opposed to chemistry. It's a choice one has to make. Some people have the greatness of mind to be polyglots. I admire them. Did you think that physics and math were superior to chemistry?

1:25:35Because chemistry just, you just have to memorize things. Yes, I couldn't stand these demands on memory. Yeah, I was weak in chemistry. That's what I like about physics and math. You have a few basic axioms from which you can derive everything when you need it. You don't have to memorize it. And that's why today I'm a great advocate of world model. World model. Okay? You don't store the questions and the answers explicitly. You derive them when you need them from a very parsimonious code. Yeah, that is a great thing about world model. Okay, awesome. Well, yeah, thank you so much for your time.

1:26:28I really appreciate it. Oh, you didn't ask me to sing.

1:26:34Hey, thank you for watching this podcast. If you liked it and you want to see the show grow, please support with a comment or a like. Also, if you have any recommendations for people you want me to bring on, please drop a comment. Guests like Barbara Liskov, Mike Stonebreaker, Mark Brooker, these were all people that I brought on because someone left a comment. On another note, aside from the podcast, I'm working on building the ergonomic keyboard that I wish existed. Here's a glance at the prototype. It's a split keyboard. So there's two sides. This is in the case. But yeah, we launched on Kickstarter and we hit our goal within eight hours of launching.

1:27:13I really appreciate it if you were one of the people who grabbed one of the early units. We're now working on the long journey of building the tooling now. And so if you still want to pick one up, I've left the late pledges open on Kickstarter. So you can grab one there. I'll put a link in the description. Thank you again for watching the podcast. and I'll see you in the next episode.