As a lot of the excitement around large language models was showing up, we were teaching this kind of broad overview of AI kind of the underlying kind of technology in terms of terms of symbolics and sub symbolic processing, but then also looking at applications and so forth. And my background is in I used to work in venture capital, lived in the Middle East and in Germany, and I'm just very interested in, like, new technologies, where do new things come from? And as somebody who's as kind of a strategy scholar, I'm interested in, you know, kind of novel theories about technology and and so forth. And I hope that that'll kind of come through. So I've got a timer here, so I'm going to try to make sure I stick to stick to my time and and then there'll be time for questions, and I know there'll be some good conversation there. So this might be already sort of evident to you, but you know, artificial intelligence already is beating human capability on lots of different dimensions. So this is just from the kind of the chat, GPT, technological background paper and and and so chess, checkers go and so forth. But even like professional, you know, exams, a CTL sat CPA exam, which is for, you know, for accounting, bar exam, for lawyers, some areas of medical diagnosis. It hasn't sort of been quite as big as people anticipated, but it's slowly creeping into that as well. And the broader question here that I'm interested in is also in terms of strategic decision making, and so are we our strategists? Are entrepreneurs and so forth. Is going to be replaced by artificial intelligence and maybe even researchers themselves, right? And so there's some arguments in that, in that direction as well. So I'll channel some of that as well. So here's Jeff Hinton, and I would highly recommend this. He gives a fantastic talk at the University of Toronto, I think it was earlier this year, and he gave kind of a similar talk at Oxford, and a couple other places where he says, basically, we're at this tipping point where digital intelligence maybe already surpasses biological intelligence and so and he argues that large language models are already there, and as we input more data and more compute, we're gonna, we're definitely gonna get there. And so that, that's the question. And we have lots of books that kind of capture this intuition. The first one here with Thomas Tom Griffis is a great book about kind of computational models of the of the human mind. It's worth reading. I'll talk about some stuff related to that. And the other other book that I'll touch on is this book by AJ Agarwal et al on prediction machines. And I'll talk about kind of prediction a little bit there as well. So I well. So here's, here's kind of the money quote that we kind of thought or focus this paper on, as we're thinking about about the future. My background is kind of in in, in some sense, Herbert Simon kind of doubted rationality and decision making that's kind of foundational to me. And obviously Danny Kahneman is a direct follow on to that. And this is a quote from him, and this is from from 2018 he says, Will there be anything that is reserved for human beings? Frankly, I don't see any reason set limits on what AI can do. So it's very difficult to imagine that with sufficient data, there will remain things that only humans can do. You should replace humans by algorithms whenever possible. So are we going to, are we? Are we sort of looking at this end state where eventually human decision making will be replaced by by algorithms and and AI, and this is this. This quote is just not, not only Kahneman, who, sadly, by the way, passed away. I think it was earlier this year. I had a chance to interact with him just via email on a couple occasions, on several papers, and I have to say, just a consummate gentleman and fantastic person. But anyway, the intuition has been echoed by others as well, more recently, in different venues that AI we need to, you know, replace human decision making with with AI. Here's a host of papers. This is just a small sample. So first one you have by some economist that says we're going to, you know that large language models in particular are going to, essentially, they serve as scientists, and they can even serve as subjects, and eventually they're going to sort of automate this process of social science. There's another paper by some folks at UBC and elsewhere at a soft AI startup that make the argument for an AI scientist, and they have kind of an algorithm or way of thinking about that. And then Felipe Cezar and harsh and Hanjin have a recent paper coming out in strategy science, which is probably, in some sense, like the opposite of this paper. And so they make a very strong argument that that artificial intelligence might be better than humans in terms of coming up with business plans specifically. And if I have time, I'll, I'll talk about that a little bit. So there's so lots of excitement around this, maybe to get to the heart of the issue, a lot of the AI focus is actually on prediction. And young lacoon has this nice quote. He says, prediction is the essence of intelligence. And when you look at this, and again, it's a great book, by by by by this team here, by Josh ganz and others. Fundamentally, if you sort of boil it down, what AI is good at is it's a prediction technology. So it's good at predicting things, and they have this sort of kind of quasi kind of model. You might think about where you go from having data to some kind of information to prediction to some kind of decision. And this is what I would argue is kind of a data first approach. And I'm going to try to push back on this data first approach and say that maybe we don't always want to start with data. In some situations, this makes sense, perfect sense, but there's many situations where this doesn't make sense. And in this in this book, they argue that even in strategic so, even in sort of decision making under uncertainty, it makes sense to follow this logic for of data, information prediction to decision. And I want to sort of push back on that a little bit, to argue that that might not be the right approach. And if this is the right approach, incidentally, then you do want to use AI like aI definitely is better at sort of the logic of moving from extant data to some form of information, prediction and decision. But I argue that in decisions of uncertainty, that's not that's not what we want to be thinking about, necessarily. It's interesting. Chris Anderson, I don't know if you've seen this, but this actually been cited across many disciplines. He wrote this piece. He was the editor of Wired, and he wrote this piece in 2008 called the end of theory, the data deluge makes the scientific method obsolete. My co author actually knows him, and he's got affiliation with Oxford, and so there's some interactions there. But in some ways, this kind of sort of foreshadowed a lot of the current moment where we say we need more and more data and more and more compute, and we don't actually need theories anymore. And so just to capture sort of the zeitgeist of this intuition here, he says, with enough data, the numbers speak for themselves. Faced with massive data, this approach to science, hypothesized model test is becoming obsolete, and correlation is enough. Correlation supersedes causation. So that's kind of some of the key intuition that I'm going to be, you know, pushing, pushing, pushing back on here. So, okay, I had a little bit of a back and forth with some computational geneticists as well. I just want to sort of point this out, so they wrote a piece called
the title of it is here a Hypo hypothesis is a liability. But this logic is showing up in in the natural sciences as well, where we don't need theories per se anymore, and with a team they're listed here. There's a cosmologist and a biologist and somebody from computer science with a the five of us sort of put together, kind of a response that channels and links to some of the things that I'll talk about here as well, so and that are talked about in the paper. So what I want to kind of at a highlight is that there's this general model of cognition that sees cognition, whether it's in terms of machines or humans, as being commensurate, as essentially doing the same thing. And if there's this general model of cognition, it suggests that certainly computers will be better at that than humans. And this to summarize lots of different literatures, we end up citing a lot of different things, probably too many in the piece itself, for some of you, if you happen to read the paper, but we wanted to be thorough in terms of our summary of the literature. But this general model of cognition is essentially this basic kind of information processing model where you have some form of input. So think about it as data, information stimuli, the environment, experience some some form of inputs that are are utilized, and then there's something that happens in the middle, information processing, computation, representation, something that happens in the middle, and then we get some kind of output, whether it's a prediction or response or decision behavior and so forth. And this is actually in the paper itself. This part used to be a lot longer, but we went even further back so we looked at the original Dartmouth conference, the papers that came out of it in 1956 or, you know, they were published subsequently. But even looked back at Turing in terms of how he thought about this process of cognition. And in his 48 and 50 papers, he has this notion of thinking about the computer, intelligent machinery, as he calls it, as as this untrained inference that has is sort of blank sheets and and we need to engage in this analogous kind of teaching process of of giving it inputs right. And so, in the context of large language models, what the data is is, you know, this data is words. And I'll give you a specific example and talk about just at a high level in terms of how a large language model is trained. Most of you probably have background on that, or you can go in the paper, and we cite lots of the kind of most recent pieces in that space. But the idea is that, again, that we start with data which is similar to the Agrawal prediction machines point that I was making making earlier, and this general model cognition, it's been around since the 50s, with Alan Newell in his kind of unified theories of cognition. So he Here's a quote from, from from him, from this book. He says psychology is right at the possibility of unified theories of cognition, where AI provides a theoretical infrastructure for the study of human cognition. And Simon, as you know, Herbert, Simon has several pieces where he talks about this general information process. General Information Processing paradigm. And again, a lot of this goes back to this 1956 conference, and even the original call was about, we're trying to understand the human mind, and we think that this notion of of thinking about it as a computer, will be a useful metaphor, and to help us in that process, there's a great, fantastic piece that sort of summarizes 40 years of cognitive architectures. They go through like 200 different cognitive architectures and so forth. And so I'm just putting this up there for anybody that's interested. It was published in AI magazine a couple years ago. But anyway, it's a nice, nice summary. So in the past, before we get to the current, kind of AI moment, that where we get large language model and when where some sub symbolic processing kind of takes off. In the past, the problems were largely the solutions were limited to highly specific domains, right? So it never kind of quite attained the generality that was hoped for. So Simon and Newell, for example, had this general problem solver, but it wasn't very general. It was being applied to very specific kind of areas. And in this, this piece, this is from the piece that I just cited, highlights some different domains, but there was very little evidence specifically in terms of the symbolic processing literature's AI of general problem solving, maybe the most successful applications came from my sin, which was this, if then rules for recommending antibiotics for infections. And so it was kind of this expert system, basically, sort of quasi artificial doctor that is fed in with details and then spits out a recommendation. And it was never actually implemented, but it was really influential in terms of thinking about, okay, how do we put in rules and heuristics for some kind of decision making or intelligence? And this was with with regard to, you know, medicine specifically, or in terms of, you know, antibiotics. And I think they had a total of, sort of, they're looking at, like, 50 different antibiotics and other possible infections and and so forth. But it was sort of basic, kind of, if then rules. That was how it was specified, um, obviously here so far I've talked about kind of these rule based heuristic systems of intelligence. But the big, sort of big turn probably in the, you know, sort of, I guess, early 2000 certainly 2012 2013 was with the sub symbolic systems, so artificial neural networks and and and so forth. So this is a picture of Frank Rosenblatt and and his early sort of modeling of what's, you know, the earliest sort of model of an artificial kind of neuron, or artificial neural network. But the idea was that, rather than impose top down, some kind of rules, we need to learn directly from the data. And again, this is kind of the idea of percept perceptron, and this is what's led now to artificial neural networks and machine learning, and the work by Hinton lacoon and suskova and many, many, many, many others. So let me just, rather than talking about this at a high level, give you one example of a of this. And I think get most of us are probably very familiar with
AI, as it's kind of instantiated in large language model. I'm a heavy, heavy, heavy user of these models, and so anytime there's a new update with open AI stuff, I'm on it. So I'm constantly using those also Claude and all the other different types of systems. And I don't want this paper to be read as sort of skepticism or lack of excitement. Both Matthias and I are very excited about these, and are heavy users of these. It's just a question of, is the argument about sort of these being eventual kind of super intelligences, particularly in terms of strategic decision making and so forth. The right argument to make in terms of even the current architecture, which people are saying right now. So just to give you some background, so we could, we could kind of contrast a large language model in terms of its inputs, and so a large language model is trained with this is kind of the data. So input to output, lots of, you know, 13 trillion tokens, everything from the internet and lots of books and Reddit posts and so forth. And it's now trained within a matter of weeks. It used to be a matter of months, but trained within a matter of months. And it's essentially, you know, ingesting and you know, it's input with all of this information that is, it's training data. And in terms of capability, what it's fantastic at is next word prediction. So it'll give you fluent text, as we all experience, right? So it's not misspelled, and it's well, fluently written, and so forth. I mean, there is argument around, sort of this convergence around what that writing looks like, and we could, we could, kind of talk about those types of issues. I guess one of the things I'm going to come back to this point is that as far as, sort of, as we look at these models, they are inherently backward looking, right? So they're stochastically sampling from the existing training training data set, right? And and so what they have, in some sense, is kind of a static Wikipedia level knowledge of the past. And I'll give you kind of a thought experiment for how to how to think about that here, here, in a minute. And if you think about humans in terms of how they are trained, or, you know, if we take the Turing idea of a human being trained in terms of words, at least, now, you know, we have a footnote around the multi modality of how humans like like, we take in lots of information visually. You know, the question of how we take this information, do we actually, you know, how we take that is, you know, a very deep and important one as well, but, but in terms of comparing it to a large language model, you know, we're encountering, and we kind of cite existing research and linguistics and child development, you know, a child encounters 18k words per day, and it would take 2 million years for it to sort of encounter These 13, you know, these 13 trillion word tokens. So it's not the same activity in terms of, at least in terms of words, but you but what we do sort of see is remarkable sort of generative speech that isn't directly necessarily tied to the inputs, right? And so I'm going to come back to that point, and there's some recent psychological research by Allison gopnik that I'll cite, or the cited in the paper that I'll highlight, that highlights how it's different linguistic views is different from a large language, while even from a three to seven year old. Okay? And the key point I'm going to emphasize is that there's something that goes on in terms of human cognition that I would call kind of forward looking theorizing or causal reasoning. And I'll give you an example of that metaphor. I want you to maybe come away with, is that the large language model is some kind of mirror on the past. It's a generative, clever mirror, right, right? So it's kind of repackaging, you know what it's seen in the past, but it's, but it's still a mirror versus in some sense, humans have a different mechanisms at play in terms of cognition, and I'll highlight what that is here in a second. So large language model. The big sort of turn was with this paper by Vaswani, at all attention is all you need. And our title is kind of a riff on, you know, theory is all you need, is a riff. Riff on that. Obviously, theory isn't all you need, and attention isn't all you need. You need you need many other things. But it's sort of to emphasize the importance of, in their case, attention in our case, you know, theory. But the important point is this, this breakthrough came from translation, essentially, right? So, so they have these existing benchmarks on, you know, how accurately something is translated from German to say English or English to German. And the transformer sort of technology here in this paper had a big jump in terms of how well it met these benchmarks in terms of a professional translator versus an actual sort of just computational system, right? But what the what the tool did, is essentially generalized translation. And one way to think about translation is taking one way of saying things and matching it with another way of saying things. So I want you to kind of, I guess, keep that in in mind as I, as I, as I, as I talk, make sure I keep track of time here as well. Okay, and so what llms do? Well, just to summarize a couple things, next word prediction, they're fantastic at that. Fluid sentences, stochastically, drawing from the training data. Again, the capability they'll they could write indefinite, well written articles about Andre or about about Donald Trump or Kamala Harris or whatever, right? And they wouldn't just be plagiarizing. They wouldn't be word to word snippets from existing texts. They would be novel sentences. And we kind of call this small g generativity, right? And so they're fantastic at that, right? And one way to think about it is they're a really good memory engine. They memorize far more. They can memorize far more things than a human can. Again, inputting something with that, you know, trillions of tokens is a it's quite, quite, quite the feat in terms of being able to sort of output that in kind of interesting, interesting ways. So the question that we ask is, what, what does an LLM know specifically? And so Truth For The LLM is a function of how frequently something has been said in the training data. So we have this kind of thought experiment, and it's trying to take it out of this current AI hoopla and excitement and to say, Okay, imagine this thought experiment. So imagine a large language model that was input with all text through 1633 so it was trained with all of science up to that point, right? And now you have somebody like Galileo come up. What would it say about Galileo? And the argument that we make in the paper is that it would actually say that this is not a thing, right? Because it's sampling from the existing training data set. It has no way to kind of bootstrap somehow beyond that. And even though there are people like, I don't know, Giordano, Bruno was the sort of philosopher that anticipated a lot of how we think about kind of the cosmos. Now, a lot of his texts were banned, but even if they were sort of in the training data set, they'd be dwarfed by all the textbooks that kind of reinforce the kind of the earth centric view of the universe, right? And so that's that's kind of one, one way to one way to think about what an large language model knows. And actually, what was considered science at the time of Galileo, for Tycho Brahe, for example, a cosmologist, wrote a lot of great things about cosmology, but he also wrote about astrology, and he wrote about Astra astrological things in scientific ways, right? And so the LLM wouldn't know any different. It would assume that, you know, how stars move, impact, how are my day is going, right? You know, horoscope type stuff, right? And this is, this was seen as science at the time, but it wouldn't have any way to differentiate it right now, there's all kinds of sort of patches and things that different models are now doing ensemble and retrieval, augmented generation of different things like that, and but they fundamentally, you're still stuck with the training data that you're using right and so that's sort of one example. So the limitations of AI kind of, when I didn't have a picture, I just kind of had chat GPT generate some images to kind of add a little bit of color. It doesn't have to give them color, but context here. So anyway, so one way to think about the limitations is it's a small g generative MIRROR OF ITS Training data. A large language model is reasoning, is simply mimicking. So when you do give, this is a really important point. Do give a large language model or reasoning task, it does great, right? So there's things like the waste and selection task, or the the one that Simon used a lot was missionaries and cannibals, or money the money Hall problem, so forth. There's lots of different sort of reasoning tasks. And when you give the AI a reasoning task, if you don't change the words, it does great. If you change the words, it gets very confused, right? So if you change the names of the protagonists or the change the units, or things like that, it doesn't give you a good answer. And again, they're putting in patches to try to change that. But maybe the biggest kicker is Francois scholt, who's a researcher at Google, has developed this, what he calls the abstract reasoning corpus, and he's basically hiding problems from large language models in terms of them being able to train on it okay? And by not by hiding these problems, he's able to not have them, memorize answers to have them, actually, genuinely not lie reason with these problems. And these are problems that a child looks at and is able to solve immediately, right? So now he's partnered with an entrepreneur, and they, I think, have a $1 million prize for anybody who's able to generate algorithms that are able to on the fly, not based on past draining data, solve new problems, right? And so this, this illustrates kind of, one of the one of the issues here that we have with, with with, with these types of models. Again, another limitation is kind of so can stochastic engine for producing text and, again, maybe one way to summarize it, we have a system for prediction based on past data,
I guess more generally, in the paper or effort. And I do work with folks in cognition. So I work with Jan kuenderick, who's a computer scientist and cognitive scientist, and Joachim Krueger, who's a brown who does, you know cognitive science. And so I'm very interested in just computational models in general, specifically applied to cognition. And this recent book by by Steve pinker summarizes a lot of that, lot of that research and, and in that, in that, he says, okay, data and evidence provide us with justified belief and knowledge, okay, and, and he has this interesting quote that we're going to push back, that we think is at the very sort of core of human cognition in terms of heterogeneity. And so this is the quote he says, I don't believe in anything. You have to believe in, right? And so the idea here is that we don't believe things. We have evidence for things, right? And I'm going to work through kind of a thought experiment to highlight that that might be problematic at the very forefront of knowledge, and certainly in the context of sort of novel technologies, entrepreneurship and strategy, which is where we're applying a lot of these things. And so I'm going to push back on that a little bit. But more generally, we have these kind of evidence based kind of decision making and Bayesian updating. These shows up in different ways. This is a this is from Garrett Gigerenzer, who's a cognitive scientist at this institute, but the brain is kind of this Bayesian computer. And so humans are statisticians and cosmetics and to be an American Economic Review, summarize a lot of the kind of evolutionary arguments around this. And this shows up in venabou and shows up in top econ journals as well. This notion of kind of Bayesian updating based on data so, and which is what we're going to kind of push back a little bit here. So this, this is probably familiar to you. This is just a basic sort of Bayesian kind of model. And here we have some kind of prior, the prior of a given hypothesis. Give you an example that we're going to work through. And here's the posterior. So once we've sort of worked through and thought about E is evidence, evidence for the hypothesis. If we believe that, if we believe the hypothesis is true, how do we get to updating our beliefs about the world? Essentially, and this is the logic I'm gonna work through. But this is sort of at a high level, what's happening in computational systems where we're talking about models, or any, any other types of types of models. I won't go through this. This is sort of basic definitions around each of these. And I'm sure you've seen Bayesian intuition in different forms before. We didn't have it. We don't have these types of equations in the paper itself. But I thought, I'll just add these. We thought about including them, but, but we what. We didn't. So let me work through this logic. So let's think, think about the probability of something. And so this is the posterior that we're trying to come up with. So this is on the left hand of the kind of the Bayesian formula. So the bed probably heavier than air flight in the fall of 1903 so what's the what's the evidence and the data? And so we want to be rational Bayesians and and not believe things just for the sake of believing things. And so the evidence is, at the time, zero successful attempts at flight. Scientific consensus for physics, flight is not possible. And this wasn't just like a few people. This was the people that you'd want to listen to. So Newcomb, Simon. Newcomb was Johns Hopkins, the top astronomer. Mathisian. Kelvin was Lord Kelvin was the head of the British Royal Society. Joseph Lacon was at Berkeley and was became the head of the the largest US scientific Association as well a few years later. So these are all points that we want to update our belief on, right? So points of evidence in some sense. So aviation pioneer Otto Lilienthal died during a flight attempt in 1896 and maybe so, if we're thinking about the probability of our us believing about the plausibility of flight in fall of 1803 the biggest kicker, maybe, or is Samuel Langley, also a scientist, had gotten money from, I think it was the Navy, lots of money. And he had this very public display. He had two days where he said, I'm going to show you how flight, you know, works. I've used all these funding. He had lots of these sort of the scientists lined up along the Potomac River there in Washington, DC and and including New York Times and things like that. And He came off this ramp, and he fell straight down on day one. The next day, he did the same thing, straight down. And so all of these things suggest, okay, we probably wouldn't assign a very high plausibility. In fact, it looked delusional at this stage, looking at this, these bits of evidence, right? And then there was this sort of persistent nagging, kind of colloquial or folk psychology or something that said, you know, people said, you know, people said, Well, what about birds? And so Joseph Lacon, who was Berkeley, said, Okay, well, I'm gonna put that to rest. He says he wrote a, he wrote a paper in scientific monthly, which later became part of science. And he said, I'm gonna, I'm gonna put this to the to the test. He's like, What about birds? And I'm gonna show you why this doesn't make sense. And he basically, this is actually chat gpts image, but you can see it from it has problems with, you know, text, as you see here. But he basically did this. He arrayed all birds, and he said, here we got large birds, and here we have small birds, then we have insects as well. Insects fly, and Simon Newcom had mentioned that the physicist, but it turns out that these large birds don't fly, and maybe there's just a limit of 15. He said, this is, this is conclusive data, that we shouldn't think about birds flying as some kind of evidence, but that we should update our our beliefs of on and, and. And so he tried to sort of put to bed this idea of flight, right? And so here's, here's the, here's a, there's a, here's a, you know, a picture from Samuel Langley falling into the Potomac River, right? And so here's quotes from kelvin and Newcomb saying that this is impossible. New York Times wrote an article a couple weeks later. It says they think maybe flight is possible in one to 10 billion years. I mean, again, ironically, now it was, I don't know, nine weeks later that we get the Wright brothers. So nine weeks later we get this. The question is, how do you get this right? And so here we have an example of, you know, something that a few months before, was delusional based on all sort of sources of evidence and data. How do you get to this? Is there some mechanism? And I'd argue that this notion of having some kind of theory Trumps data. Okay, it's not to say that you don't need data. Of course, you need data, but the question is, what's the right data? And so Kurt Lewin, who was a social psychologist, I like this quote a lot. It's nothing so practical as a good theory. So what does the Wright Brothers theory look like? So one way to summarize it is that if we take the same equation and we say, okay, we think the evidence matters, but maybe we can actually, like, causally intervene in the world and generate new data, right? The question is, Can artificial systems I was talking to Andre, Andre Andre, right before this, can artificial systems do this based on past data? Right? In our case, in terms of, you know, humans theorizing, we can generate causal intervention experiments that generate new data and evidence. And so what does this look like in the context of Wright Brothers? Interesting enough, I'm here. You have not, you know, I think one of them got like the high school equivalent, or something like a diploma, or something like that. But I think the other one was not, hadn't gone through high school, but they're bicycle mechanics, and they have these sort of sheepish, sheepish letters that I quote, quickly, that we quote quickly in the paper where they were writing to people like Samuel Langley and top scientists are saying, I know we're delusional, but we still think this is a thing. But what's interesting is that sort of okay, in some case, they're, you know, ignoring all this data, but it's much more specific, where they say, okay, their specific focus is on, what are the problems that we need to solve in order to generate this data, right? And so one problem that they identified is we need to figure out Lyft, okay? And so they went back and they requested all the data from Otto Lilienthal, who's the guy who had died in Germany, and looked at lots of the data and did careful analysis on themselves. And they quickly figured out, okay, we can't, we don't need to, you know, sort of test lift a scale where we risk dying, and so we can engage in creating experiments with wind tunnels and propulsion. They thought of the propeller as this kind of rotating, rotating wing. And so maybe that can solve the problem of rotation. They actually tried to work with existing engine manufacturers, but ended up having a bicycle mechanic. Wanted somebody that worked for them built their own engine for the plane itself, and then they looked at solutions for three for kind of the three axes, sort of pitch on role in terms of how to solve solve that. And there's details in the paper around that. And so now you have a very different logic for how to think about the role of data in terms of forward looking decision making, right? So here you have somebody who says, okay, we can actually engage in agency, puzzle experimentation. And I think where we go wrong a lot is we think, you know, kind of data is the answer, and there's lots of excitement around, evidently, decision making so forth. But again, data is only as good as your theory, right? And and your ability to kind of intervene in settings and so forth, make sure I'm managing my time here. So okay, I got eight more minutes here.
So the way that we're at a high level thinking about human cognition and think about the input output model that I was summarizing in terms of large language models, is the human is not just taking in information, but making decisions about what's the relevant data. So the arrow is going outward to data, information, and humans have theories and are have an ability, sort of causally reason and experiment. By the way, this is, this is a more broader biological point, and we have a quick footnote, and I just touch on this. And I've been in touch with a neuroscientist, Henry yen, who's at Duke around some of these issues, because it's not even that clear that sort of in terms of what humans are doing within their spheres, like the fruit fly, for example. So this is Heisenberg, not the Heisenberg in terms of physics. This is a son who actually, this is one of the last pieces he wrote, and he studied fruit flies for 4050, years. And he said, Okay, people have taught about thought about like fruit flies, of these simple systems information processing. He's like, they're nothing of the sort. Like there's really creative problem solving going on on the part of the fruit fly, in terms of within their kind of niche and in their environment. And so anyway, we have a quick nod to that, because we think that that's important as well, but, but the important point here is this arrow from thinking about how these theories and kind of contrary beliefs or, or, you know, beliefs that look delusional enable us to look toward the world or create, create data for, for, for novel uses, right? And so the pitch that we make in the paper and this links to other work that I'll cite here in a second, around thinking about humans as theorizing kind of, you know, scientists, and I've done a little bit of work with Stu Kauffman, who's a complexity scientist in Santa Fe, around these issues and how evolution has essentially given us mechanisms. And this shows up in places like hunting or having causal theories that enable us to find what's the relevant data, rather than sort of just being this input output information processing machine, which is a very different model of cognition, right? But more generally, like children do this, like Elizabeth Spelke and Allison gopnik have had fantastic studies around this and and so what we do is we create data and intervene the world through causal experimentation. And a lot of this work builds on joint work that I've been doing with for several years, with Todd singer, but also Alfonso Gamber della, Ronaldo camufa and Elena Novelli are doing work in this domain as well. I mean, interestingly enough, they're also, you know, we're looking at various types of AI, sort of supplements to human reasoning as well, and we can have a longer conversation, maybe, about so AI human hybrids, in terms of how each with their respective strengths might complement each other. But at this stage, given the current state of AI, I've looked at all the recent sort of causal AI models and the GEPA model that Hinton has come up with around planning and things like that, I would argue that there's still a unique kind of human Vantage here in terms of engaging in this, particularly in terms of forward looking kind of causal reasoning and experimentation, right? And so what we argue is Central is this, is this data belief asymmetry, which is kind of the root of new knowledge and value. And so asymmetric belief are central for new knowledge. So this is the Galileo example in the Wright brothers. But I'd argue, like I said, I worked in venture capital, and I often wrestle with, okay, how is it that I'm projecting into the future I only have data for now, and what mechanisms do we have as humans for making decisions that are forward looking, and it's only sort of over the last sort of, I guess, decade, in some sense, sort of, I've had a light bulb go off around, kind of the key processes around this issue. I guess an important point is that we're not, you know, this is not bounded rationality. This is not kind of processing information, how we're delimited in terms of that it, you know, like the Wright brothers did the right thing by ignoring some of the deaths in some way, like maybe they didn't ignore them, right? So they actually knew that Otto Lillian thought died in 1986 and or 18, sorry, 1896 so it's not that they ignored him, but the question was, what is relevant in that data point, right? And so it you know, if you take the LaConte data about small birds fly, and large birds fly. Is that like or the or that it seems that things that fly have feathers right, like, what's, what's the right thing to anchor on in terms of solving a particular problem? And we think that kind of this move toward intervention, experimentation is critical. And so what, what these beliefs can do is they can enable causal raising, about a path for realizing some kind of novelty. And there's a general sort of framework for kind of startups that we've developed with Alfonso and Todd around how to think about this, if anybody's interested, interest, interested, interested. But like I said, we're thinking about AI and human interfaces right now and right now as well. But that's an open access piece if somebody's interested in the space. But the question is, again, the right what's the right data versus, you know, thinking about just having more data, more data, more compute, versus the right theory, the right experiment and the right data right? So here's a quote by Einstein that captures this intuition in terms of seeing the right thing. Whether you can observe a thing or not depends on the theory which you use. It is a theory which decides what can be observed, right? Um, just quickly to Paul, provide some sort of context and from, from, from, from, from, from psychology. So Allison gopnik, with several of her of her colleagues, looked at LMS versus seven, three to seven year olds, and highlights how there's very different mechanisms, causal learning, tool use, understanding the application of physical properties and problem solving. Anyway, I'd highly recommend that piece if people are interested in more work in that space. I think there's practical implications for this data, belief asymmetry, idea, because we there's a lot of emphasis on Bayesian like right now, uh, Scott Stern, for example, has a heavy push around, kind of Bayesian entrepreneurship and things like that. But I think that you know, the evidence for the probability of something, it doesn't just kind of array itself, right? It doesn't sort of present itself. It needs to somehow be generated in a lot of instances. And we need to, need to know what's the relevant data. And we might even ignore data that looks to be conclusive in terms of saying that flight isn't possible, like the data used, you know, was for in the context of the Wright brothers. And, okay, I mentioned this point already, and I mentioned this point as well, so I think I did a couple minutes, and so I'm just gonna jump through and get to the key point. So just a quick I mean, I think that this plays out in lots of different contexts. Airbnb, for me, provides a nice example. So Fred Wilson, sort of famously, he released his emails around this, but he famously, in several emails, he's like, if he invested in Airbnb, it would have made, I think he said, somewhere, more money than all of those investments combined. But he didn't invest in it, and he felt like this is only for couch surfers and hippie hippies. Hotel Oxford experts experts seem to say that this isn't a viable business, and so they didn't invest. And so here's an example you should update, right? You're getting feedback data. If you'd asked me in 2007 No, I wouldn't rent my home to a stranger that's visiting from out of town. When I traveled to Milan, what I wouldn't strain. Would say, stay with training strangers. Now, retrospect, I actually do like mostly I, you know, I stay, I stay in Airbnbs because it's much easier with my kids than getting, you know, two or whatever, three hotel rooms.
And so no, there's no evidence of data, but they engage in some form of causal reasoning. He says, If we solve specific problems, think about Wright Brothers, lift propulsion and steering their case, trust matching and transaction escrow. This might happen, right? We're gonna, we're gonna, we're gonna create these experiments and interventions that might enable this, this outcome that we're looking for, right? I'm gonna skip through these examples just to get to the so I would highly recommend Felipe et al's paper, because it sort of provides the antithesis of what I'm talking about. I would argue that in terms of the kind of, the foundational, fundamental mechanisms of what AI is doing right now, in terms of prediction, I would say that's not strategy. I think that AI like can do strategy. If strategy is Porter's Five Forces, I've got a mistake there, using Python, Porters five forces, you know, framework, then, okay, it does a good job of that. Strategy is writing a business plan. It does a good job of that. But if strategy is about three or theorizing and kind of novel causal reasoning and problem solving, then not right now, at least right. And so I would push back on that. So I'll just summarize, just with a litany of key key points I've just gone over. And so my key points here are for now. AI is constrained by its training data. It's backward looking. It can't gage in causal reasoning. Again, there's models. I play around with all of these. I'm very suspicious of them, like the new AI scientist paper or the the MIT folks that recently published this piece has has lots of issues. And I'd argue that human, you know, cognition, isn't computation or prediction or even based on past data. And these data beliefs asymmetries, which are counter like this is the these look like conspiracy theories. They're delusions, but they're actually kind of central for new knowledge. But what we do is not just have these beliefs, but we can engage in some kind of interventions and experiments that enables us to kind of create value in some way. I think that there's be lots of new things that come out in terms of AI hybrid technologies that will be aI human hybrid technologies that will be central. And we're sort of taking, you know, the current state of AI, right? Who knows where this goes. And so I'm talking to lots of computer science scientists and colleagues around this. But like I said, this came out of Matthias, and I carefully, kind of teaching this AI class and thinking about, okay, the hype around this, and this isn't sort of hype just pushed by startups. This is, like, you know, really serious scientists, Jeff Hinton and Benji and, you know, others can we sort of look at that in terms of how that relates to decision making in humans. And so with that, thank you so much. If you have hard questions, you can email my co author. It's it's 1am in Oxford right now. And so he wanted to be here, but he couldn't, so he's sleeping. But anyway, I'm happy to answer questions, and I know that Andrea some Andrea has some thoughts as well. So with that, I will close my screen and stop sharing. Thank
you very much, Temple. Now I give the thought to Andre to give a quick discussion.
Yeah, absolutely. So I'm just gonna try to make this quick so I have lots of thoughts, mainly because I absolutely love the paper. So the way I would recommend everybody read the paper. The way I think about it is that the paper is basically well reason and saw the foundation that you can use to counter anybody that's an AI maximalist or alarmist who thinks that anything that we do as humans will be replaced by AI. Now, even if you don't agree with the key thesis of the paper, what tepo just presented, that there's basically something that human cognition does that is fundamentally, fundamentally inaccessible to AI, I think you'll still what's really cool about is it forces you to think quite deeply and carefully about what the cognitive processes are they're involved in generating new ideas and whether or not they can be programmed. So the key, and I think contrarian point of the paper, is that human cognition is not just data processing, as you know, many, like many people claim, and this is where so I have there are two things I want to do. I want to sort of spell out how I understand this based on the paper, and then tempo, you can correct me if I'm wrong. And now, essentially, what I want to do is basically build the best possible counter from the other side and sort of get typos reaction to that. So the way I understand it is the following. It's basically, I would say what we do that's different from data processing, essentially, is what I would call taking theory, lips leaps of faith, by going beyond or against existing data. So I would try to think like, okay, fine, but what is it exactly like? What are the steps that we take to do that? So my sense is that there's a few elements. Number one is that our perception of the world is not just words like tempo show like what an LLM does. Obviously, our perception of the world is much richer that sensory perception. There's lots of different things now that comes in handy, because it allows us to build intuitions and obviously come up with theories, basically based on associations, based on all the sensory inputs. And some of these theories are actually unsupported by data, and they represent in that regard, in that regard, complete leaps of faith. They call it intuition, so they still have bearing in the real world, but we don't know if they're true or not. And this is where the we have this capacity for biases, cognitive biases, which are normally a bad thing, you know, persisting despite prevailing evidence, engage in motivated reasoning. It means try to find evidence that supports what you already think. So these are usually bad. However, in some cases, they can be very helpful, because they push you to uncover something novel and true which conventional wisdom would say it's impossible or untrue. For example, human flight, right? So this doesn't mean obviously, to be completely irrational. It's just basically almost like selectively rational. It's like, once you have this contrarian theory, you still need to go through some sort of scientific process to select the data, which data to pay attention to, and what actions to take in the real world to actually create or validating or invalidating experiments. Now I think what's really interesting about this is obviously it's relevant to invention. So think of coming up with new theories like human flight or theory like Darwin's theory of evolution. And the other big one is obviously entrepreneurship, like startups that initially looks too good until they don't right. So initially, like a lot of reasonable people say Airbnb and others are silly. But then if you think about like you don't really have the data. So it turns out, when you actually have the data, turns out it's, it's actually a better idea. So yeah, I'm fairly, I'm completely convinced by the argument of the paper. But I think it's useful to sort of ask tempo to go, you know, to sort of counter very specifically. What I would imagine is a common counter from the other side. So the way I would frame that counter is the following. Is like, Why can't this what I just described is like, step by step cognitive process for coming up with new ideas for humans. Why can't we just program this into an AI? So let me be very specific. I can take a machine so I can take an AI or machines like, why can't I do basically the following? So first step, ask an LLM, for instance, to come up with some new theories. The constraints should be, of course, that the theories have some bearing on reality, and it's not but, but the theories, importantly, they shouldn't be constrained by existing data. And by the way, the latest, the chatgpt 01 that launched last week, is actually pretty good at coming up. It's kind of sci fi, but it's realistic sci fi. So you can definitely play that game. Step number one. Step number two, you can ask it, what experiments do I need to be to run in order to prove or disprove this theory? So I think it's, I mean, I would find it interesting to think about, like, why is that not kind of what we're doing, basically? So another way to put this is to say, well, when we come up with new ideas, either for inventions or for companies, all we do is, basically we follow our drift, like, you know, irrespective of current evidence, and we ask a lot of what if questions. Again, we don't really have the evidence. But then we start taking experimental action, prove or disprove the initial intuition. Why Can't We? I mean, it seems like something we should be able to program. And I guess what will be interesting is to like, pinpoint exactly. Why is that difficult? Or like, at least today, or like for the for the foreseeable future? The other point that I want to make is, okay, so if our competitive advantage is that we take these leaps of faith and data, well, where do we get the ability to take leaps of faith? Like, are our leaps of faith based on some sort of past data? I mean, granted the past data, but like, data for us is multi dimensional. Again, it's not just words, it's sensory perceptions and feelings and all that. But isn't it to at some higher level, still, data processing? So I think that's another question that I want to ask, and the last thing I would say is, so I'll start with what I think is my favorite quote from the paper, which is I wrote it here. There are times when being seemingly irrational, ignoring evidence, disagreeing about its interpretation, or selectively looking for the right data turns out to be the right course of action. So I think it's very nice, and I completely agree with it. But again, same question, why can't this be programmed into an AI using some sort of evolutionary mechanism where you built in some randomness, and then, just like genes, like mutations, you know, lead to mutation, leads to new genes, if the gene So, if the mutation is adaptive, adapted to the environment, then it stays. If not, it's basically wiped out. So you have to get feedback from the environment. So again, it's like, couldn't this be programmed into an AI? And then, fine, okay, final. Final point is, hippo maybe would be interesting to talk I'm sure that some people have heard about this new chatgpt, or one that I just mentioned, that came out last week. In particular, what I think is interesting is that it claims, some people claim that it now has reasoning capabilities, and interesting to see your take on like, is this like on the path to real reasoning that we're talking about, or is it still just better, more efficient prediction, as opposed to theoretical leaps of faith? So let me stop here. I had a lot more notes, but I can send them to you by email.
So so Andre sent me, like, two pages of notes that were, like, really on point so, and these are actually great, fantastic points, building on the notes that he sent earlier today. So this is really, like, super valuable to us. Let me just react to a couple of things. I won't be able to kind of react to all of them, but I appreciate you actually sort of highlighting that the bias point, because I used to teach kind of decision making and strategy and entrepreneurship as, like, I used to just, like, teach biases, and I would teach, like, 188 biases. And then it was like, and I thought, like, if you sort of fully understood all these you would actually just be frozen, like you would be frozen in time, and you wouldn't do anything, because the biases are just so severe, right? And so I thought, Well, my and the point to sort of anchor on here is that might these biases actually be like features rather than bugs, like confirmation bias turned out to be really useful in the context of the Wright Brothers, right? It looked delusional until it didn't. But that's exactly how I thought about venture capital, like we missed some investments to just look delusional, and then it didn't like somebody, they actually made it happen. And so there's something there that's, for me, really interesting, where it's sort of kind of, you know, motivated reasoning, wishful thinking, confirmation bias, and then you can make an
argument evolutionary state. But, like, basically, they're evolutionary adaptive, right? Like, those things actually can lead to like in some rare circumstances. But still, that works.
No, that's a powerful point, and I need to think about that, because I talk to people who do sort of the evolutionary kind of psychology and things like that, and I don't think that that's been kind of worked out, that, you know, most of the focus in particularly kind of evolutionary psychology that's shown up in economics, the stuff that was published in aer and so forth, says, Okay, we have this, you know, we we've had these problems in our in the past, on the Pleistocene Savannah, where we sort of when we were hunting and so forth and and we know how to solve these problems, because we've encountered them over and over and over, right? But the question is, like, how do we, how does somebody like Bootstrap and all of a sudden just do something differently, where they dig, like, I'll just give you an example of one evolutionary huge transition. At one point, somebody dug a huge pit in the ground and started to catch mammoths, where you get a lot more calories than trying to chase a deer like the people used to sort of persistent hunt, which was actually you would run until the deer, deer. Deer can't sweat. You would run until the deer would fall over and and so anyway. But there were these sort of innovations. And so what are the mechanisms through which where these come from? Is a key question. What your key point? I think, in terms of like, why can't we put this into computers? I guess we're sort of reflecting on the existing state of, of, you know, large language models, and this, you know, the jepa model that Hinton has, and, and, you know, all the, all the new models that are coming out. And your question was, okay, why can't we, sort of, you know, they can come up with new, new theories like, so, if the architecture is this, using past data, take the large language model for example. I actually haven't been able to get into the reasoning tool of of chat, GPT or open AI to understand kind of what's behind the black box. Like, and maybe I haven't been giving it to the right prompts either, because I haven't found it to, like, give me super novel and interesting things. The question is, can it sort of somehow come out of this correlate, you know, it's built on this kind of correlational matrix that it's that it's that it's sampling on that's based on the past, right? And so how would it somehow bootstrap itself out of that? And, again, it's using, sort of, it's being now more careful about kind of chain of thought reasoning, and, you know, different things like that. But I'm not sure that it does the leaps that you're talking about, per se. And I think that is central to human reasoning. And these leaps can be talked about in very sort of highfalutin ways, imagination or whatever. But I guess the key intuition for us is that, you know, there's actual causal mechanism and experiments and interventions that you can do. And I think this is, this is central for value creation, is, is sort of laying that path out. And again, we're trying to use this example of the Wright brothers as a convenient sort of retro retrospective story, where it worked out clearly, probably many examples of where it didn't work out as well, right? But I, you know, I, like I said, I use these models just as much as the next person. Alfonso Gamber della is developing with his he has some computer scientists working for him now, models that are trying to enable humans to do this better. So I think these human AI interfaces will be really exciting, but, but, but we'll, we'll see, well, I guess we'll, we'll see where, where, where it goes. I'm trying to think of what else to react. And I have so many notes here that
can ask on this, on this last point, I find it very interesting. So when you say so, there's the leap of faith, and then you have to be engaged, take action in the world, to experiments, if you have, like, which one of those do you think is more unique to humans and it's hard to replicate by Is it the initial leap of faith, or is it the fact that afterwards, okay, I just came up with this idea, and now I have to go and, like, validate it. Like, create some experiments, take action to validate it.
Yeah, some of these are almost like empirical questions. I mean, that's why I'm talking like Felipe's Azar and others, because, you know, he's having AIS generate business models and things like that. And you can imagine where there's some aspects of this, the humans are better, like, you can imagine an AI that's good at coming up with, like, tons of different but then how do you select from all those different options? Right? Then the selection is really critical, right? Somehow humans get to Michael Polanyi has this quote that humans somehow get to the right hypothesis, because we don't have time to go through all these millions of combinations. And so what is it? Charles purse, who's this logician as well, who talked about this as well, that humans has this kind of natural propensity that enables us to do this. But I think coupled with the computational abilities, or maybe the selection abilities of AI, or the ability to point to the right types of experiments of AI, I don't know, those are all things that are like the very forefront of the types of things I'm thinking about right now, because I think that that's that's critical. But I think a lot of this is, you know, you know, we'll see where things go in the future. And this article, as we went through this process with reviewers, they're like, well, this article could be dated tomorrow. Like it could be a dumb article tomorrow. Yeah. So we went the rounds with a, you know, with, with this just got accepted last week from by strategy science, but we're still editing it. So this, this, this, these points of feedback are definitely feeding into it, because we still have to turn in our final draft. But as we went the rounds with it, with the reviewers, they're like, Well, what about the next, next thing? And I'm like, Well, I don't know. Like, we're looking at this with, you know, we don't have any kind of crystal ball for the future. We're looking at the current capability, and we're looking at what it does. And this notion of using past data to predict a future, we think, isn't the architecture, you know, it's an amazing architecture, all kinds of things, but we don't think it's the right architecture for for novelty. You
