The Mysterious Math Behind LLMs - Anil Ananthaswamy - #537

WANTED: Developers and STEM experts! Get paid to create benchmarks and improve AI models. Sign up for Alignerr using our link: https://alignerr.com/?referral-source=briankeating One of the most powerful AI systems we’ve ever built is succeeding for reasons we still don’t understand. And worse, they may succeed for reasons that might lock us into the wrong future for humanity. Today’s guest is Anil Ananthaswamy, an award-winning science writer and one of the clearest thinkers on the mathematical foundations of machine learning. In this conversation, we’re not just talking about new demos, incremental improvements, or updates on new models being released. We’re asking even harder questions: Why does the mathematics of machine learning work at all? How do these models succeed when they suffer from problems like overparameterization and lack of training data? And are large language models revealing deep structure, or are they just producing very convincing illusions and causing us to face an increasingly AI-slop-driven future? KEY TAKEAWAYS 00:00 — Book explores why ML works through math 02:47 — Perceptron proof shows simple math guarantees learning 05:11 — Early AI failed due to single-layer limits 07:12 — Nonlinear limits caused the first AI winter 09:04 — Backpropagation revived neural networks 10:59 — GPUs + big data enabled deep learning 15:25 — AI success risks technological lock-in 17:30 — LLMs lack human-like learning and embodiment 22:57 — High-dimensional spaces power ML behavior 27:36 — Data saturation may slow future gains 31:11 — Continual learning is still missing in AI 33:46 — Neuromorphic chips promise energy efficiency 41:49 — Overparameterized models still generalize well 45:05 — SGD succeeds via randomness in complex landscapes 48:27 — Perceptrons remain the core of modern neural net - Additional resources: Anil's NEW Book "Why Machines Learn: The Elegant Math Behind Modern AI": https://www.amazon.com/Why-Machines-Learn-Elegant-Behind/dp/0593185749 Get My NEW Book: Focus Like a Nobel Prize Winner: https://www.amazon.com/dp/B0FN8DH6SX?ref_=pe_93986420_775043100 Please join my mailing list here 👉 https://briankeating.com/yt to win a meteorite 💥 - Join this channel to get access to perks like monthly Office Hours: https://www.youtube.com/channel/UCmXH_moPhfkqCk6S3b9RWuw/join 📚 Get a copy of my books: Think Like a Nobel Prize Winner, with life changing interviews with 9 Nobel Prizewinners: https://a.co/d/03ezQFu My tell-all cosmic memoir Losing the Nobel Prize: http://amzn.to/2sa5UpA The first-ever audiobook from Galileo: Dialogue Concerning the Two Chief World Systems: Ptolemaic and Copernican https://a.co/d/iZPi9Un 📺 Watch my most popular videos:📺 Neil Turok https://www.youtube.com/watch?v=Dt5cFLN65fI Frank Wilczek https://youtu.be/3z8RqKMQHe0?sub_confirmation=1 Eric Weinstein vs. Stephen Wolfram https://www.youtube.com/watch?v=OI0AZ4Y4Ip4?sub_confirmation=1 Sir Roger Penrose: https://youtu.be/AMuqyAvX7Wo Sabine Hossenfelder: https://youtu.be/g00ilS6tBvs Avi Loeb: https://youtu.be/N9lUceHsLRw Follow me to ask questions of my guests: 🏄‍♂️ Twitter: https://twitter.com/DrBrianKeating 🔔 Subscribe https://www.youtube.com/DrBrianKeating?sub_confirmation=1 📝 Join my mailing list; just click here http://briankeating.com/list ✍️ Detailed Blog posts here: https://briankeating.com/blog 🎙️ Listen on audio-only platforms: https://briankeating.com/podcast #universe #podcast #briankeating #intotheimpossible #science #astronomy #cosmology #cosmicmicrowavebackground #intotheimpossible #briankeating #AnilAnanthaswamy Learn more about your ad choices. Visit megaphone.fm/adchoices

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Speaker 00

  In the world of expense management, SAP Concur is here to help. From cost control slipping to travel chaos, SAP Concur has an integrated travel and expense solution to help you overcome problems like overburdened teams, decisions looming.

• 0:19Speaker 00

  When travel, expense and AP are connected with a single AI-supported solution, SAP Concur keeps you ahead of problems. Visit concur.com to learn more.

• 0:31Speaker 01

  Hey Spotify listeners. When you travel with American Airlines, you'll enjoy comfort and convenience at every altitude. Before takeoff, unwind in the Admirals Club Lounge with curated self-serve offerings of hearty bites and refreshments. In flight, stay connected with high-speed Wi-Fi. Sponsored by AT&T. Your journey is designed to feel easy from the moment you check in to the moment you arrive. No matter where you're going, American gets you there. Book now at aa.com.

• 0:56Brian Keating?

  What if the most powerful AI systems we've ever built are succeeding for reasons we still don't understand? And worse, they may succeed for reasons that might lock us in for the wrong future for humanity. Today's guest is Anil Ananthaswamy, an award-winning science writer and one of the clearest thinkers on the mathematical foundations of machine learning. In this conversation, we're not just talking about new demos or incremental improvements or updates on new models being released. We're asking even harder questions. Why does the mathematics of machine learning work at all? How do these models succeed when they suffer from problems like over parameterization and lack of input training data? Are large language models revealing deep structure? Or are they just producing very convincing

• 1:39Brian Keating?

  illusions and causing us to face an increasingly AI-slop-driven future? Thank you so much for joining us all the way from Bangalore. This is so exciting.

• 1:48Speaker 03

  Well, Brian, thank you very much for having me. It's a pleasure.

• 1:51Brian Keating?

  It's really a wonderful book. We're going to judge the book by its cover, as I like to do later on. It's entitled Why Machines Learn? And the first question I want to ask you, Anil, is I was taught as a physicist, you can never ask why questions. That's the first word of your title. What made you want to? Explore why and not how or what machines learn instead of why.

• 2:15Speaker 03

  It's funny, I I answered this exact question yesterday at a panel discussion about the very same doubts that people had. This is just a writerly conceit, I must admit.

• 2:29Speaker 03

  The, you know, the title came about because I, when I was trying to learn the mathematics of machine learning, I encountered very early on this amazing proof that uses very simple linear algebra to show that a single layer neural network, something called a perceptron from the 1950s, will converge to a solution in finite time if a solution exists. And, you know, in the late 1950s, the algorithm was first developed, which was essentially, it was essentially a simple neural network that could, you know, do linear classification. The algorithm is very simple, and that to me is the how.

• 3:10Speaker 03

  And a few years after the algorithm was developed, people started mathematically proving that the algorithm would converge to a solution in finite time if a solution existed. And to me, in, you know, in my head as a former software engineer, the math became the why. And of course, if you were to ask a physicist, they would just, you know, I, I, it's funny, because about a couple of months ago in Bangalore, David Gross was visiting the Nobel Laureate. And he had the exact same question about the title of the book. And I tried to give him my rationale, and he did not buy that one bit. He said, no, there's no why here. It's how. So, yeah, it's just a writer's conceit, you know, to me, how is the algorithm?

• 3:55Speaker 03

  And because the book is about the mathematics, and I feel like the math kind of gives you a rationale for why these algorithms do what they do. So that's how the title came about.

• 4:06Brian Keating?

  What was the first mathematical idea that you encountered in machine learning in research that you did on the book that made you stop and think that this is genuinely beautiful as I find it to be?

• 4:17Speaker 03

  Oh, it was exactly this perceptron convergence proof. So maybe we can kind of talk a little bit about how that perceptron came about, right? In the late nineteen fifty's, when Frank Rosenblatt, who was a Cornell University psychologist, he designed what was the first kind of artificial neural network. And it was a single-layer neural network. And, and like I just said, you know, the initial work was simply developing the algorithm and, and showing that it worked. It did pattern classification. It was able to take two categories of data.

• 4:58Speaker 03

  And if these two categories were linearly separable in some, you know, mathematical space, the algorithm would find the linear divide between the two clusters of data. And subsequent to the invention of the algorithm, people started mathematically showing why this was powerful and why this, this classifier even worked. And you have to think back to the nineteen fifty's, you know, when somebody gave you a mathematical proof saying an algorithm would find a solution in finite time if a solution existed, that was like gold dust, right? That was really, but when you look at the proof, it's very, very simple linear algebra, right? It is just manipulating vectors and matrices.

• 5:44Speaker 03

  There is, there is nothing more than that. And it is so beautiful. So that was the, that was the proof that made me kind of say, oh, hang on. Until then, I was just learning it for the sake of learning. I was not thinking about a book. I did not, you know, this was just me trying to get under the skin of machine learning to try and understand it for myself. But when I encountered that proof, it's, that's when kind of a light bulb went off saying, hang on, there are all these beautiful things that we should be communicating to readers. And so that set me off on a journey looking for other theorems and proofs that exist in machine learning that could then form the backbone of a mathematically oriented narrative, historical narrative of machine learning.

• 6:27Brian Keating?

  Why do you think it took so long for the tools and the techniques of machine learning, the mathematics, which as you say is very simple, I mean, I teach it to, you know, my undergraduates and even high schoolers that I happen to know. So how is it that it took so long for it to, you know, develop into this incredibly dominant part of our economy?

• 6:49Speaker 03

  Yeah, I mean, so the machine learning and AI that we talk about these days is what is called, you know, deep learning or deep neural networks. And these are very, very massive artificial neural networks, where a neural network is simply, you know, a whole bunch of artificial neurons interconnected together. And an artificial neuron, you can think of it as a computational unit. Some inputs are coming in, it does a weighted sum of the inputs, and if that input exceeds the threshold, it does something on the output side. So it does that kind of computation. And that's an artificial neuron and a whole, you know, bunch of these things interconnected together for a neural network.

• 7:33Speaker 03

  In the 1950s, when this started, the only way we could do anything with these networks is if they were single-layer neural networks, which meant that inputs are coming in, the neurons are there, and they do the computation and produce the output, right? Just one layer of neurons. The training algorithms that like Frank Rosenblatt had developed, or even somebody like Bernie Widrow had developed the least mean square algorithm, these things only work for single-layer neural networks. And the moment you put another layer, you know, after the input layer, the training algorithm didn't work. So the, and essentially, the moment you had multi-layer neural networks, the algorithm that they had to train the network was ineffective.

• 8:17Speaker 03

  And so that was in the beginning a stumbling block. So in the 1960s, people kind of realized mathematically that these single-layer neural networks actually were

• 8:28Speaker 03

  Good at what they were trying to do, which was linear classification. But they were really no good for anything that involved finding a nonlinear boundary between two clusters of data. And around, towards the end of the decade in the nineteen sixties, Marvin Minsky and Seymour Papert wrote this amazing book called Perceptrons in honor of Frank Rosenglatt, who had developed the first neural networks. And in that, in that book, they A had a very elegant proof for why the perceptron would converge to a solution. But they also had another proof, which showed that if the solution involved a nonlinear boundary, then the perceptron would fail at very, very simple tasks.

• 9:10Speaker 03

  And that kind of put a big damper on, you know, research, because people thought, oh, if it can't even solve this problem, which was, you know, just literally taking four data points arranged on the X Y plane such that a single straight line could not separate, you know, the circles from the triangles, the circles and triangles between being two different kinds of data. And the other thing Minsky and Papert did was they kind of insinuated without mathematical proof that even if you have multi-layer neural networks, they would still not be able to solve this, these simple nonlinear problems. They never, it was not a proof, which is very obvious now, because of course, these things solve nonlinear problems, but at the time, people kind of took that at face value

• 9:53Speaker 03

  and the research interest in this topic dried off, funding dried off. So, so neural network research kind of fell off a cliff during the nineteen seventies. That was the first AI winter. And, but there were other kinds of machine learning going on, non-neural network based, you know, machine learning techniques that were like the nearest neighbor algorithm, which was really popular for a long time. They were, you know, base class, Bayesian classifiers, all of this stuff was getting developed and studied and support vector machines, which came about a bit later. The real reason for why neural networks never really took off until, let's say, the next phase was the eighties, when people like Hinton and John Hoffield, both of whom got the Nobel last year for their work during the eighties.

• 10:38Speaker 03

  They they kind of, you know, uh, invig- reinvigorated interest in these uh neural networks. Hopfield designed what are now called Hopfield networks. Uh, they're not used anymore today for what we are doing with AI, but they were a big deal in the early nineteen eighties. Uh, and then Hinton, along with David Rumelhart and Ronald Williams. They wrote the first paper on the back propagation algorithm, or at least they put uh everything together to show how a deep neural network could be trained using something called the back propagation algorithm. So until then, we just didn't know how to train these networks. So that was a big sort of huge uh sort of gap between the early neural networks in the late fifties, all the way to the mid eighties.

• 11:22Speaker 03

  But even then, even once we figured out how to train these networks, it was still not enough because these networks are extremely data hungry, right? They require a lot of data to learn about patterns that exist in data. We just did not have the data in the in the eighties. Uh, and the other thing is that they're also very compute hungry. You you you need a lot of computing power to train these networks. And of course, we just didn't have that either in the eighties. Uh, so even though, you know, late nine, late eighties and early nineties, we had these things called convolutional neural networks that Jan LeCun and his team had developed. Okay, they went nowhere because of the lack of data and the lack of compute.

• 12:03Speaker 03

  And uh and traditional machine learning methods continued to flourish, you know, throughout the nineties and early two thousands. We had support vector machines, which were a big deal. Uh, and it was really the invention of the internet, the availability of extremely large scale data that you could essentially use the internet to collect all that data, uh, images or text or whatever. Um, and uh the realization that you could use the compute that was the the computing power that had been developed in the form of graphical processing units, GPUs. Everyone knows about GPUs today uh as being the backbone of what's happening with AI. But really, these things were developed for video gaming.

• 12:48Speaker 03

  You know, so, uh, you know, when you think about what, uh, uh, you know, video games need to do, they need to refresh your screen at a very fast rate and, and, and the screen is essentially a matrix of numbers, the pixel values, right? So they're very, very good at manipulating matrices of numbers uh, in order to, you know, refresh the screen for the purposes of, uh, you know, running a video game. Uh, and people like Hinton and, and many other people realize that, okay, we can, uh, you know, use these GPUs to do matrix, uh, manipulations, which are also the backbone of, uh, machine learning. So when, when a machine learning algorithm transforms an input vector into an output vector,

• 13:31Speaker 03

  the, the inside the black box, it's essentially matrix manipulations that are happening. So they, you know, so it was the, it was the combination of having enough compute in the form of GPUs and having the data that then in the, in the two thousands, late two thousands and early twenty teens, that things began to change dramatically.

• 13:49Brian Keating?

  Today's video is sponsored by my friends at Aligner. If you've asked AI a tough question and got back gobbledygook, that's not entirely the fault of the AI, but the frustration that you feel could actually be worth up to a hundred and fifty dollars per hour. Behind every AI breakthrough is a network of experts actually teaching these systems how to think. And my friends at Aligner are connecting brilliant people, mathematicians, scientists, engineers, geniuses, just like you, to make sure AI works for all of us. Aligner has specifically partnered with the Into the Impossible podcast to find geniuses from my network to give AI models expert feedback. Your job, if you accept it, is to evaluate AI outputs. That's it. Design problems that even today's best models can't solve. Your job is to grade their attempt at quantum mechanics, topology, advanced coding.

• 14:34Brian Keating?

  You're literally teaching AI the difference between right and wrong, undergraduate mistakes and doctoral level thinking. That's why they're partnering with me. Listen, I know that many of you have done unpaid internships, shall we say, been lab rats running someone else's experiments, but now is your turn. You don't have to grind as test particles in someone else's lab ever again. This is different. It can be done all remotely, timing is flexible, and you get paid weekly up to a hundred and fifty dollars per hour. Liner is selective. They need to be in order to get the best results, right? They only accept people who can genuinely push AI forward. Most applicants won't make the cut, so check out alignerr.com using my link below. AI has already consumed the internet and lightly wasted a lot of your time, as it has with mine, with incorrect answers, logical flaws, or poorly worked out solutions.

• 15:17Brian Keating?

  This is your chance to get it right for the future of science and to get paid while you're at it. Click the link below. And we'll get to, you know, get back to the kind of historical overview of it and even some of the nuts and bolts of how a perceptron works and some of the matrix algebra. You know, it's remarkable, Neil, there's a famous quote attributed to Stephen Hawking that every equation in your book cuts the readership in half. But if that's true, I shouldn't have even read this, but I mean, it's got over 400 equations and incredible detailed illustrations. It's really, it's sort of this hybrid between a textbook and a thriller, a historical thriller, and I just think you're to be congratulated for doing it.

• 16:01Brian Keating?

  I listened to it, which I don't know if I recommend the audiobook compared to the printed book. I really love the printed book. I'm actually giving it to one of my kids who's very, very precocious and wants to learn calculus. I figured maybe he can learn calculus from machine learning that you describe in this book.

• 16:22Speaker 00

  In the world of expense management, SAP Concur is here to help. From cost control slipping to travel chaos, SAP Concur has an integrated travel and expense solution to help you overcome problems like overburdened teams, decisions looming.

• 16:41Speaker 00

  When travel, expense, and AP are connected with a single AI-supported solution, SAP Concur keeps you ahead of problems. Visit concur.com to learn more.

• 16:51Brian Keating?

  But you mentioned, you know, this kind of, to paraphrase Marc Andreessen, you know, AI is eating software and software is eating the world. I'm going to talk about this phenomenon, which I've done a little bit of research on for fun for the podcast. It's called lock-in and I'm not sure if you're familiar with it, but I'll just describe what it, what it is. It's the phenomenon by which an early technology becomes super dominant, cannibalizes everything that came before it because it enables some new efficiency or new capability that heretofore didn't exist. And, you know, there's a couple of classic examples. One is the QWERTY keyboard, which is not optimal and it's not efficient from a human,

• 17:36Brian Keating?

  you know, from a frequency of words and typing perspective. But it was invented because the typewriters that were early, early adopted had this problem that the keys, the mechanical hammers would stick together if they were used to, to often next to each other. So they wanted to space letters apart that were, so that they wouldn't be pressed at the same time and you wouldn't have this lockup, not lock-in, but lockup. Another example is the, you know, the, the quality of the Hubble deep field image is, is great. It's breathtaking, but it could have been, you know, as good as, as the web telescope images, which are, you know, 10 times better, if not for the fact that the backside of a horse is about a meter across.

• 18:18Brian Keating?

  So when the Romans designed chariots to be pulled by two horses, that was set by the width of the horse's rear end. And because of that, the, the roads and the train tracks that later took precedence over the roads had a width of about two, two, you know, two to four meters to accommodate two chariots going back and forth. And because of that, and because of the fact that the space shuttle was built, its boosters were built in Utah and the launches were in Florida, they had to transport these massive rockets through train tunnels all the way from Utah in the US to Florida, which meant it had to go through a train tunnel, which meant it couldn't be bigger than a certain diameter, which meant that the specific impulse, the thrust couldn't be above a certain amount,

• 19:03Brian Keating?

  which meant they couldn't get to a high enough altitude that it could have taken a better image. Okay. These are examples of lock-in, that some early technology establishes the, basically dooms the future into, into this, you know, kind of irrevocable prison that it can't escape from. And I'm wondering, the success, this transition inflection point with L L M's plus G P U's, I'm worried it's another type of lock-in. And as successful as it is, I'm worried that we won't get the things that I'm most interested in, which, you know, new laws of physics and, and, and new descriptions of mathematical reality, et cetera. Do you worry about the, the success, not, not the, the, the failure, not the A I winters and

• 19:44Brian Keating?

  stuff, but do you worry about the, the summers being so bountiful that it will crowd out essentially any competing and possibly better technology?

• 19:53Speaker 03

  Yes, I, I think you're spot on, because if you, and, and here the lock-in weirdly is the incredible amount of data that we have been able to scrape off the internet, right? And also, and the presence of G P U's. Now, the, the G P U's, one can argue that they're just a computing element which haven't necessarily locked, locked us in. But I, I think this L L M revolution has been made possible because of this extraordinary amount of data on the internet, right? And, and we have managed to somehow create these models that are learning about, you know, the knowledge and the, the sort of syntax of human written

• 20:37Speaker 03

  language. And kind of, it's an intelligence that is imposed from the top down. These, these machines are not learning things from the ground up, the way, let's say, humans do, or animals do. And our general intelligence very much is a property of the fact that nervous systems have evolved over evolutionary time, and nervous systems have encountered things in their environment and have enabled the, you know, the development of brain structures and algorithms that operate in those brains from the ground up. And, and

• 21:14Speaker 03

  I hadn't thought of it in the way that you're framing it, but it makes complete sense that the economic incentives now to succeed in this arena is so high that there's so much money that is being poured into building these L L M's and they're getting bigger and bigger. People have bought into the argument that scaling up is going to unlock more and more quote unquote intelligent behaviors. So, yes, at this moment in time, we are certainly locked into this particular form of, you know, A I.

• 21:57Speaker 03

  So much so that I'm sure there are many, many, many smart people who otherwise could have been doing other kinds of research into, you know, different kinds of models that would potentially learn how to generalize better, be much more sample efficient like our brains are, use much less energy than these L L M's do, et cetera. And all of those areas of research have probably been kind of squeezed of funding because of the money that's going into developing L L M's. So, yeah, entirely, entirely possible that we are in a phase of lock-in because of this current trend.

• 22:34Brian Keating?

  And, you know, as I said before, to me, the greatest thing would be to get, you know, a theory of quantum gravity, you know, that no human has been able to come up with. And I want to draw your attention to a statement made by a different Nobel Prize winner. It's Albert Einstein, who said that his greatest thought, his happiest thought was that an observer in free fall would experience no gravitational force. And he literally said it gave him tingles up his spine, basically. And, you know, I wonder to what extent, and that allowed him to create the, you know, principle of general relativity and equivalence principles and so forth that we credit to him.

• 23:15Brian Keating?

  But I wonder, you know, can a computer experience a tingle down its spine? Conversely, can it experience pain? Can it have a Happiest thought. And if not, what does that portend for its ability to create new laws of physics that humans are incapable of creating with this, you know, three pound, you know, neural network, uh, pink neural network that we have in our, in our brains. What, to what extent, in your opinion, is embodiment is, you know, kind of unique human sensations, what we call qualia? Are those important for making breakthroughs that really matter to, to scientists, say, like me?

• 23:49Speaker 03

  Oh, that's a huge question, right? I think it comes down at some very basic level to what we think is human consciousness. And, you know, whether our intelligence and our consciousness can be thought of in materialist terms.

• 24:05Speaker 03

  You know, so, so for people who take the view that everything about our consciousness and intelligence can be explained eventually in computational terms. And even, even if it is computational, then the computation also is substrate independent. If that's the case, if, if, if everything that we are, and it's a big, big if, right? If everything that we are is something that can be boiled down to computational principles, substrate independent computational principles, then I don't see any in principle reason why machines cannot be built to perform those very computations and have the same kinds of experiences, et cetera, that we are privy to, right? But there's a big if there, you know, it's, and that's a huge one.

• 24:48Speaker 03

  Can L L Ms have those, again, I mean, a lot of this comes down to agreeing or disagreeing upon what we think is happening within us.

• 24:57Brian Keating?

  That's right. Yeah, I almost thought as I was reading this, I hope Daniel writes a book, Why Humans Learn.

• 25:03Speaker 03

  Yes. I mean, that's a big question for right now, for even machine learning people and computational neuroscientists, right? We don't have full-fledged answers to, you know, why we do what we do. So our intelligence, what kinds of algorithms are running in our brains, for instance, are, is everything finally describable in terms of computation? Even that question is not answered. The answer to your overarching questions about whether machines can, you know, eventually feel and have feelings the way we do hinges upon.

• 25:34Speaker 03

  Answers to questions about our own intelligence and our own consciousness. If if everything that we are can be talked of in materialistic terms, can be reduced to the workings of matter. And if, you know, if all of what we are is somehow captured by computations. And the computations have to be substrate independent, it doesn't require biology, it could happen, you know, in in silicon material. Then, yes, why not? And embodiment would be just another axis on which these machines would function. But without knowing the answers to questions about human intelligence and consciousness. It's really hard to answer what will happen with machines.

• 26:13Speaker 03

  I I don't think we are in a position right now to definitively say that we will be able to build machines that will feel. You know, and have conscious experiences. It all depends on our definition of consciousness. And then there are people even today who would say that, yes, machines are very definitively going to be conscious. And there are, you'll find as many people who will completely say, no, that's absolutely not possible. So I think it's an open question. Whether whether the whether conscious experiences are eventually necessary for the kind of breakthroughs that, you know, we're talking about. You know, coming up with the theory of relativity without having any prior knowledge of that stuff, you know. That I'm not so sure consciousness is necessary there.

• 26:58Speaker 03

  To me, they're orthogonal problems, like intelligence and consciousness are, you you can have them varying on orthogonal axes. So you could potentially have a system that is capable of coming up with something new. But have no quote unquote conscious experience of it, hence no joy, no pain, whatever.

• 27:17Brian Keating?

  What do you think is the most underappreciated and over, you know, kind of emphasized aspects of machine learning that you've encountered?

• 27:25Speaker 03

  Underappreciated. I think for me, after having written this book about the mathematics of machine learning, I, the thing that I find most fascinating and that is really underappreciated, and I think it's hard for someone who hasn't encountered the math to even appreciate, is the high dimensional mathematical spaces in which these machines operate, right? I mean, these are all, these machines are doing their thing in vector spaces. And, and it's extraordinary when you look at the dimensionality of, of these mathematical spaces in which these calculations are happening and, and the properties of these mathematical spaces that lead to the properties of these machine learning algorithms.

• 28:09Speaker 03

  That is really fascinating, but I don't know how something like that could be appreciated or even, you know, communicated without explaining a whole bunch of stuff about vector spaces and things like that. So there is something very beautiful that is happening in these mathematical spaces. And it's entirely possible that our brains are also functioning similarly, you know, navigating high dimensional spaces to do the things that they do. And to me, that's the most fascinating part.

• 28:36Brian Keating?

  And yeah, you mentioned this phenomenon of emergence, which is, you know, like the Supreme Court in America said about pornography, which is, you know, you know it when you see it, but it's, it's very hard to, to define how these phenomena really do come about. It, it is really was, was not truly clear to me until I read your book. And in terms of, you know, the details of how these algorithms work, but also the import of these, of the training data and how important that, really crucially important that is. You go over the, some of the restricted training data, you know, the US Postal Service data that was used for, you know, recognizing numbers and so forth. And then, you know, we don't look at the post office as the model of efficiency,

• 29:20Brian Keating?

  but it does do this incredibly well in optical character recognition and all sorts of other techniques that they pioneered, that you mentioned in the book and other countries as well. But it seems to me, you know, kind of this very strange phenomenon to be in that we, we've ingested most of the internet, you know, we have these huge, huge number of tokens and, and parameter models that you can put, you know, on, on your local desktop and soon on your phone will be, you know, not far behind if, if it's already not here. But but that, you know, what what is what is left to be ingested? You know, when I talked to Yann Le Cun last year, you know, he was saying, well, a cat, you know, can can take in, you know, four terabytes of data per second.

• 30:03Brian Keating?

  But, you know, if these if these algorithms are waiting for the next, you know, Avatar movie to come out, so it can ingest in more language and more data into its training set, if that's allowed even. Um, it seems to me like we're just going to slowly asymptotically converge to everything has the same information because there's only one internet out there. And yes, it's hard to characterize at all. Could it be that the very enabling feature of the success of these models will be its downfall because eventually there'll be no advantage? Everything will have the same data. We're all have access to the same internet and there'll be no advantage to any of these models and they should just all have the same outputs and, you know, given some predictive input. So what do you make of the, you know, the, the, you know, kind of, again, a lock-in

• 30:47Brian Keating?

  phenomenon that having all this training data was crucial, but now we're kind of saturated? And maybe that means we'll we'll asymptotically improve only only very slowly in the future?

• 30:57Speaker 03

  Entirely possible that because the lack of sort of freely available data is very obvious now. I think, you know, all that has been already scraped and taken in. There's a lot of data still locked in behind firewalls in, you know, within institutions and corporations and in private hands. And that's actually very, very high quality data as opposed to the stuff that we have scraped off the internet, which is relatively low quality data. But there's a lot of structured data that exists in company databases and institutions and that there is still value to be unlocked there. There's also this idea that these, we could have synthetic data generation, right?

• 31:41Speaker 03

  Now, that has the danger that we will end up sort of, you know, AI is generating data and then kind of, there's a very interesting, very evocative phrase that was used by someone, I forget who it was. They said that eventually these models will choke on their own exhaust, right?

• 31:58Speaker 00

  In the world of expense management, SAP Concur is here to help. From cost control slipping to travel chaos, SAP Concur has an integrated travel and expense solution to help you overcome problems like overburdened teams, decisions looming.

• 32:18Speaker 00

  When travel, expense and AP are connected with a single AI-supported solution. SAP Concur keeps you ahead of problems. Visit Concur.com to learn more.

• 32:29Brian Keating?

  I call it, sorry to toot my horn, but I remember the mad cow disease of the nineteen nineties in the UK when basically all meat was tainted because cows were fed cows. So I call it, I call it mad bot disease, you know, where they're taking their own data and then, you know, using it to regurgitate to something new. But I like the exhaust as well, but go on.

• 32:51Speaker 03

  Yeah, no, so you're, and this is a valid concern, right? People have this concern that maybe we are saturating and, but, but it's also true that if, even if we just continue the same paradigm of training on more and more data, there is still very, very high quality data that is available and we just haven't, you know, used those. And I may not, it's possible we may not be able to use them for, you know, publicly usable LLMs because this will be copyrighted data and private data. And there will be all sorts of concerns about, you know, privacy of the people whose data it is, et cetera. So I'm not sure it can be unlocked that easily, but there's, there is good data out there.

• 33:30Speaker 03

  My sense is that, and, and, and Jan Lekun is right about this, that there are ways in which animals and humans learn that there's something, we're doing something very different than LLMs. You know, we don't require, even though as a child or, you know, as a cat, we encounter a lot of data. There's a lot of structure in the, in the environment that we are encountering and there is something about the algorithms that we have that are operating inside our brains that are much more sample efficient. We just don't require that many examples of some instance of, you know, a pattern for us to learn about what it is. And then we are able to generalize so much easier, right?

• 34:15Speaker 03

  We, we learn abstractions about some, some problem and then we use the learned abstractions to then solve a problem in a completely different domain. And machine learning algorithms are not there yet, even these L L Ms, they can't generalize the way we do. So my suspicion is that even if L L Ms and the current approach saturates on this data problem, the breakthroughs might come in the form of new algorithms that learn very differently. And they learn continually, right? We know. So current machine learning models, especially L L Ms, they don't have this feature of continual learning. They're not, you know, you you train a model and then you you freeze that model. That's it. The weights of the model don't change after that. And you can use it as much as you want,

• 35:00Speaker 03

  but it is what it is. And you get a snapshot in time of the knowledge that it has ingested. And it's not a continually learning machine. And we are, of course, we are learning all the time. And even though when we learn new things, we don't mess up things we have already learned, or we don't forget the things that we learned before. Machine learning algorithms are not like that right now. So somebody is going to figure out how to come up with machine learning algorithms that are, you know, capable of continual learning, are more sample efficient, energy efficient, et cetera, and are able to generalize better. Then then the data problem will will be will not be as acute.

• 35:41Brian Keating?

  And what, you know, kind of alternatives, if you had to take, you know, the Schrodinger versus Heisenberg, you know, from your previous previous explorations in in physics and through two doors at once, what sort of, you know, competitors to the GPU plus L L Ms are there? Even if it's, you know, it's kind of the ninety eight pound weakling versus the, you know, the behemoth. What what's sort of the David to the L L M plus GPU Goliath right now?

• 36:10Speaker 03

  Yeah, I'm I'm not so sure the GPU part is really the the issue, because, you know, even any kind of computations that are happening in these machine learning models finally will involve matrix manipulations. So the GPU is going to be important. Whether you require as many GPUs for other algorithms, that's a different question. We don't know the answer.

• 36:30Brian Keating?

  Let me just break in. What about these tensor TPUs? What's their fundamental? You know, advantage or, or, you know, comparative difference between those and GPUs.

• 36:39Speaker 03

  And I, I have lack of knowledge there about the exact differences between TPUs and GPUs. I mean, they're still doing matrix manipulations, but the, you know, tensors are obviously a more general form of matrices, so they're manipulating these more general forms. I don't know the exact details about how a TPU works, so I would.

• 36:58Brian Keating?

  I believe in practice that, that, yes, I also am not, you know, incredibly familiar with it, but that, that, you know, Google has adopted the, you know, the tensor, the TPU approach and has used no Nvidia, you know, GPUs, whereas, you know, Nvidia is used by almost everybody and it's the most valuable company in the world and it has the stock market capitalization of, of all of UK and, and India and, and Germany put together. So it's, it's kind of astonishing that Google could be considered this, you know, kind of David, as I said before, but, but, okay. So, so then in, in terms of alternative model, you know, applications of ML, what, what are some alternatives to, to, I've heard of these things like Grok with a Q

• 37:40Brian Keating?

  and other, you know, kind of neuromorphic but not actual LLM. What are some of the, you know, kind of alternative algorithms that run on some form of matrix manipulating computational device?

• 37:52Speaker 03

  I, I think in terms of making these things more energy efficient, right now, when you look at these artificial neurons, you know, they are of course being simulated in software and so you have inputs coming into a neuron. It does some computation and based on the computation it produces an output. But in the context of a, a software simulation, the neuron has some real valued output that is always present, right? If you were to then implement that in hardware, that would be the equivalent of a neuron consistently having a voltage signal.

• 38:32Speaker 03

  On its output side, which means it's consuming energy all the time. Whereas our brains are what are called, you know, they have what are called spiking neurons. Where our neurons essentially collect information that coming through the dendrites. They do some computation and every so often or very, you know, very infrequently they'll fire. And that, you know, an occasional signal will go out on the axon in the form of spike trains, voltage spike trains. And a biological neuron for the most part is very silent. It's really not doing, you know, producing any output. It's just doing the computations but staying silent. And when it does produce a signal, it's a spike train which consumes very little energy.

• 39:11Speaker 03

  And we are now, just now, beginning to figure out how to build sort of artificial neural networks where the individual neurons are spiking neurons. And, uh, and then, once we have figured out how to train large artificial neural networks made of spiking neurons, if we then implement them in hardware through these so-called neuromorphic chips, then we can potentially have very energy efficient neural networks. Take a couple of orders of magnitude or more, uh, you know, in terms of energy efficiency. Uh, so that's definitely one thing to look out for. Uh, you know, you could still build L L Ms using that architecture, but it would be significantly

• 39:55Speaker 03

  lower in energy consumption, but we still haven't, uh, cracked the problem of how to build these things at scale and train them at scale. Uh, so that, that is one, uh, big sort of research area. The, the thing that I have been most intrigued by are efforts to get machine learning models to, um, learn about the environment in which they are functioning and, you know, so essentially learning models of the world, uh, in the form of abstractions. So they use, they kind of build abstract models of the world and then use those abstract models to make predictions about, uh, you know, what's happening outside. And this whole approach is how we think our brains work.

• 40:37Speaker 03

  Our brains, we think, work by constructing world models and situating ourselves. As agents, models inside those world models. And and then anytime we need to make a perception. Our brains are essentially using these world models to hypothesize about what might be there outside. That is causing the sensations that fall on our eyes or on our ears. And and and it's these hypotheses that we perceive as things that are out there. And then the brain has to do a whole bunch of processing over many, many layers. In order to make sure that what it is hypothesizing is out there is actually out there in in the form of. You know, making sure that the predictions it's making about the sensory consequences of whatever might be outside.

• 41:22Speaker 03

  Is exactly what was received by our senses. So there's a whole bunch of error processing going on, but fundamentally, you know, it has built these very sophisticated and complicated and abstract world models. And AIs that are beginning to do that might show us the way towards functioning more like the human brain. Then current L L Ms. So they also would potentially have the capacity to be more sample efficient, they requiring less data. Because when you think about our sort of cognition, our cognitive capacity. When you have a problem, you're not constantly waiting for external sensory data. You are, you know, capable of running internal simulations, counterfactuals, right?

• 42:05Speaker 03

  So we are essentially generating so much data internally for our own neural networks. So it's entirely possible that if we can figure out how machine learning models can do the same. They could also become much more data efficient. So that's something to watch out for because there are those things are going to do things differently than L L Ms.

• 42:26Brian Keating?

  Yes, right. Well, I promised that we would review the cover, judge the cover of the book. And now we'll do that. So we have a special jingle, which is generated by machine learning techniques that we'll insert here. We're going to judge a book by its cover.

• 42:54Brian Keating?

  All right, so Neil, so take us through the title of the book, the subtitle of the book, and the cover artwork, please.

• 43:00Speaker 03

  So the title is, of course, Why Machines Learn, and that was just a title, strangely enough, that just popped into my head when I was first conceiving of the book. It came about because I was learning about a particular algorithm called the perceptron, learning algorithm, which is used for training a single-layer neural networks. And as I was learning the math of why the algorithm works, it was the beauty of the math that made me think of, oh, there is a book to be written about why all these algorithms do what they do from the perspective of the mathematics. So the why was just my, you know, writerly sort of conceit, right?

• 43:45Speaker 03

  You could just have, just as he said, how, and it would have been a fine title. But the why seems to grab people's attention, right or wrong. And in my mind, it was more why than how.

• 43:59Speaker 03

  And the subtitle, again, it's just elaborating on this exact idea that there is a lot of very beautiful and relatively simple mathematics underlying this extremely powerful moment in time that we find ourselves in, right? And it's like, maybe high school or first-year undergraduate level, linear algebra, calculus, probability and statistics, and some optimization techniques, right? It's not at all, sort of, it's not the kind of physics that most sort of graduate students in physics or electrical engineering would do. They would do much more sophisticated math than what is required to understand.

• 44:43Speaker 03

  You know, again, there is a simplicity in the math for understanding how these machines or why these machines do what they do, but it's a very different level of math that you need if you are the one designing these algorithms. So that is a different ballgame, right? The cover art on the book is a variation of some M.C. Asher etching. Uh, so uh, completely due to my publishers. Yeah, so they, I think that is, I think it's an M.C. Asher etching called Three Spheres. And then they have gone ahead and added a fourth one and made it.

• 45:15Brian Keating?

  Yes, mesmerizing and you know, kind of reminiscent of uh, other curvilinear shapes and and things like uh, the 3D printed brain my kid made me. Um, in the book, you you emphasize something that wasn't obvious to me, but it seems, you know, kind of, if I were to set out on a journey to recreate machine learning techniques, I might stop because of this problem of what's called over over parameterization. And you make the case in the book that, you know, classically, there's there's something in, you know, classical statistical uh, analysis that that if you have an over over parameterized model, you should overfit the data and therefore your model should fail or or you're representative of it.

• 46:00Brian Keating?

  But deep learning seems to not only succeed but thrive on having more and more parameters. I mean, every week we're getting inundated with new models and foundations and and this number of billions of parameters and and soon it'll be trillions. I'm convinced it'll be trillions. So what's the least kind of hand wavy explanation for how this even works at all? Given the, you know, in classical statistics, that over parameterization should kill your reliability and therefore make it completely worthless. But in fact, it's it's one of the most useful tools ever created by humans.

• 46:36Speaker 03

  So I think mathematically, we are still trying to figure it out, right? You're right, the old statistical learning, machine learning techniques kind of made it very, very clear that if you over parameterize your model, you will end up memorizing it, which means you you'll end up memorizing your training data uh, or or fitting it. And then when you're when you're encountering new data, you won't be able to generalize to that new data. And uh, so people used to make sure that their models were optimally param parametrized so that you were not overfitting. And then along come neural networks. Uh, and we noticed this empirically, you know, so this is not something that was worked out theoretically. They just noticed that if they just made the, you know, networks bigger and bigger,

• 47:18Speaker 03

  the A, they work better and they really noticed that these things were not overfitting. Uh, the consensus, well, I don't, I don't think there's any consensus at this point about why it's, you know, it is still an active area of research. Why do, uh, deep neural networks, despite being heavily over parameterized, generalize so well as, as well as they do and, and the fact that they don't overfit? Um, there, there is, uh, some, uh, thought about that there might be some implicit regularization going on, uh, in these networks, that they do end up pruning themselves so that, uh, it's, it's not as heavily parameterized as, uh, as it seems at first blush.

• 48:00Speaker 03

  Um, but still, uh, you know, these, these networks have brought us into a regime of parameterization that was not the, uh, not the regime in which traditional machine learning function. And, uh, what has been very interesting is not that people have figured out why, uh, neural networks are doing what they do, they've started noticing that other traditional machine learning techniques, um, like kernel methods, uh, and, uh, you know, support vector machines combined with kernel methods and others, that also had hints of the same behavior, but they were never really pushed. Uh, uh, you know, early on, people just assumed that, oh, you know, over parameterization was not to be done.

• 48:45Speaker 03

  And now there are hints in earlier papers that if you go look at some early machine learning papers, they were seeing this behavior in non-neural network, uh, machine learning methods. And, but they were never explored. So what the artificial neural networks have done is they have kind of opened our eyes to the fact that there is this completely new regime of operation, which potentially even traditional machine learning methods could benefit from. And, uh, so now the math is being worked out and there is no clear answer to this yet.

• 49:15Speaker 04

  The one that asks my guests to judge their books by their covers. But today I'm asking myself to judge my own book by its cover. My newest book, Focus Like a Nobel Prize Winner, is chock full of advice, life tips, and focus and productivity tips from nine of the world's greatest minds, Nobel laureates ranging from economics to peace to physics, of course.

• 49:37Speaker 04

  I will go check it out, and my publisher's got an Amazon to run a special. So go to Amazon and get the Kindle copy today.

• 49:45Brian Keating?

  So another feature of this book is the incredible care and diligence by which you describe the nuts and bolts of how this field has come to be so successful and the mathematics of it. As I said, there's a thousand equations, there's hundreds of illustrations, there's interviews. It's an incredible book. As I said, it's sort of this hybrid new paradigm that's a blend between textbook and thriller, historical thriller, and kind of modern day application of these tools. But one of the heroes in the book is a technique called stochastic gradient descent. And I certainly wasn't familiar with it. I knew it was gradient descent. I've known about it since the time of Isaac Newton.

• 50:31Brian Keating?

  But the question of how it works so well, given that these landscapes that you describe, we can only visualize in the book two dimensional, three dimensional projections of two dimensional things. How is it possible in millions or trillions of dimensional landscapes that this S G D method works so well? First of all, could you explain it for the audience? And then how is it that they work so well and it's become to be this kind of the superhero of M L techniques today?

• 51:01Speaker 03

  Yeah, I think so. The high dimensional landscape you are referring to are what are called loss landscapes, that the error that a network makes.

• 51:11Speaker 03

  And it is an error as a function of the number of parameters. So if it was, for instance, one parameter, you would just have a curve, a 1D curve. But if it's two parameters, then you have some sort of Surface, three D surface. Uh, of course, these things have hundreds of billions, if not these days, close to a trillion plus parameters. So the the loss function is in some extremely high dimensional space. And also there are lots of nonlinearities in the network. So the the shape of the lost landscape is not convex. So it's not some sort of simple bowl shaped uh surface where if you start off at some high point on the surface,

• 51:54Speaker 03

  you can just do simple gradient descent and be guaranteed of coming, you know, finding the global minimum. And we don't even know if these things have something called a global minimum. So the these are extremely high dimensional surfaces with lots of hills and valleys. And the weird thing is, if you just did gradient descent, if you just went, you know, small step by small step down the lost landscape, trying to find the region of that landscape where the error that the network is making is very low, you might then end up getting stuck in some deep local minima and never be able to get out of it. Right? So stochastic gradient descent is this idea that you you kind of do a drunkard's walk down that uh

• 52:40Speaker 03

  down that slope. And you're taking steps not just always in the direction of the negative of the gradient, but you are taking steps that have some sort of stochasticity. And it's that stochasticity that potentially allows you to escape these local minima and end up finding what might be an optimal minima, even though we don't know if it'll find a global minimum or even if one exists, but it it does end up finding uh some sort of optimal minimum, which represents a state where the where the network is making a low enough loss so that a low enough error so that is actually functioning the way you want it to.

• 53:23Speaker 03

  So it's the stochasticity that seems to be allowing us to navigate this extremely complex lost landscape and escape local minima.

• 53:31Brian Keating?

  And the other thing that, you know, kind of resonates very highly is, of course, the perceptron. I think that is the, you know, main character energy of of this book. Can you give a some sort of? Description for maybe a layperson of how these things were conceived of and what they fundamentally do in terms of simplifying, you know, these massive, you know, kind of data sets or whatever into a tractable problem, maybe not always soluble, but at least tractable through very simple mathematics. But what is the fundamental job of a perceptron? I viewed it, you know, before as sort of this, you know, kind of black box, literally,

• 54:14Brian Keating?

  you know, black box. But now I see it more as kind of the transistor, the qubit, the element of ML. So can you describe that for the audience and whether or not you think that we'll still be talking about them and using them in 50 years from now?

• 54:28Speaker 03

  So the perceptron is the name given to the first artificial neural network, right? And it was Frank Rosenblatt who designed this artificial neuron. And the artificial neuron is very, very simplified version of what we think is happening in our biological neurons. So the biological neurons have a whole bunch of inputs that come in through the dendrites, the biological neuron does some computation, and then based on the results of that computation, it produces an output. And you can think of this same thing now implemented as a piece of software, which is what a perceptron is. You have, you have, imagine, you know, a circular figure, which is the body of the artificial neuron, inputs

• 55:13Speaker 03

  are coming in. Let's say you have, you know, three inputs coming in, X1, X2, and X3. And what the artificial neuron does is it basically does a weighted sum of these inputs. So each of these inputs has associated with it a strength or a weight, like W1 for X1, W2 for X2, and W3 for X3. So it'll do a weighted sum of the inputs. And if the weighted sum exceeds some threshold, it will output a plus one. If the weighted sum is less than a threshold, it outputs a minus one. This Was the computational unit that was the perceptron, right? Now, and it was amazing that something this simple could then be used to do,

• 55:58Speaker 03

  for instance, classifying two sets of images into images that are cats and images that are dogs. Right? So, think about, think about 10 by 10 images, for argument's sake, where you can, these images are black and white and they represent either, they are either images of cats or images of dogs. Then some human has painstaking, he looked at these images and said, oh, if it's a cat, we're going to call it plus one. If it's a dog, we're going to call it minus one, right? Now, you take each one of these images. If it's 10 by 10, that means there are 100 pixels. You turn the image into a, a single vector that is 100 elements long, where each element of that vector represents one pixel of information.

• 56:43Speaker 03

  You feed, you feed these 100 pixels into the perceptron. So, now the perceptron, instead of having three inputs, is going to have 100 inputs, because it, there are 100 pixels coming in. And you're training the perceptron to learn the weights that are necessary in order to take a certain image and output a minus one or a plus one. Right? And as long as these images are separable in 100 dimensional space, into cats and dogs, where cats are in one part of the 100 dimensional space and dogs are in another part of 100 dimensional space, and there's a clear gap between the two in this mathematical space, the perceptron will find the plane, hyperplane,

• 57:26Speaker 03

  that is capable of separating the dogs from the cats. Right? And then, when you have a new image and you want to know whether it's a cat or a dog, all you have to do is take the image, linearize it into the 100 pixels, feed it into the perceptron that you've trained, and it's going to say, oh, this is plus one or minus one. It doesn't know dog from cat. All it knows is, this side of the hyperplane, I'm going to call it cats. This side of the hyperplane, I'm going to call it dogs or whatever. And then,

• 57:53Speaker 03

  This was the beginning and even today's neural networks are just slightly more sophisticated forms of the artificial neuron that Rosenblatt came up with in the 1950s. And they just have additional elements that bring in nonlinearities and allow you to train multi-layered neural networks and things like that. But in essence, they are still simple computational units that are nowhere near as complicated as water biological neuron is. And yet they do amazing things because of the fact that we can interconnect hundreds of billions of these things together. Will they be there 50 years from now? Oh my, that's...

• 58:37Speaker 03

  I don't see why not, because we have an existence proof of a machine that does something really well and which is our brain. And our brain, one thing we can definitely say about our brain is it is made up of a whole bunch of neural networks. Right? I mean, there are 86 billion neurons in there with 100 trillion connections and even a very simplified model of that is a very complicated neural network. And it's obviously doing amazing things. So no reason to think that neural networks won't be around. But we might come up with ways of interconnecting these neurons that are very different from the ways we do it today. So the architecture of these neural networks might be very different 50 years from now.

• 59:22Speaker 03

  But the idea that we'll have neural networks, I think they'll survive.

• 59:25Brian Keating?

  Oh, yeah.

• 59:26Speaker 00

  In the world of expense management, SAP Concur is here to help. From cost control slipping to travel chaos, SAP Concur has an integrated travel and expense solution to help you overcome problems like overburdened teams, decisions looming.

• 59:46Speaker 00

  When travel, expense and AP are connected with a single AI-supported solution. SAP Concur keeps you ahead of problems. Visit Concur.com to learn more.

• 59:57Brian Keating?

  So I know it's late there, but if you'll indulge me with sort of a two-part question where the first part of the final question is just a relatively rapid-fire question, which is Richard Feynman, you probably know, was asked, you know, like, what is the nature of a scientific model of reality? And he gave an example where if an alien species, you know, looked at the Earth, the planet Earth, with its atmosphere and with the water cycle and so forth, it would, if it were, had all the knowledge of the laws of physics, it would know that we have this phenomenon called rainbows, right? And he basically said, if you're, you know, understood basic physics, Maxwell's equations,

• 1:00:42Brian Keating?

  and, you know, at a high level, as we do, you could make predictions just from observations of the basic ingredients of a system. I want to ask you, before I turn to the final question, which will also be about this phenomenon, do you think, like, a smart alien species, you know, looking at LLMs plus GPUs plus machine learning plus all the great stuff you write about in this book, could predict that these models would hallucinate and that they'd be sycophantic? And I'll tie that into one of your earlier books in just a second, but do you think it was inevitable, in other words, that these things would have these pernicious, in some sense, and very dangerous phenomena, potentially, of hallucinating?

• 1:01:26Brian Keating?

  You know, where I asked it recently, you know, what books has Brian Keating written? And it said, you know, Losing the Nobel Prize and Into the Impossible and A Brief History of Time. And it's like, well, that's nice. I wish I had, you know, a couple percent of the book sales of Stephen Hawking, but I don't. So tell me, would a, you know, kind of an intelligent alien looking at these models and so forth, would they be able to predict that they would eventually come to have these pernicious phenomena like hallucination and sycophancy?

• 1:01:55Speaker 03

  I think so. I think if you, if, you know, assuming that the aliens can look at the math, which if they are smart, they will obviously be able to. It's no different from us knowing, you know, if you look at the math, it's very, very obvious why they're going to hallucinate, right? I mean, these, these next token prediction machines are essentially probabilistic. They're, they're, they're always trying to generate a probability distribution over their vocabulary to say, what is the most likely next word? It's, it's not a hundred percent certain about what has to be produced. And it has learned about patterns that exist in data that is not, you know, a definitive amount of data for any particular problem.

• 1:02:37Speaker 03

  And the way these things are constructed, they will always output something that they think is the most likely one. And right or wrong, right? I mean, it's just the nature of the beast. It's just constructed in a way where, yes, if you look at the math, it's so obvious that, you know, the, I find the word hallucination itself problematic because the procedure that generates correct answers or answers that look correct to us is exactly the same procedure that results in hallucinations. So there should be, anyone who can peek at the math will immediately say, yes, of course, these things will hallucinate.

• 1:03:13Brian Keating?

  Yes, and in some sense, you know, as I said, it could be useful for me, you know, to at least to buoy my confidence. But, but of, of course, we do want things to, to, to give us factual information. But of course, to the extent that they mimic the human mind, you know, this is perhaps, perhaps inevitable. So I want to follow up with a final question. You've been very generous with your time in a late evening there in Bangalore. In The Man Who Wasn't There, you, you write about these patients, these who exhibit. These sorts of very strange phenomena and including what you define, I think, in that book as maladies of the self. And these are confabulations and hallucinations, not

• 1:03:57Brian Keating?

  as much about sycophancy perhaps as LLMs are prone to, but but they lose a sense of themselves. And I'm wondering, has, you know, the explorations that you've done in LLMs, have that, has that kind of refined the way that you think about the way the human mind works? And we kind of mentioned this, and my hope is that you'll write a book about human mind, but you kind of did earlier. So what have you learned about the human condition, these these unusual traits that, you know, you you talk about in in your in your earlier book, that that that these, these, by the nature of of being complex systems that have emergent phenomena, that you you sort of will get strange behaviors,

• 1:04:38Brian Keating?

  maybe even worse than hallucinations and and sycophancy, maybe true disorders, maladies, true maladies of the self. What do you make of that as a kind of a learning that you encountered in writing this new sec, this new book, you know, after having written this incredible book, The Man Who Wasn't There?

• 1:04:56Speaker 03

  I think for me, the writing of The Man Who Wasn't There was very instrumental in making me think of what is happening within us in computational terms. It's kind of, when you view our perception of our bodies, of our cognitive selves, et cetera, through the lens of the brain creating models of the environment, models of us embedded in that environment, using the models to make predictions about what is out there, including making predictions about our body, and the fact that what we perceive at any moment are predictions that the brain is making. Once you view everything in that, within

• 1:05:36Speaker 03

  that framework, it again becomes very clear that while on average, and most of the time, the brain is doing what it's supposed to be doing, and whatever we are perceiving is more or less congruent with physical reality, so we are not being, we're not hallucinating, we're not being psychotic. The fact that it is a computational process and the fact that there is stochasticity in that process means that these computational systems are prone themselves to making wrong predictions. And because what we perceive is the brain's prediction at any given moment and we take

• 1:06:16Speaker 03

  that prediction to be, you know, real and truthful, even if the prediction is wrong, it will feel like real to us. So it's very easy to understand why we end up having states of psychosis or states of hallucination. So now, when you think of what's happening with machine learning models and the fact that we are seeing some of these processes, you know, in very minor ways being duplicated in machines, it starts, the connections become more and more obvious that, you know, we might even end up building machines that function like us, which will themselves be prone to psychosis, which will be no different than, I mean, right now, we complain about the hallucinations that an L L M makes because it represents wrong answers to us.

• 1:06:58Brian Keating?

  Right.

• 1:06:59Speaker 03

  But imagine building a machine that is using its internal predictive mechanisms to understand its own state and its behavior in the world. And if those predictions about its own state are wrong, it is essentially hallucinating about itself. We're not too far away from building at least simple versions of such machines. And, you know, I don't even want to imagine where that will go. But the parallels are pretty striking between, if we take a computational view of what's happening in our brains and the things that we're doing when we build these machines, the parallels are striking.

• 1:07:34Brian Keating?

  Anil Ananthaswamy, thank you so much for writing this wonderful book. It's really one of my favorite books. I only regret that I didn't read it earlier and I have interviewed, you know, dozens and dozens of people from both, you know, pro-AI, anti-AI and L L M's, but understanding the details behind, you know, what's underneath the hood was a real treat. And you approach it as you do with all your writing in such a beautiful, eloquent and careful way that I just can't. Thank you enough for this and the opportunity to interview you and for you to stay up late on this late December evening for you over there in Bangalore. Thank you so much. This has been a real pleasure.

• 1:08:16Speaker 03

  Well, Brian, thank you very much for having me on your podcast. It's been my pleasure entirely. Thank you.

• 1:08:20Brian Keating?

  If you enjoyed this conversation with Anil, you'll want to check out the follow-up interview I did with Jan LaKoon, where we tackle many of these same questions, but from the perspective of someone building the systems themselves. Jan's no AI doomer. That episode is linked right here. Watch it right now. And if this conversation helped sharpen how you think about AI, not just what to believe, but how to question it and how to understand what it's actually doing, then please do me a favor. Like this video, subscribe to the channel and leave a comment with a question you think the AI community is still avoiding. I read them all. See you in the next episode.

• 1:08:53Speaker 05

  What's the secret to being a great host this season? Keeping things deliciously simple, like mixing Simply Cranberry into a festive drink as guests arrive, or serving up a warm mug of Simply Apple on a chilling night. That means less time behind the scenes and more time in the moment with everyone. Crafted with real fruit, Simply juices and fruit drinks make any holiday drink effortless, inviting, and completely delicious. Simply. Say yes to simple. Stonefire Naan is made with real buttermilk and ghee. It's the perfect flatbread for any

• 1:09:27Speaker 06

  occasion. I put out Stonefire Naan with dips at every dinner party. I make naan grilled cheese when my daughter has friends over. I use Stonefire Naan to reward the humans who do my bidding. Right. It's perfect for camping trips. Dance for it, feeble mortal. Sorry, is he talking to me or? Dance.

• 1:09:48Speaker 06

  From our tradition to yours, Stonefire. Together we Naan.