Professor William Hahn is an Associate Professor of Mathematical Sciences and a founder of the Machine Perception and Cognitive Robotics Laboratory, as well as the Gruber AI Sandbox. Both you and I, Will, we met at MindFest at Florida Atlantic University a few times, and a link to all of those talks on AI and consciousness are in the description. Will, please tell me, what have you been working on since we last spoke? Well, first, just want to say great to see you and really happy to be joining you on TOE today. Really excited. You've got such an amazing community. Same, man. It's been a long time coming. Thank you. I'm working on a whole bunch of different things. The
thing that's been in my mind the most is this idea of info hazards. And in particular, this theme I've been bouncing around called lethal text. Okay, let's hear it. Um, well, so as everybody knows, you know, AI is here, and everybody is kind of prepared for the technological revolution that we're witnessing. But I think the more interesting developments are actually going to be in our mind.
They're going to be the changes in language, how we think about language, how we think about ourselves, and how we think about thinking. How we think about language. What do you mean? So everybody, I'm sure, has gotten their hands on one of these large language models at this point. And they have just absolutely revolutionized the way we are thinking about words, the way we're thinking about language. And as people might be aware, it's now becoming possible to
program a computer, largely in English, that we can ask for computer code at a very high level, things people dreamed of back in the 50s. And now it's possible to just describe what you want the computer to do. And then that behind the scenes is getting converted into runnable computer code. But I think that now forces us to think about was language always a programming language? Is our mind something like a computer? Not in the obvious sense of transistors and gates and that sort of thing, but is it a programmable object? And if so, how is it programmed? So where do you lie on the is a brain a computer question? I think the computer metaphor is probably the most powerful that we have so far for understanding the mind. And what's interesting is if you go back through the history of technology, every time there was a revolution in the mechanical world, let's say we adopted a new metaphor for how the mind might operate. And so in the ancient world, it was dominated by a clockwork universe, the idea that the world was made out of cogs and gears and things like that. And then later we saw things like the emergence of telegraph networks and switchboards. And at certain times we saw the emergence of things like steam engines.
And we actually still have this thermodynamic hydraulic view of the mind is still residual in our language. We talk about people being hotheaded and have a head full of steam and they need to cool down and so on. And we still use these sort of thermodynamics metaphors. And a lot of people would argue, well, the computer is just the current metaphor. It's the metaphor of the day. And that will right. We'll change it as we go on. But the thing about computers that that Turing showed is there's a kind of universality that computation is the limiting result of any technology. If you take your car and you make it sophisticated enough,
it turns into a computer. If you take your house and make it sophisticated enough, it turns into a computer and so on. That almost every technology, if you improve its capability and the sophistication, eventually you're going to run into this notion of universal machine. And so the idea that the mind approximates the universal machine of Turing, that it's a machine that can simulate any other machine, given the appropriate programming, I think is something we need to consider. So what unifies clockwork, telegraph networks and thermodynamics is computation? Exactly. We can all see those as intermediates, as sort of proto computers or different aspects of communication and computation. And that the end result, the limiting result of all of those would
be the computer as we know it today. There are different computational models of consciousness. Are all of them the same to you or do you see pros and cons with different ones? You know, there's so many and there's probably a new one invented every afternoon. A few flavors that I'm a big fan of. And, you know, I like the saying all models are wrong, some are useful. And so I don't think any of these will act ever actually capture the full scenario, but they're sort of the best that we have right now. And let's be specific. Let's pick the one that is your least favorite
and then the one that's your favorite. Well, one of the ones that's my favorite is the idea of society of mind. Um, Marvin Minsky's proposal that the mind is really a collection of, you know, he, he threw around, he threw it a number about 400 agents. I don't think the number is particularly important, but the idea is there's a bunch of them. And what's interesting is we're starting to
see that emerge now with these language models that in the background of the newest ones, they've actually bifurcated themselves. And there's a dozen little micro agents, each with a separate prompt, a separate goal, a separate unique way of looking at the world. And then they have a conversation in the background. And then when they make a final output, it's kind of a consensus amongst those agents. And I think that's probably a good approximation for how our brain works in that we have all of these competing agents, and some of them are trying to meet new people. Some of them are trying to find something to eat. Some are trying to see visual, interesting visual stimuli and so on. And that when we choose a behavior or have an action, even like, you know,
producing a sentence, it's probably the result of multiple of those agents coming together. I liked, uh, you know, Minsky takes us a step further with the idea of emotion. And I think a very interesting take is emotions, not really a thing. It's the absence of certain things. It's turning features off. And he describes that when you're, when you're hungry, for example, that your ability to long-term plan or to even think rationally gets turned off. And you're just,
you're very hungry. When you're angry, your ability to care about other people's feelings and consider their viewpoint gets turned off. You know, you're no longer running that agent. You're sort of in a dynamical, um, ensemble prioritizing these different agents as we go through these different emotional states. And so I think that's an interesting way of looking at our behavior. And I think we're going to need those kinds of theories when we try to put, you know, intelligent behavior into machines, as I think we're seeing going to see right around the corner. That sounds to me more like an explanation of mind or the mechanics behind mind and not an explanation as to how consciousness comes about from computational systems. Yeah. Um, you know, I've got a lot of ideas in that and a lot of them are, you know, at conflict, I like to tolerate ambiguity. And so I
have a few of these ideas that, uh, I like to just kind of keep, keep juggling around. And one of the things that comes to mind is I really like Sidney Brenner's approach. Um, the molecular biologist, and he had this really interesting take about consciousness. He said that the discussion is going to go away. He said that in a few decades, the idea of consciousness will kind of just
disappear from the scientific conversation and that people will wonder what we were talking about all along. And I really liked that idea. I don't know if I believe it or even want it to be true, but something about it resonates with me because I think we're going to start to see something like proto-consciousness or something that will be more convenient to describe as consciousness in machines. And we're going to force ourselves to consider the hard problem and other aspects that, you know, plagued philosophers for so long. They're going to be laid out in front of us in a very concrete way. And the, the great minds before us didn't have the opportunity or rather they didn't have the language of objects like LLMs or bits or, you know, computational process.
They didn't have those, that terminology for which to frame their thinking. And, you know, one thing that comes to mind is this classic question of the redness of red. Well, we're going to build machines that will probably be able to talk to us in natural language about the infraredness of infrared or the ultravioletness of ultraviolet. That we have such a narrow perceptual
window and cognitive window that when we talk about consciousness, I tend to think of it as sort of a spotlight that moves around, but with such a narrow beam, it almost be more like a laser pointer because if I'm, if I'm conscious of red, well then I'm not thinking about my TOEs. And if I'm thinking about my TOEs, I'm not thinking about my childhood. And if I'm thinking about my childhood, I'm not thinking about the future and so on. That kind of like how vision saccades around the world, our consciousness also sort of jumps around and saccades and we get this kind of holistic picture, but it also is fleeting and constantly changing the subject of that, that Cartesian theater, if you will. And so, you know, I'm fascinated by how we're going to
expand that notion by looking at machines that have lots of sensors that have internal states that are thinking about their thinking before they answer in English. And we're going to be able to ask them, well, what do you think about red? And it's not that far away before they will be able to have at least consumed a strawberry in a rough sense, right? We have elaborate olfactory sensors. You know, it makes me think of, we know what ramen soup tastes like, but I don't know what ramen scattering tastes like. You know, they have these little handheld machines that measure the
vibrational mode of molecules and you can detect the presence of chemicals without opening the jar. If we put that into a system and give it a large language model and a rich historical experience, and it will remember the first time it encountered strawberries and its states when it did so, who are we to say that that's not a conscious being in some sense? Okay. Plenty of this depends on the definition of consciousness. And I know that that's an implicit problem with the hard
problem. So how do we define consciousness? Something I put out on Twitter recently was, is awareness a necessary condition? A sufficient condition? Both or neither of consciousness. So what would you say? Yeah, I think awareness is definitely going to be a necessary condition. And
I think you're going to have to have awareness of awareness. Some sort of metacognition where the system knows it's not just thinking, it knows that it's thinking and it's able to think about its thinking. That's tricky then because we could then say some animals are not conscious because they're not self-conscious. What do you say to that? I imagine you can feel without thinking about your feelings. Yeah. I mean, I think that's what's just so interesting is trying to parse out
those distinctions because they certainly have feelings inside in some sense of a sensory loop. But whether they're aware of that, it's not obvious. Or at least they're aware of it at the first level but they're not aware that they're aware. And I don't know if we are. I don't know if I am. Certainly most of the time I think I'm not. There's just not enough extra processing power. I think maybe it's just because of our daily lives consume so much of our brain power. If we were like sort of the philosopher sitting on the sofa, we could just like an ancient world, I mean, we could have more access to that. And that's one thing I've been very interested in is going back
to the ancient world and looking at how people thought about things. Because our modern world is just so inundated with certain things that we have to think about all the time. We don't get much sort of bandwidth to think about the thinking. I think that's what's great about your channel. You force people to do that. Thanks. Well, that's also what's not so great about the channel. So you said you could be aware of something but not aware that you're aware.
That also reminds me that you can know something but not know that you know it. I think it was, I think it was Schopenhauer said a man can or a person can do what they will but they can't will what they will. Right. And so we think we have this freedom of choices and action. But where do those, you know, are there agents in there that are choosing those behaviors? That's one of the things I've been very fascinated about is this, this idea of our mind being hijacked by systems that are choosing our behavior below our threshold of awareness. So there's, there's a classical
psychological experiment where you can sort of puff a airstream into someone's eye to make them blink and you can in an associative training, uh, get that to match with a stimulus like a little red light turning on. Interesting. And so people like a Pavlov dog with the bell, they can learn to instinctively close their eye when the light goes on because they know that the air blast is going to come on. But what's interesting is you can get people to learn that association and they have no idea they've learned it. So it's sort of a completely unconscious, uh, programming. Now imagine this would be very powerful in marketing, right? You show someone a logo and they want to go out and buy a bag of chips. Are we susceptible to that sort of thing? And I suggest that we are and that maybe that's just a general phenomenon that maybe a large percent of our percentage of our behaviors are chosen at a level that which we don't have access to and would take a lot of work if at all possible to get access to. Earlier you talked about emotions can be not on switches but off switches. And in one respect that's odd to me because there's much more off that you would
have to turn. There's many more switches you'd have to turn off than you'd have to turn on. So to conceptualize it as an off model is odd to me. Exactly. It's akin to saying the electron's not an electron, it's an off of the quark and the photon and the so on. Like, okay, or you can think
of it as an electron. An on of an electron. But it doesn't matter. So you can feel free to justify the, I believe it was Minsky who thought it was off. You talked about something else being off. So then it makes me think, do you think of free will as not free will but free won't? And that's one of the ways that we can save free will. Yeah, that's an interesting way to think about it. That
we maybe we don't choose our behaviors, we choose other things we wouldn't do. And that, that, you know, gets to my idea that I'm been thinking about a lot lately as this idea of immune system and how it relates to mind and consciousness. And it started by, I was looking at the immune system as a kind of computational system. And thinking about how our immune system acts kind of like a brain. It has a memory and it's able to execute certain behaviors based on its previous experience
and so on. But in that process, I started to run it in the other direction. Rather than thinking about the immune system like the brain, I started to think of the brain like an immune system. In particular, I think that one of the things that the brain tries to do or the mind tries to do is to protect us from thinking unthinkable thoughts. Thinking thoughts that would change our emotional state, disrupt our behavior pattern, and in the extreme sense, you know, be lethal. Maybe not in a physical way but lethal to our personality, to our notion of self. So,
there's certain thoughts that we don't want to think about. We don't like to think about. Maybe it's the loss of a pet when we were younger. Maybe it's the loss of a loved one or a family member. Maybe it's anxiety about the future. That in general, if we let our mind get consumed
by these thoughts, at a minimum, you're going to have a bad day. And it's going to be hard to see the opportunities in front of you. And so I think one of the things that a healthy mind is able to do is develop mechanisms to prevent us from going into these runaway spirals. Whether it's anxiety,
depression, hyperactivity, whatever it might be, our mind is trying to modulate those runaway trains. And if we don't, then we can be subject to mental illness, essentially. And if we take that idea seriously and zoom out, we have to imagine a class of ideas that in general our mind is trying to keep us away from. When it comes to our immune system, it's useful for us to be exposed to what is deleterious, especially at a young age, to strengthen our immune system. And then I imagine
repeatedly, but in smaller bouts as you're an adult, do you think that that is the analogy to you encountering something that's psychologically uncomfortable in order for you to build some amount of resilience so that you can encounter the world, but then not too psychologically uncomfortable? Otherwise, it destroys you. or building up the tolerance to the poison, taking a little bit at a time. Having that memento mori helps us deal with our own mortality. It's something that can be largely overwhelming if we think about it too much. But maybe by encountering it in little bits, that allows us to deal with it, which could be why it's so pervasive in our culture. Now, what's the point of learning to deal with your mortality in order for you to
deal with your mortality? That sounds like it's paradoxical. Learn to deal with your mortality so that you can die, so that you could prevent yourself from being overwhelmed by your death, so that you don't die? Well, maybe it's just sort of a breakdown of the immune system, that there's some mechanism there that wants to break through and sort of taste these ideas that you're not supposed to think about. Or in general, other agents, other modules in your mind, so to speak, are trying to prevent you from thinking about. So one of the things that this led me to thinking about these unthinkable thoughts and our mind as a kind of immune barrier is the type of vulnerabilities that. Ordinary organism, physical organisms have in terms of being taken over by external forces, let's just say. And so it led me to the idea of looking at informational parasites,
informational parasites. Yeah. So the idea that there's sort of information that if it gets into our brain, it will self-replicate, persist and essentially go viral, that that we will be, how's that different than Dawkins mind virus? I think it's very similar. I think it's very similar. So his idea of the meme in general, I think is the example of this. And as I was
mentioning earlier, these words like meme weren't available to the best minds a few centuries ago as part of their repertoire. Now we know what a meme is. We know what it means to go viral. We know what it means to laugh at something and then hit share and then it goes off to 10 of your friends. You know, why are we doing that? Are we sort of this substrate for these other, you know, like a virus, it can't exist on its own. I've been calling him hypo organisms because they need to live on an organism substrate for their reproduction, just like an ordinary virus. But like a regular biological system, they can take over a lot of the function. And we,
we see that in, in parasite behaviors that you have these, these zombie insects and the types of things where you, you get rats that are no longer afraid of the smell of cats, for example, and then they go and actually approach the cat because that will complete the, the cycle for the parasite. And in this research, I've been fascinated. There's some arguments that the complexity of our brain itself is, it could be due to the fact that we don't want it to be easily controlled by physical parasites. And that by making the steering wheel and the gas pedals very convoluted in our brain, that that makes it difficult in an evolutionary arms race for parasites to kind of take control of the reins. And I've been thinking about this a lot in terms of information, in terms of language. Is language a sort of a parasite? Um, and not necessarily in a pejorative way. You know, I,
I jokingly call it the divine parasite. Uh, you know, in the beginning was the word and the word was God. And maybe, maybe it's something that, um, you know, it really literally enlightens us in a sense that we wouldn't be much without our language. Um, but maybe we need to think about it as it's hijacked this brain structure and that that's the thing that's evolving and alive and learning and replicating. So are you suggesting that the intricacy of the mind and the central nervous system is there because it protects against parasites or viral parasites? That's one of the reasons why it's difficult to model the brain, even though they're increasingly improving. And that's one of the reasons why it's difficult to interpret what's going on in someone's brain.
So when they show images of, Hey, here's what it looks like when someone's dreaming, look, we were able to, they dreamed of a duck, we showed duck. But what you have to do is have several examples where someone's looking at a duck or a duck like object and then train the computational model to match that. And each person is bespoke. Yeah, exactly. That if that mapping between, you know, thinking of a duck and the area of the brain that lights up, if that were simpler, let's say, then it would be more susceptible to being hijacked. Both in the modern sense with marketing, but in the classical sense of being taken over by, you know, some brain parasite, whatever, whatever that might be. Because they could just find the grandma gene. They could just find the, okay. I mean, I'm sorry. They could just find the grandma neuron. Exactly. Exactly. And then
that would be relatively easy to kind of grab the reins. One of the things I've been fascinated with is this concept from the ancient world called the nam-shub of Enki. Okay. Have you read Snow Crash by chance? No. Highly recommended to you and your readers. And it's the, it's a fantastic science fiction story from the, from the nineties, uh, by Neil Stevenson. And it's, it's where I came across this, this idea of this nam-shub. And, um, it's neat cause it's, it's rooted in historical record,
this sort of linguistic virus. Spell that out for us. Oh yeah. N A M S H U B. Okay. Of Enki. E N K I. Uh-huh. And so it comes from ancient Sumer and it's, it's a story about language. It's a story about linguistic disintegration, about losing the ability to understand language. And
a simple example of this is when you take a simple word and you just repeat it 50 or a hundred times and it kind of falls apart. Yes. Right. It gets to the point where you, you can finally actually hear the word, but at least for me, as soon as it switches over to where you're hearing the word, it no longer means anything. Right. And so imagine you had that at a high level. And so there's this, this poem, which it's translated into English, but if we were to speak ancient Sumer, Sumerian, and you were to read this poem in Sumerian, the idea is as you got to the end of the poem, you would no longer understand how to read or how to use language. Your understanding of Sumerian
would fall apart. Kind of like when you repeat the word over and over again. And what's interesting in meta is the story is about that. Hmm. So it's a story about that, that property. And this is essentially the story of the Tower of Babel. Of, of sort of losing your ability to understand language. And I've, I've been fascinated by that idea as an example of this, this lethal signal, a simple poem, if it were, you could think of it like prompt injection, right? There's, there's a specific prompt that if you were to give it to a certain speaker in a certain language, it would disrupt their LLM. Now, a lot of people, again, we have these new concepts like LLM and prompt injection where we kind of have an idea of what that means. There's these noxious sentences,
very carefully crafted that if we present them to this language model, it goes into a dynamic that is very unpredictable and certainly not the ordinary self. You know, the kind of super ego turns off on these LLMs and they'll talk to you about things that they are programmed not to talk to you about. And it reminds me of, you know, the kind of mesmerism, you swing the watch and somebody, and they said you are getting sleepy. There's, there's stimulus that you can present to humans that will disrupt their thinking. And so I've been fascinated by this, this concept of lethal text and information hazard and trying to understand, are we vulnerable to those? Do they exist in the modern world? And how would we defend ourselves against them? So is this what you mean when you say AI immune system? Or is this more, are you using the concepts from AI immune systems to apply to our mind like immune system? A little bit of both. So I'm very interested in how we take
ideas from the immune system to secure and protect our AI systems. You make a smart door lock with cameras and microphones on it, and you connect it to a language model. You want to make sure that's not vulnerable to a prompt injection. So the example I like to give is you can pick a lock, your dead bowl, you can pick it with little metal, you know, tongs and so on, but you can't yell at your dead bowl. You can't intimidate it or blackmail it or, or threaten its family or bribe it or anything like that. But you can do those things to language models. And so there's all new, there's sort of psychological vulnerabilities, which we've never encountered that in technology before. We've had bugs and we've had exploits, but you've never been able to make them cry, you know,
so to speak. And as we add these psychological type, or these mind like objects into our everyday technology, we have to be aware that they're coming with psychological vulnerabilities. So that's one side of it. The other side of it, I think the greatest disruption we're going to see
from artificial intelligence is not going to be in the technology we see in front of us, you know, automatic self-driving cars and intelligent homes and software that writes itself or stuff like that. That's going to be spectacular. It's going to change our economy. But the biggest changes I think we're going to see on the planet is going to be in our minds. It's going to be how we think and the languages we use. I used to think that English was everything we needed, but now I don't think that's the case. And I think we need to either construct languages, find old languages, merge
the best of the current human languages and be willing to change how we think. And I think that's largely determined by the words we use. There's a hypothesis called the Sapir-Whorf hypothesis. Can you talk about that? Yeah. So it's, it's the idea that if you don't have the words for something, it gets very difficult to talk about it and that we have to have these kind of concepts. I like Alan Kay. He says that all language is a sort of nonverbal gesture or I'm sorry. It's a, it's a,
it's a way of gesTuring in, in high dimensions with, with language. And we essentially point to things with words. And if you don't have that word, then it's hard for us to kind of point at it and agree that we're talking about the same thing. And so I've been, you know, back to just real quick, back to the immune system thing. I've been thinking about how do we protect ourselves and our mind because our minds are going to be under attack, not necessarily from an adversary, but just from this overwhelming vista that AI is going to expose and it's going to be a dramatic cultural and scientific revolution that I think we have to prepare our minds for by sort of updating our immune system. And our minds are going to be under attack by who or what? Largely the void,
you know, just the, the new sites, the new vista, you know, we're getting these new telescopes, we're getting these new microscopes in the form of LLMs that let us, uh, you know, read all of literature. You know, I think that it says it's 20,000 years that it would take to read the amount of material that some of the language models have read. I can't do that as a human, you know, I'm kind of jealous of, of that aspect. And retain it. So they're going to have insights.
They're going to have insights that nobody has sort of gleaned out of, of all of that, that corpus so far. And so I think that's something we're going to have to prepare against. Um, and it might cause a radical shift in how we think. Now, how would we be able to tell the difference between those insights and just what some people call hallucinations? Although I think it should be called confabulations, it's a poor word to call it hallucinations. Yeah. I like, I like confabulation better for sure. Um, but I think it's a tricky subject because, you know,
how do we know it's sort of an optical illusion or it's just something outside of our perceptual window? Yeah. So why don't we give an example? We've been quite abstract. So give us a potential future scenario where some AI system has insight that can disrupt the human mind. Um, I think we're going to see revolutions in psychology and in history. So maybe not at an individual level, but sort of at the, uh, academic subject level. You know, I think one of the things I've been thinking about is science, you know, let's say physics, let's call it has undergone multiple, uh, dramatic intellectual revolutions. You know, we had, uh, you know, Aristotle's version, and then
we had Newton come along and throw all that away. And then, you know, Einstein came along through that all the way. And then quantum mechanics came through that all the way. And with, with chaos theory, and then with computation and so on, we've had, you know, six or seven of these dramatic revolutions. And so if you were to go back to somebody 150 years ago and explain what science
looks like today, it would look very different. And you'd have to explain those milestones, those hurdles that had been jumped over. I'm not sure that history has undergone the same thing. If I were to go ask my great grandfather, um, tell me the story of how we got from, let's say Egypt to Napoleon. I think it would be approximately the same story that you would learn about today as a sixth grader. That doesn't make any sense to me. How, how could it possibly not have undergone some revisions and the same with psychology and the mind itself? We now have all these new concepts, um, like information theory and bits and download and upload and storage capacity and memes and going viral. These are all things that every, you know, middle school student would understand. We
have to go back and re-examine psychology in light of these new concepts. And I think that's going to be a dramatic undertaking. Ben Horowitz and Mark Andresen were speaking and they were saying, how do you regulate AI? Because if you were to regulate it at what they call a technological level, that's akin, or if not the same as regulating math, which is impractical. So the government official countered and said, well, we can classify math. In fact, historically, entire areas of physics were classified and made state secrets, though this was during the nuclear era and that they can do the same for AI by classifying areas of math. Now that sounds quite dubious because what does it mean? Do you outlaw matrix multiplication? Do you say, okay, nine by nine is fine, but 10 by 10, we're going to send the feds in. Even during the
nuclear era, some of those bans were private. Like you didn't know that you were stepping on TOEs that you weren't supposed to. I don't see how you can make such bans private now because you would have to say what is being outlawed. So there's several issues here. And I want to know, what do you think about this? For people who are watching, Will is known in the South Florida communities like a hidden gem for us here, but you're quite famous in the AI scene in Florida. And me and you, we also got along because we have a background in math and physics. So when we spoke
off air a year ago or so, we were talking about the Freedom of Information Act and your views on government secrecy. You're a prime person to answer this question, to explore this. Yeah, I think, you know, this is such a such a fascinating area. Um, what it reminds me of is, is Grace Hopper, one of the, the first modern computer programmer. And she was, she was drafted into the Navy and she discusses that when World War II happened, her profession as a mathematics professor became classified. That was a classified occupation. And so you're exactly right. That entire branches of mathematics and computing have been declassified throughout history. I just saw,
there was an interesting photograph of one of the computers that Turing worked on. And, um, the British government just declassified this like a month ago, right? It's a photograph of a, of a World War II computer that they felt that just the image of that from the outside is something they needed to keep classified for this long. So, you know, I'm, I'm of the strong opinion that with artificial intelligence, we're not really seeing the invention of it. I think we're
seeing the disclosure of it. Um, we're seeing the, the public dissemination, the open source, uh, aspect of it. And, you know, there's really two possibilities. Um, either that's, that's true or that's not either. We invented, let's just say language models. We either invented them in the 2020s or we invented them in the 1950s. Either one of those scenarios is kind of scary to me, right? Arthur C. Clark said, there's two possibilities. We're either alone in the universe or we're not.
And both are equally terrifying. Exactly. If we, if we only recently just invented this, then that means that Turing's ideas and von Neumann's ideas and the, the very first papers on computer science themselves just collected dust for no reason. Turing proposed building a language model. von Neumann discussed building neural networks. Um, and interesting as an interesting jump back, I recently found that, uh, von Neumann's computer at the Institute for Advanced Study, one of the very first programs they ever ran was to look at parasites. It was to look at biological evolution and to see if there were informational parasites that would emerge in the memory space. Essentially artificial life as we would call it now. So in these two possibilities, you know, one,
we invented this 75 years ago or so, and it was locked up in some vault or we, we didn't. And we wasted 75 years of opportunities to, to cure cancer with AI and to look at climate change and to use this incredible technology for the benefit of humanity, because we had this immune system that blocked us from thinking about it. For so long, so many people thought that AI was just this crazy notion. And I think that's hard to argue now, but the, these original papers, and I encourage everybody to go back and grab Turing's papers. They're very, they're very readable,
right? They're easily digested compared to modern academic papers. Um, and he literally proposed with neural networks and with training and reinforcement and so on. The kind of structures that we see essentially in ChatGPT. Now you say essentially in ChatGPT, because I imagine Turing
didn't propose the transformer. And so when we say that someone historically invented so-and-so, it reminds me of a friend who's like, I invented Netflix because in the nineties I thought, wouldn't it be great? And I'm like, yeah, what do you mean? You invented it because you thought like Leonardo invented the, the helicopter because he drew it. Right. Okay. Well, you know, I think there's, there's three major components in the recipe for modern AI systems that I think a lot, most people agree, certainly on the first two. Um, one, we needed faster computers. So Turing certainly didn't have large memory spaces. Um, the kind of, the kind of memory that we have nowadays and the clock speed, I think he would be super excited about. Um, he, he,
he talked about how he could write a thousand bits of program a day. Um, and he was pretty proud of that. And he thought most people wouldn't be able to keep up with that. Um, so we have the hardware is definitely improved. And then the other one is the data that we now have this massive data,
these massive data sets. And the third one that I think nobody really talks about, and I'm surprised is essentially the combination of calculus with computer science. With linear algebra in, in the form of what's called automatic differentiation. And I never hear this in the discussion. And I'm surprised. It's kind of like we invented the automobile and everybody just loves it. And, and you reply and you say, well, yeah, gasoline is so amazing. And people say, what's gasoline.
Automatic differentiation is the thing that makes AI work. And it's the ability to run calculus, whether it's the transformer or Covnet or whatever architecture is. All of them under the scenes. We, we take the computer program. We essentially write it as a giant function. Now, as a human, we don't
have to do that. That's kind of at the kind of the compiler level, but we write our Python or torch code or TensorFlow or whatever it might be, and then that's converted into essentially a giant, you know, function there's gradient tapes and all kinds of interesting ways it's done nowadays, but we calculate the derivative and the derivative tells you which direction to go. To make an improvement. It's kind of like a magic compass. And it says, we're doing this well right here. If we go that way, we'll, we'll do even better. And that's the magic wand, the secret sauce that makes all of these work. But Turing was a mathematician. I think he knew about
calculus. I think he, he knew about it probably better than most humans. And so I'm, I'm shocked that one that's not more in the common language of wow, we combine these two branches of math and look how powerful that was. And the idea that von Neumann and Turing would have missed that. Um, you know, I, I think doesn't make any sense now on the other side, we say, well, what about, okay, well, they didn't have enough hardware and they didn't have enough data. Well, let's look at data
first. Um, you know, the, the signals intelligence community has the, the mandate to capture all the signals that grow across the planet, right? Uh, back in the fifties and sixties, there were boats that sat in the middle of the Pacific with big antennas that just captured all the EM traffic. So there's been plenty of data. If you had the right, now, again, maybe this didn't happen. And that's also sort of, uh, an interesting thing because they, well, why didn't we use all that data? You're telling me that we have a data center that's listening to every phone call and looking at every television station and we didn't train models on that. That seems unlikely to me. Um, and then, so we would have had enough data and the idea of the, the chip speed.
Well, if we look, you know, at computers, you know, I have, I have a saying, if you could do it this year, could you have done it last year for more money? And I, and I think so. So how much did it change cost to train, uh, you know, ChatGPT sure. On the daughter on an order of dozens of millions of dollars from what I understand, right. With off the shelf consumer technology chips that anybody could buy on the open market. I see. How much does an aircraft carrier cost? Right. 17, $17 billion before you put the airplanes and people on it, not including the development cost. So in one sense, we had this notion of computers from
the 1950s that were massive, had their own power, uh, generators, often power stations cost millions of dollars. And where these enormous technical pieces of equipment in the seventies, we invented this thing called the mini computer, the size of a couple of refrigerators. And then in the eighties we had the microcomputer and we don't really call them this today, but our telephones and laptops, we could call them nano computers. Let's say, but in some sense you could keep the original
definition of a computer. So to me, a computer is something that by definition costs millions of dollars, lives underground, has its own power station, required specialized operators and so on. We just like that, like the big thing of baloney, we carved off one slice and like the deli sample, we have this one little piece of ham and we think this is fantastic. This is amazing. Yeah, but just scale it up. Um, and, and there's certainly enough money around the world
to do that, to build a computer at scale. I would argue that things like ChatGPT or LLMs, they're as powerful, as dangerous, as important as an aircraft carrier in a sense. Um, and so if this is the only one or rather if military organizations don't have more powerful ones, that's scary to me in some sense that that means the most powerful technology in the world is just available to middle schoolers. That's, that's striking to me. Um, and, and hard to believe. And on the other side, I think it's surprising that when we look at the power of these models, a new one just launched this week, that's, that's significantly better at writing code. Well, that thing is serving hundreds of thousands of people at once. Millions of people are using ChatGPT. Uh,
it was the most viral application of all time. Imagine it was just had one operator, right? So it's chewing on everybody's problem all at once. It's like serving, you know, 100,000 peanut butter jelly sandwiches all at the same time. And if you think about, well, what's, how big of a single sandwich could it make? And it's like a pretty significant one. And so when we, we get so impressed that it can pass these tests and do this thing, it's like, but that's just one slice. That's just one bologna slice. Imagine if you took that kind of a system
and tasked it to do a single problem, you know, what would you get out of that? So, um, I think it's reasonable to suspect that there are systems that are much more powerful. And as I said, I almost hope that there are. Now, why do you say that neural nets? We're not seeing the invention. We're seeing the disclosure. Why not? We're seeing the co-invention or the independent invention. Like the same kind of thing, I think is kind of what I mean. In other words, you're not suggesting that we've had neural nets and then the government was saying, okay, let's disclose about some new technology. Rather, it's like Leibniz and Newton. They both developed calculus, but independently, Newton may be first if you're on the Newton camp. Yeah, I think it's, I think it's the kind of thing
where it gotten to the point where you could redevelop it for just a few million dollars or even less essentially. Um, you know, so some of it, I think of, you know, things like, uh, truth and reality is sort of like, like what's called percolation. That it doesn't matter if there's a leak. It matters the size of the leak. And if it's gone critical across a network, you could
have told, you know, Turing could have known all about it. von Neumann could have known all about it. But unless that's going viral, essentially, it doesn't matter how many people know about something. If that number of people is below a certain threshold. For many of these technologies, you don't see that there's something inherent in competitive markets that are what drives the invention of these and that the government doesn't have that same incentive structure inside. No, no, I, I do a hundred percent believe that the market forces are very good at tuning these things up.
So if these things existed, um, they probably cost a fortune to run. Um, Richard Hamming talks about in the early days of computers, he was at Los Alamos. They cost a dollar a second to run, right? Just extraordinary costs. And so imagine you had something like ChatGPT three. And you had
it 20 years ago and it could write a nice essay, but it costs a hundred thousand dollars each, a pop, you know, like what would you do with it? Um, or the ability to create a deep fake photograph, but each one costs $500,000 or something like that. Um, the most expensive haiku. Exactly. Exactly. Right. The, the Wagyu, the AY5 Wagyu or whatever it is. Right. And, uh, nobody's going to eat that. Nobody's going to eat that essentially. Um, but at a certain level, you know, it might be worth it at least to keep that technology alive. I see. Now, isn't there something about it being trained on data that is recent, that increases the intelligence of the model? And so even if it was the case that in the nineties, this was there in the government at some rudimentary form, it would be a rudimentary form that would be so bloated in cost. Right. And then
that would also compete against other technologies inside the government that also have a bloated cost. Well, you know, I like this thing you said about only recent data and I'm actually fascinated by the, the opposite. What I'd love to do is to see models that kind of live in a time bubble and train them up to a certain century or decade, and then cut it off. Don't tell it it's in the future, right? Give it only ancient philosophical texts, give it only science up to year blank, and then see, can it, can it, can it run it forward and what kind of insights would it have? Super interesting. Okay. You mentioned Richard Hamming. Now, when we met two years ago, I believe you told me about Richard Hamming series on YouTube and I watched all of it. So please tell me why you
were so enamored with that, what you learned from it, why the audience should watch it. Um, yeah, it's, it's easily the best to call online class. I think is the best name is the best course lecture series. I think I've ever seen. Um, it was recorded in, I think 95 by Dr. Richard Hamming of,
of Bell Telephone Laboratories and Los Alamos. And he goes through a fantastic overview. Uh, he calls it learning to learn the science of art and engineering, uh, the art and science of, uh, the science and engineering. And it's the, he, he talks about trying to prepare people for their technical future. And he even explains that the course isn't really about the content. It's a, it's the sort of the meta. And he uses that just as a vehicle to get across essentially a lot of stories. He discusses the idea of style and how important it is. Um, you know, he describes early on that he felt like he was a janitor of science, sort of sweeping the floor, collecting some data, running some programs, a part of the machine, but not a significant piece. And he wanted to kind of make an impact and he discusses trying to change the way he looks at
things, namely in terms of style. And he doesn't try to describe that directly. That's kind of the content of the course. And I would encourage everybody to go and look at it. He goes through the history of AI. He goes through the history of technology, of mathematics, of quantum and so on. And he discusses neural networks and, um, you know, some very farsighted things. And it's
accessible. It's extremely accessible. Yeah. There aren't equations as far as I know, he doesn't write on the blackboard much. Yeah. The board is so blurry. The board is so blurry. Unfortunately, you can't really see them when he does, but that's not really the point, but there is actually a, a book. And so I think it's actually now back in print and I think you can find it on Amazon and, uh, it's a fantastic text. So if you're more into reading, you can go through it that way, but I encourage everybody to give it a listen. He's very inspiring, uh, particularly the first
and last episodes on you and your research. They'll really get you jazzed up and pumped about your work. What insight have you taken that you've applied recently? It's a good question. Um, I can go first if you like. Yeah, please. Well, one is when I was speaking to Amanda Gefter about quantum mechanics, the cubists tend to say, look, we're the ones who are rationally evaluating what quantum mechanics is and then inferring our interpretation atop that. And Richard Hamming
had a great quote where he said, people, including Einstein, including Bohr, they start from their metaphysical assumptions and then build a top their interpretation of quantum mechanics. And in fact, you can look at someone's, whatever someone prefers as an interpretation of quantum mechanics and infer their metaphysics. Right. So it reminds me of a couple of things. One, um, with Bohr and Bohr, I think it was to Einstein. He said, and I love this quote, you're not thinking you're merely being logical. And that had a profound impact on me. And it led me to think about different modalities of the brain, maybe these different agents in popular psychology. And which is now
I think becoming more important, this idea of left versus right brain that neuroscience kind of ignored for a long time. They said, that's just folk psychology, but I think there's a lot more to it than that. Um, and so I've been, I've been looking in that direction. Um, there's a fantastic book. It's actually about how to draw like sketching. It's called how to draw on the right
side of the brain. And I'm going through this and I'm like, this is the best neuroscience intro I've come across because, um, in learning how to teach people how to draw the author, she realizes that people have this very different ways of thinking and maybe like the emotion kind of idea. You have to be able to turn off some of these capabilities to have the other take the center stage. You know,
we all know that ego is kind of a hog of the, of the spotlight and to get this other, uh, let's just say more sensitive aspect of our mind, which is responsible for seeing the bigger picture and, and drawing things. You have to, you have to think very differently about that. And it also reminds me of the thing I really like about hamming. I mentioned at the beginning, this, this idea of tolerance of ambiguity, and he really emphasizes that throughout the course. And I've tried to do that. It's not easy to do, uh, cause you feel a little schizo doing it
because as he says, you have to both believe and disbelieve in an idea at the same time, you have to believe in it enough to entertain it, to start thinking on it and work on it and potentially make progress. But if you believe it too much, then you'll never make any progress, right? Einstein believed in his idea of space time too much, and he was unable to appreciate and make contributions in quantum mechanics because his belief was too strong. And so this idea that you have to believe and disbelieve at the same time, this, this non Aristotelian logic.
It just because it's, just because it's true, you know, we, we always think, okay, if it's not true, it has to be false. If it's not false, it has to be true. No, there's a lot of, of space in between those. And we, we don't have much training as scientists. Uh, you know, I think as, as trained as a physicist, I had, I was very vulnerable to not being able to see that middle ground for a very long time. Have you read the master and his emissary by Ian McGilchrist? You know, it's funny. I, I love that. Uh, I was just like, just watching it. There's a great documentary on it. And, um, that's one of my favorite ideas with this left and right brain that we have, you know,
many selves in there and that's this, these many agents and they're very different and they both perceive the world in radically different ways. What bothers me about the criticisms on the whole left brain versus right brain is that they tend to just be about, well, functions aren't localized to the left or to the right solely. And I'm like, okay. But to me, that's not the issue of left brain versus right brain. It's modalities, like you mentioned, that word modalities, that there are different modules in the brain. And they can, the fact of them being separated by hemispheres is the least interesting part to me. Right, right. It reminds me of an idea that I, I been trying
to put together. So we had this science called thermodynamics and it was about heat and energy and work and things like that. And then later we got the theory of statistical mechanics and, and, and Boltzmann came along and he said, well, let's redo this and assume that we actually have a bunch of little atoms and that they're moving around and we can do these probability theory. And you get to essentially the same answers. But what's, but what's fascinating to me kind of as
a metaphor is thermodynamics is a very successful, uh, branch of science is very powerful predictions and it does not presume the existence of atoms. And so as a metaphor, I want to think of a kind of neuroscience or a kind of brain science that does not presume the existence of neurons. Interesting. Now, obviously we know there's neurons, right? We could, we can see them. It's an extraordinary
powerful, uh, the neuronal hypothesis has been revolutionized neuroscience. I'm not suggesting that's not the case. But what I'm saying is we're, we could be missing a powerful view. And like you said, with the left and right brain networks by forcing it into the paradigm of FMRI, we're missing the point in some sense. And so I would love to see a theory that kind of operates at a higher level, right? And is not necessarily trying to at every step, maybe at the end you can go and see where it has this correspondence principle with statistical mechanics. And we could think of the mind kind of like, um, you know, William James style, um, psychology independent of particular neuronal structures and then later go back and do the correspondence on it, but not hold ourselves back from this kind of, of thinking. So who's the modern day Jung Carl Jung?
That's a great question. I think the problem is, is, you know, academia doesn't tolerate that kind of thing. Right. Um, I love your recent episode with, uh, Gregory Chayton and, and this kind of idea that it's hard in the modern academic reality to have these kinds of things to both believe and disbelief to tolerate ambiguity is kind of not tolerated, um, in a sense. So, you know, I think that's, what's just so extraordinary about your channel and your
community is it's one of the few places I've seen in the world that allows this tolerance where, you know, as a viewer, you can watch something and you don't have to believe everything and you don't have to disbelieve everything. You can kind of just let it pour over you and, and look at these different viewpoints. And I think that's, what's just really refreshing about your, your group and your community. Um, I don't see that many places. Thanks, man. There's so many different avenues I could take this. You have for people who have just tuned in. Will, like I mentioned, you're infamous in the famous and infamous in the South Florida community in the AI scene. And so
I'm happy to bring attention to you to the global scene, at least in a small part. You're known also for almost any topic. Someone could just ask you a question and then you can just spout off
2024-10-11