"I Worked for DARPA & Here's the Truth..." | Lee Cronin

Hey guys, if you're not following me on Spotify, please hit that follow button and leave a five-star review. They're both a huge, huge help. Thank you. It's great to have you back, man. Last time you were here episode 289, it was like a year and a half ago. You closed up, like we literally stopped the cameras and then you're like, "Yeah, I was doing so much for DARPA as well." And you just lay that one on us and I'm like, "Well, you supposed to say that at the beginning of the podcast, but now you're telling me like, "Ah, it's not that big a deal." How's DARPA not a big deal? I mean, I'm like a big DARPA fan. Yeah, so I've had maybe, let me get this right, so how many grants have I had from DARPA? So DARPA is a funder, right? They're a funder. They're a grant funder, they give people money. I think the reason they were established, you know why they were established? I think you can check this, so their mandate was like to not, when Sputnik went up into space, the US were not expecting that satellite. The Soviets launched a satellite. I believe it was in the 50s. I think that's right. And basically DARPA was established to eliminate strategic surprise. So basically that one load of money, fund people do crazy things and also to have a mandate to do stuff. Arguably, DARPA might have had something to do with, say, the internet, self-driving cars. And a few years ago, DARPA started looking at chemistry and AI. Many years ago, way before the hype now, and there was a program they started called Make It, actually. And the person who started it, you know, they kind of used some terminology from one of my papers. And I was kind of a bit grumpy. They didn't offer me any funding, right? Because they gave money to MIT, to Stanford Research Institute. DARPA plagiarized Lee Kronin? I didn't, you used the P word. They were inspired. I said, anyway, I remember they got a call with me and I was like complaining. But, Ellie, the fact that they used this paragraph of, you know, what I was trying to do. And they're like, Dr. Kronin, we're phoning you up because we were on the talk trip out that. And I was like, okay. And basically what it turned out is they wanted me to put in a white paper. And I said, okay, I want this, you know, resilient dollars. And they said, well, we're not asked for a little bit less to do this thing that fits into our program. You know, it's US taxpayers and you're one of the, you know, few groups overseas doing this thing. And we think it would be valuable for us. And what was the thing specifically? So it was to, it was to, it was part of their Make It program. Can you make molecules on demand in theater or, you know, on the moon in Antarctica, whatever, right? We do chemistry robotically. And then when they asked about that, and I've been building chemical robots for years. You've been building chemical robots for years. Yeah, yeah, yeah. That's a dense statement for, for, since 2012, right? So about, yeah, about 14 years. How would you define a chemical robot for people out there who are unfamiliar? A robot that you give instructions to and it will do chemistry. So, yeah, it's a, it's out, so I made, I can tell you about it. But let me answer the DARPA question because you asked it. So that was the first grant they gave me and make it and they carried on giving money there. And then they were, they had a one on computation using molecules to make molecular computers. And I, I got money from DARPA there and that was in a US consortium. So there was a couple of the groups in the US, right? And then there was another grant they gave me for kind of a vent using AI for discovery. So they were the three DARPA grant. I think there's three DARPA grants. What year did they give you the AI grant? Oh, I can't remember. I mean, Bob Park. 2019, 2018, 2019. And then it carried on through the pandemic. It was a bit tougher in the pandemic. But DARPA brilliant organization, right? They go where the experts are. They don't give grants to people in countries that are probably anti-America, I would say, right? So they're in fun bin Laden. And also they want to start new fields and they want to start fields that are going to obviously be of strategic importance, right? And the UK and the US have done a lot of hookups in different scientific disciplines for a long while. So I thought that was quite good. And then molecular, I made a molecular computer chemicals to do computation as well. What does that mean? Yeah, I don't know. Yeah. I mean. So. So well, I know what it means now. So what does a computer do? So a computer. We're around going everywhere. But anyway, let's start. But what is a chemical computer? A chemical computer is a chemistry set where you could put in inputs. The data, you input the data into the chemistry somehow. And the chemistry would then take that data and process it using chemical reactions, run transistors on silicon. And then you'd read it out using it some method. And I was like, I built a chemical computer using a thing called the blues off saberskinsky reaction, which is a chemical clock. And all it goes tick tock, tick tock or red, blue, red, blue, red, blue. Okay. So basically, put a lot of chemicals in the pot. You could just take, you know, maybe just a beaker, which is like a glass, put in half a fill it up halfway, put in some chemicals. And without doing anything, just leaving, watching it, it would flash red, blue, red, blue, red, blue, red, blue whilst you're stirring it. When you stop stirring it, it would basically stop. But then spontaneously flash one color or another, what's called an excitable media. So a bit like how, think about how neurons flash. So that's a cool idea. Why don't I just basically make a grid of stirers? So I 3D printed a grid, basically about the size of a small book. And in that grid, there was, it was seven by seven grid. So 49 little wells. And I put a sterile bar in each one and have put a little motor under it and stirred them and put the chemical reaction in. And all the colors went across the grid and they used the web camera to basically read out the colors. And so when the stirers were on, I could, that, that was a, a one, when the stirers off, that was a zero. And I basically then put, basically printed in ones and zeros to represent some input. And then just read it out and then worked out what it was doing. We kind of just made it up to start with. Yeah, wouldn't that tell us when you're done? Well, that's. Hey, it's a DARPA project. It's supposed to be crazy bad. No, so what it was supposed to do is, is could I, could I somehow use the, because remember the BZ is a clock tick tock tick tock tick tock. If you put a load of clocks together and you allow them to synchronize. As they desynchronize, you could process some information. So we actually used it to classify and we used it to make a primitive neural net. And the idea was the neural net, you literally use no power and was quite good at error correction and could like refine images. So the same way you would use chat GPT or you'd use a kind of Nvidia's architecture to do gradient descent. And what is gradient descent, or it's basically that it changes the weightings in the neural network. And to basically, you know, kind of maximize its training capability, because obviously you have the weights and the, and the activations two different things. But basically this was a chemical neural net. It was the idea. It was the first step. I'm now making brain gels, but that's another scar. You're making brain gels now because we put that and put it in a gel because it was all liquid and sloppy. And the memory wasn't that good, but if you put it in a gel, then maybe you could then basically use the polymer gel to flip to switch in such a way you could teach the system to learn for a longer period of time. One of, if not the most common DMs I get on Instagram are many of you asking me what my favorite books are. Well, today's your lucky day, because if you use the link in my description below to join my newsletter, I'm going to give you my top 10 book recommendations right now. Also, if you'd like to join my clipping community for a chance to make money producing content from the show, you can do so via the Discord link in my description below. And finally, you can join my Patreon for early uncensored releases of the full episodes via the Patreon link in my description below. As always, please take one second to hit that subscribe button. If you haven't already, I really appreciate that. Also, the like button's a big help. And you can join my newsletter via the link in the description below. How does that work scientifically? No idea. That's why I did it. So you're trying to figure it out. It basically, yeah, I mean, look, we're again, we're all over the place, but why not? I was inspired by the episode of Pickle Rick and Rick and Morty. And I was like, well, and so. So it's like, I took a Gerkin, right, and attached the Gerkin to the mains to 240 volts and then put in low, low, you know, my workshop, I have at home. And used a bunch of electrodes and I tried to program the Gerkin to recognize difference between a circle and a square, but it kept exploding. It kept exploding because it kept heating up with them. So I had so long. So I took the Gerkin, I thought the Gerkin could be a good. It's a good idea. It's like a Gerkin computer. Literally, you could just be making this up. Right. It's like, yeah, just anyway. So Gerkin heated up to put it in the. It sounds ridiculous. You heat it up by putting it in the 240 volts. What would happen is sodium ions, potassium ions, because it's in salt. So you got this Gerkin, you soak it in the salt. So it has this salt in it. It's conducting when you put power, when you put electricity in it. to it. When you put alternate and current high voltage, it heats up, and basically the ions can move around. So it literally makes an excitable media, a bit like the BZ reaction, like flashing on or off. They're here with the gyrkin. And then when I put in the little electrodes, I put in the little grid, and I was then trying to basically train the gyrkin to tell the difference between a square and a circle, because it was made malleable. It's a bit like taking some cement and making it liquid again. And then you make it liquid again, and it'll learn something, and you'll set it. But the gyrkin kept exploding before I could actually train it. And this is in your lab at your house? I did at home, yeah. I mean, I don't think the university would be too happy with it. Now fast forward to the brain gel, I was like, well, look, if I could take a gel and then make the gel conducts electricity, a bit like a wire, and then put in electrodes on one face and the other. And program one surface, using a camera, so I could just take a visual feed and plug it into the gel, and then basically read out as I show the camera to different objects, can the gel tell the difference between the edges on the objects and train it in the same way you'd use OpenCV or some kind of silicon-based system for edge detection, right? It's a bit like you use a Jetson Nano or something, right? And I was just playing around if I could use material to compute. Turns out you can, but now you're like when you said, well, how's that work? And I laughed. Well, it's not because I'm making it up, but because the way you program material is the material responds in time, according to its stimuli. So it's very much trial and error. So it's a bit like how the brain is evolved, like the brain has evolved over billions of years. And so the brain has a number of different kind of programming periods. It was programmed by evolution. So your brain, let's say, is 3.8 billion years old. So that's time, that's time number one. Then you are created by your parents. So over a period of nine months, your brain is growing in the womb. As you, as you're, and so your brain is starting to program, when you're then born, obviously, between you go through massive developmental surges. So you know, age 21, 22, your brain may be fully almost there. So then your brain is unable to process data in real time. So there's all these different developmental stages. So the brain is able to evolve, grow, be programmed, activate, and then continue to work. And we have no idea how brains work. And what are unique to brains? Well, sentience, consciousness, the purpose and a living system that will tell you that they are an individual. You talk to all humans. Most humans can communicate with, we'll claim they're an individual. So you think the brain is the thing that is strictly just programming that in and it might not necessarily mean that that's the case? No, no, I think there is something very interesting. I don't think it stops at a brain. I think biological systems are very complicated, but I think humans are unique in that they will claim to be an individual, right? With words, if we teach them language. And I think that's something that we don't understand how chemistry gives rise to consciousness and how chemistry gives rise to, you know, the fact that I will claim to be me and I have my own particular personality and do stuff. This is very important this time because we're viewing the AI as our entities are sent in and they are indeed they are not. Yeah, you've made a lot of arguments on this because you think that for example, a lot of AI tumors are making leaps that the AI can exist without human beings. Yeah, yeah, in the future and you don't agree that that's the case. Yeah, clearly, clearly, it's not, it's fantastical. Why do you say it's fantastical? Well, we've not seen any evidence of the, so the eight, if we think about the origin of AI, right, if you go all the way back, origin of life, the AI is produced by, well, let's say the origin of the term AI was at the DARPA, sorry, no, wait a minute. Don't do that. The DARPA, the DARPA meeting, right, we should know, you can check. It's like 4D and slip right there. Sure. Okay. Whatever. Let's build as many conspiracies as we can go. But I think DARPA has funded some AI work, right? Right at the beginning. So DARPA College, the term AI was kind of coined. When was that again? In the 50s? I mean, I don't know, 56. I don't know. But a long time ago, maybe a bit later. And so, but if you think about what a computer is, a computer is a deterministically built system. It's switches, right? And the way those switches work is we built a computer is built out switches. Those switches have high and low. You put those together to make logic gates. Those logic gates, the universal logic gate and the computer is a nan gate and nan gate. Yeah. And not and. Right. You can build everything with not ads, right? So not ads are literally, you can build memories, you can build everything. The way you architect those together, you build the silicon, the silicon infrastructure for everything we have today is based on that. There are maybe some differences that I'm, you know, electrical engineers watching this, listening to this might be shouting, go, no, that's wrong. There are other things called ASICs, which are basically that you might be able to do what we would call, I mean, neuromorphic computing. We're basically, you can have an analog voltage. So in a transistor, you basically, you have a voltage between zero and five. Five is high, one, zero is low, zero, and something in between. And if it's on the zero side, you call it a low. And if it's on the five volts side, you can give a little one, but I'm not a qualified electrical engineer. So I don't know how the ASICs are built today. I designed some ASICs for fun. For fun. Yeah. Well, I want to build a brain. So if I want to build a device, a chemical brain, I have to design a, I need to have an analog kind of electrical system for addressing the chemical brain, but we'll get there later. We'll come back to that. So the most important thing with a computer is like, it's fairly deterministic in some regard in what, sorry, let's not say that it's digital. And so that means I can put in the inputs in there and go on. So with AI is today, they're based upon digital computers and they use all this digital infrastructure. We take massive amounts of data and train a model and the model can do splendid things. But the dooms say that these models could run away. And I agree the models could run away with bad humans running them, right? There's always a human in the loop. The human is turning on the power station, the AIs, although they are, you know, we use terms like recursively self-programming. I mean, it's a bit like you and I could build an AI now that we can have a, we can have a loop in a program like, you know, do this thing. And if this thing is, it occurs then do that thing and carry on until it either hits the objective or you, someone turns off the computer. Now, that's always been the case, right? Those loops have always existed. So the, so the AI dooms are kind of stuck in this like this kind of fantastical reality whereby they say that AI could overtake the world and cause a nuclear explosion or something else. What I think is more likely is that bad humans using the AI to do bad things will happen. And, and I do worry about that. So I'm not dismissing AI dooms. I'm dismissing AI dooms who are saying this thing will happen because magic on its own. Yeah. Got it. And I think that is actually a major worry because they're talking about magic. And actually there's just some doom ago, some, some bad actor going, I'm going to use the AI to crack into this system and do some damage or I'm going to, or I'm going to use an AI to bot system that will basically do some social engineering on social media to cause some kind of effect, right? So I think that I haven't seen any evidence of sentience in any AI, right? The sentience comes from humans exclusively. Any evidence of sentience you get from the internet, the training on the model. If you say a Claude, are you alive? Claude will probably say, no, I'm not alive, but like I, and they'll use I and this anthropomorphization is just to hook you to be addicted to talking to the entity. What if it's a format? What if it's a technicality though? And it doesn't get the sentience. But because, and I'm going to weigh over, simplify this for a minute, we're able to eventually or create technology that isn't quite simply plugged in and you know, unplugged to be able to do that because of that a non-senscient but extremely intelligent artificial intelligence who has no ability to empathize, feel obviously feeling sort of emotion or level of humanity took actions that it doesn't understand the grave consequences of. Do they have an argument for something like that to possibly happen? So I don't, I don't, I don't quite understand the statement and but I know what you mean. So I'm not going to be too obtuse. But AI's, are they intelligent? We have to define intelligence. What really makes me very confused is there are people out there that say, in next years, the AI will be more intelligent than any human. And what does that mean? What is intelligence? If you can't measure something. So could the AI, call something to happen on its own? No. Could humans not understanding how to set up a system, call something bad to happen? Think about social engineering. I don't know, take any system like, this is kind of hard because I'm not a qualifier. But let's say we're going to I don't know, create a type of by-law in the city that says, we have to have trees for you to get planning permission in the city, every time you build a building, you have to have a tree or every corner or something. Suddenly, before you know it, there's just random trees everywhere. So what happens is when humans make decisions and policies, they get propagated and they get propagated by systems, whether it's in a corporation or in an institution or in a software algorithm, right? So those things get propagated and they're unforeseen consequences of them. And I totally buy that AI is trying to achieve an objective, could do bad things. But those bad things, humans are in the loop all the time. And I guess what I'm trying to say is the AI is not bad, the human is bad, or the training data is bad. So and that would, if I agreed with that point, and you might be right, I honestly hope you're right, and it is in control of humanity, by the way, so obviously speaking here. But let's assume though, there are some sociopaths who have their hands on the trigger with these things, which I think it's fair to say some of the technocrats we have in the world, not all of them, but some of them may fit that bill. What's to stop people to put a name on it like a Sam Altman who jokes about humanity ending? What's to stop him from being like, oh, you know, today I want to play fucking World of Warcraft, but with AI and let it go wild. So look, I have a lot of sympathy for people who are building companies and building technologies. I'm building one myself right now, which we can talk about. I think we have to kind of, we have to take a step back and say, what are we built, how are we building our technologies, how are we regulating those technologies, and how do we then get feedback from the regulation of those technologies to make sure that good things are happening to humanity? So let's take an example where this has already happened, right? We got the internet and then we got social media. So what was the critical failure in social media has given us arguably, I don't think anyone would argue that social media is particularly nice right now. What's the failure? So there's one critical thing that people should have done, which was the policymaker should have said to social media, they should be treated like publishers. Section 230 or whatever. I don't know what the US version is. So if you're treated like a publisher, that is anything you put on your platform, you have to basically hop be held accountable for. When you start putting nonsense out there, it can be taken, you have to say no, it's clearly alive, put it down. So basically Facebook and so Facebook and X and Instagram and all these entities are not accountable for the stuff they put out. Now that's for me, that's a real problem because basically that entire decision, although they're a medium, it's a bit like you can't hold the person who built the printing press liable for the bad books printed, right? Get that. The same way we might argue when you talk about my chemical robots, am I going to be liable for anyone making any bad chemistry on them? But I do think there's something we can learn from the social media and the fact we didn't regulate that, we didn't hold them accountable. And so it's very hard to know how it's going to revolve. However, having said that, if someone invents the social media, there's quite good as in a nice place to be, not a hellscape, right? If we can work out why humans are very good at, you get an anger for attention. If you create anger, you get the attention, the system wants. So, you know, if there was a social media system whereby I'm not saying it should be like, you know, some kind of Mary Poppins world where everyone's like, you know, it's all flying around and all happy. But there, I guess we have to understand the psychology of social media and the negative effects it has particularly on teenagers. When I grew up, there was no social media. That's right. And I've seen that social media has been probably a negative for a lot of teenagers. But some of my argument was like, well, in environments where they don't have access to social centers or youth clubs and things, there's people can work, play online, play games, interact one another, support one another. And we don't see all those positive things because there are only negative things that amplified. So when it comes to AI, we have a similar quandary, which is how do we, how do we let the free market and free humans dictate what's happening? Because there is one argument, if you look at the difference between say Amphropic and maybe Open AI, Amphropic basically won't let you do anything on the models because it's got some kind of higher purpose wired in. Whereas Open AI is like, no, no, I'm going to serve you as the entity. And actually, if I'm a, I want to, and if I choose to use the object to do bad things and I'm breaking the law, then I should be held accountable. Whereas like, and I think so, I'm not sure if I'm getting that correct, but it seems to me that there is a kind of a split right in this right now. But that's kind of obscured by the doom is saying, AI could run away. And I don't understand the mechanism for running away. I do understand the mechanism for tinkering. I do understand the mechanism for, but may put disinformation out there or for doing some kind of now the conspiracy theories that don't aren't yet possible. Here's one crazy thing. Look, there's no real conspiracy theories in history, not many, because humans can keep them secret. But if we had good AIs and we could just pump, you could play around with the future with the AI is pretty well. And that might be one thing to worry about. Now, you know, is there one big cause if we say that, you know, but one particular political movement or post second world war, when the winners talk about what happened, could that be one big glorified conspiracy theory, because obviously the winners write the history books, okay? There's always a percentage that's written by victors for sure. As far as like, when you look at a story, it's never 100% what was said, I take issue with when you then try to say, well, if 5% wasn't what we were told, therefore 100%, it's not, you know, you got to be careful with where you go. Absolutely. Absolutely. And I think that's really important. And in terms of nowadays, I think that critical reasoning to what can I verify and what do I find impactful to me is really important. So I think one of the things that I think AI will accelerate the teaching of critical reasoning. Right now, we're in this real, I worry a lot about how young people are going to be allowed to get their first job, you know, make mistakes, get the training, the mastery is required, have the hard work and motion frustration is the best trainer, right? But whereas like people, you don't want to get frustrated now, just use chat GPT. And so, but I think this is an aberration. I think these tools, there are actually a lot of them are really good. Like the coding tools are amazing. And so I do wonder, and this is me actually shifting my emphasis a bit, if AI might actually save us all, save us all, save us all because social media is such a clusterfuck, right? If by using AI tools, we can actually allow, but help us do critical thinking, help us verify things. If we understand that the, you know, the current AI is as generated, I'm not very good at actually class as classically trained, weren't very good at being correct, but actually they're getting better at being correct because you, you can say, hey, go out and check the literature and check if this thing actually happened and help me, you know, storm, storm and or steel man, a particular argument. And then also help me understand as objectively as possible, how I might frame that argument. Because if, if nothing else, these AIs might be able to allow you to do one to one teaching, they've got infinite time, infinite patience. And if you use them correctly, if you use them correctly, well, we might be, yeah, if we can use anything incorrectly, like I say, you know, so, so I'm kind of, I'm, as you might tell in this conversation where I'm just saying wildly contradictory things, strong opinions loosely. Yeah, yeah, I mean, for sure. But I do think that the, the, the doomers are kind of like, they're kind of out to, it's a bit like, you know, um, that, well, I would say that the, the doomers serve a very good purpose, but I just wonder if the doomers could actually help us by shifting their doom to the real issues that we need to to deal with such as like bad acts as using the systems, right? - Well, that's the thing that I'm thinking about with your world right there that you're painting. On paper, that would make sense. But again, if these different platforms, because there's as many of them, are going to actually be a positive and help teach critical thinking and things like that for a person to be able to do that themselves, then I think a requirement for humanity for all of them across the board would then be you need to open source all the code at all times, because you don't want to know, or you want to make sure that there are not human beings, you know, just slightly injecting their little opinions into coding the actual AI, I'm going to use a fake word here, but the AI organism that is going to teach us to what we do. - Well, I mean, no, AI organisms are cybernetic organisms, they are the combination of human programmers and the model working together, right? So I'm very happy to say, the sentience in the AI comes from the humans, right? This is really the nice mystery that we can talk about. But look, I'm uniquely unqualified, although to say what the models do, because I don't build, well, I mean, I build models, but I build models based on data and very basic learning algorithms and systems, because I need a first principles understand it. If I can't understand it, then I don't tend to implement it. And the reason for that is like, I need to understand what the model was doing. And a lot of these models are kind of, they've got lots of meta levels. It's a bit like having a conversation with someone and you think they're understanding, you've got completely different images in your head. And then you can leave the conversation, they're going to do something else, the completely different to what you thought they wouldn't do, and you're going to, likewise. So you can have what you've got to be able to do with these systems to somehow ground them and understand conceptually what's going on. In science, that's really important. The only things I have exception to in AI, my hard lines are, the AI's are not intelligent, right? They're not intelligent. They're not intelligent, they're not intelligent. How do you define intelligence? Yeah, that's the question that makes sense. Because people talk about AI, AGI, and superintelligence. What is superintelligence, everybody? Like, is superintelligence like some kind of magic that I don't understand? What is intelligence and what is artificial general intelligence? Let's go all the way back and just, I've got very strong positions on this, which I think will, which will, will endure the time. So intelligence, I would define to be a property of an entity to solve problems that will allow that entity to survive. So this is like a biology, right? So now, a general intelligence would be an entity that can solve almost any problem that's coming at it to survive an entity that cannot solve any problem. So you've got, you've got, basically, you could say your entity is really good at mathematics, right? Now, people will say, well, the AIs are good at mathematics, and they do solve problems. So now we're saying, right, well, I think that the AIs are very good at taking a prompt, a problem that you give it and say, please, when you've got this data set, help me interrogate this data set and give me the outcomes. And so what I'm playing with right now is trying to engage with people positively about what we mean by intelligence, right? And intelligence is a very interesting thing. And what I will say to you is that we do not know yet how to define or measure it. I'm working on a problem, a shadow problem that I'm uniquely unqualified for, or maybe uniquely qualified for, which we'll have a go at measuring it. And I'm really inspired by going back in science, looking at problems and discontinuities in problems. So the one that I really like is like the concept of temperature. We knew in the part or gravity, but let's take temperature. We know that things are hot and cold, but what is that? Oh, that's hot. Oh, that's cold. So that we didn't really know how to measure temperature. We knew some things were hot, some things were cold. It wasn't until humans, scientists were able to, well, actually technologists would say, right, what do we need to be able to do? We need to measure temperature, right? How do you measure temperature today? Well, the good old days, you'd have a thermometer. The thermometer could have a number of different liquids in it. And you want a liquid that would basically, as a temperature goes up, it would expand and hopefully have kind of a linear relationship. So whether that's mercury or alcohol. And so if you take an alcohol and you, so for that expansion, so let's go back. So let's want to say, temperature, we knew the things are hot and cold. We didn't have a scale. How do we get a scale? Well, we realized that liquids expand as they get hot and they contract as they cool down. Great. So let's build a system whereby we can measure the expansion of liquid. What technology is required when we need a glass? And what property do we need that glass? We need the glass to be uniform, because if you made it fat and thin and fanning out, maybe a triangle shape, that ain't going to work very well. You needed to make a really nice tube where the sides of the tube were parallel was a perfect cylinder, was perfected to get. Well, back in earlier times, there were some glass blowers in Italy that perfected how glass blow tubes. Suddenly the glass blow, glass blow, the tubes, and then you put the alcohol in, you could go around and measure the temperature everywhere. Suddenly, oh, it feels cold here. We'll go Antarctica, this temperature. And I'll be going here. And of course, the Americans, we're going to have Fahrenheit. I don't know where you guys got Fahrenheit. I feel like Fahrenheit a lot better. It has more units. Like a Celsius, a one degree Celsius jump is a jump. But if it goes like 82 to 83, it's, you know, I'm wearing the same thing. Sure, okay. We got that right, you got that wrong. Just take the out, please. Well, there is this temperature scale called the absolute scale and Kelvin scale. So one degree Kelvin is the same as one degree Celsius. It just goes down to -273 point, something or other. So that means zero Kelvin is the lowest temperature you can get. You can't get any colder than zero K, right? In fact, you can't achieve zero K anywhere. On Earth, we've built some really good systems that you can almost get to absolute zero. And what that would mean is the molecules of absolute zero wouldn't do anything. They wouldn't even vibrate. They'd just be like stationary. But anyway, so I digress. Brilliant thing about temperature. Technology was built to measure it, right? We knew things were hot and cold before, but we didn't know what quite what it meant. Now, intelligence, the problem is human cognition and psychology and IQ tests and all this stuff. And we have people saying, well, you know, this AI will be able to solve any problem from a human. Their misunderstanding, what intelligence is. Giving a algorithm a series of queries to solve a mathematics test is not the same thing as a human solving an unseen problem. And intelligence is actually the ability to solve an unseen problem. Now, unseen is hard, right? Because there's not in my training. I can't solve it. So this is the problem. But when I actually have the problem and I've solved it, I can train on it. Am I way oversimplifying this? If I put it in the bucket of almost what you're saying is logic versus the ability to be creative? - Well, we'll get there. Logic is so the way to computers are quite good at using logic, you can code them properly. There's some good programming languages to do it and they can do lots of things. No. What I'm saying is something super concrete that intelligence is a thing we don't know yet how to define or measure. And I think for a little hobby, 'cause I'm basically bored and I haven't got enough work. - You're bored. - Far from bored and I'm too busy. But I think someone needs to have a go at this. And of course, I'm uniquely unqualified, so maybe that'll make me. I've almost finished it, so it's not that hard. So, yeah, you're so busy, you don't change clothes. - I have 12 different, I think 12 pink shirts in play today. - All right, I'm gonna need to see some elements in it. 'Cause it looks like the same shirt is like, probably. - I don't know, you can check. Maybe the other one was a wide collar one. These are narrower collars. Evidence, I was gonna wear a different pink, anyway, doesn't matter. - Sorry, it looks sharp. - It's fine. - But you're talking about intelligence. - So the nice thing is, I don't have a cognitive load, right, when I dress, it's just pink. - It's good. - It could be like Mark Zuckerberg, and he just wears like, I don't know, blue, gray, gray, I think. - Does he have like a, like a gray. - Just wears gray. - He's kind of like based with fashion now, like he's always wearing different chains and shit. - Sure, I don't know. I mean, I mean, I mean no billionaire. - Got it. - Me just poor person. - You just raised like 70 million, Millie, I don't wanna hear it. - Sure, I mean, but I raised it to change the world and chemistry, not sort out my wardrobe. - That's right. - But fine, we can go back there. But let's go back to intelligence measuring, 'cause I wanna linger on it. You can measure temperature, and therefore we can have a scale and got there. We don't know what intelligence is. We've got the IQ issue. We've also got the race issue that goes along with it, right? There's out there, like, and also the genetics issue, like, right, which was all kind of like crazy, 'cause psychologists and cognitive psychologists and so got confused. But then we got people, people confusing, the AI is getting good at stuff as being intelligent. And so, you know, I think Elon keeps saying, "Oh, the AI GROC is going to be a PhD level in all these subjects right now, and therefore no PhDs. And that is such a fundamental fail." Why? Because PhD, so let me explain how the AI's work right now. And I'm going to convince you, and hopefully by extension, anyone listening or watching this. So, AI's today do the following thing. You get a lot of data, you take that data, and you train a model, that's right. And that model will be able to tell you what's in that data, right, and find your relationships in that data. Great. So, that data, if I say at the internet, all the problems solved in the past are encoded, they say all the problems solved in that past are captured in that dataset. So when you give the AI a problem that's been solved in the past, it can solve it, right? That's not intelligence. That is incredible retrieval, right, and retention, exactly, right. And sometimes the AI's on that good at because they hallucinate and so on. They hallucinate. Well, because of the way the structures of some of these models work, they hallucinate. And so, what happens is that because they have to give you back, back a output, the, if you say, you know, it's less problematic now because the, the frontier labs, if you can call, I don't know what they call the frontier lab, the top AI providers are very good at basically having what's called a, having a checking system, they might say, come out with an answer or check it. You're like, you go to a chat GPT and say, you know, calculate, do a, do a calculation. It used to just retrieve it using the model. And of course, the model would not get it right because the numbers, the, the training data and the internet will humans kind of put that by humans. And so it would just give you random number back. And now that it was, oh, I recognize this is a calculation. Let me bring up my calculator tool, press the buttons on the calculator and get the number back, right? So the AI is quite good at doing things like that, but, but the, the, the thing is the AI's are trained on this huge amount of data and all the problems that, that, that we've solved up till now and humanity are in there. So it's potentially that's awesome. But a PhD scientist mathematician is not about understanding the past. The train on that is about using the scientific method to solve problems. So whenever you give a new problem, there's not in the data set to the AI, it can't solve it. It has enough data and enough retention from all the things in the world that it can keep at one time that a human can't such that it can find abnormalities or holes or something to be able to then, for lack of a better way, putting it, plug it in a way, in a faster way than a human could. How's that not intelligent? So I, I would not, so because if we, again, we have to define intelligence. You can call it, I would call that very useful, right? My calculator is very useful. There are many tools I've built that are, that have been built that are useful, that enable me as an entity with, you know, with human rights, right? To do stuff. But Einstein with a pencil was much more intelligent than Einstein without a pencil. Because he was able to use the pen and paper, pencil and paper to think. So it's, it's almost like a cybernetic thing. So we get trapped in these circular arguments. Are the AIs intelligent? To are the AIs, do they have free will? And are they sentient? Right? Is where we kind of confuse it. So people use intelligence as a confusion. Intelligent entities can solve problems and want to persist in time. That's called biology. I would say intelligence is a property absolutely reserved for biological organisms. Right? Now, what are the AIs? We can call them maybe augmented, augmented informatics, that would be the word. Identified informatics. That's a lot to say Lee. I know. It doesn't sound as good as AI. Right? Artificial intelligence sounds great. Right? I'm like, wow. Right. So there is a good, there's a culture, the zeitgeist I see shared with it. But again, the AIs aren't magic. They are incredible, the incredible models. In fact, they are revealing things about human language and cognition that we didn't expect. And that that isn't like what the, the, the, the AI. So again, I'm not an expert. So if you train an AI in a certain, in a, in a certain language and then you train it in another language and you look at the representations of those entities, let's say, king or queen or brick on cement, they occupy a space that is pre-lingual, right? Which is conceptual and, and, in English, actually, what do you make of that? Oh, yeah, yeah. Basically, there, there, there was a mathematical representation beyond the language in a space, right? Okay. Which is really interesting, but not that surprising, because humans use language to represent concepts. So if I basically take the, the Spanish and I distill Spanish to find out, you know, love and hate hot and cold up and down and prop them in some space, and I did the same in English. And you overlay them. They'll, they're very close. This is why AI translation tools are amazing, right? Yeah. I don't know if, um, I haven't tried it yet, but this is great German, I, I, I, I, I learned German. I'm not very good at it now, but it's great word in German called gamutlick, gamutlick, gamutlick, which is like loosely translates as cozy. So I don't feel like gamutlick, because that's just gamutlick. Yeah. Okay. You go into, if you go into a, if you go into a pub in the UK and it's a windy, winter night and it's cold outside, there's nice fireplace and beer, you go and go, oh, that's cozy. Nice gamutlick. Yeah. Germans have it, but it doesn't quite translate, right? Like it's a, it's a different kind of words. So there are cultural representations and I'm, uh, um, the Inuits have a lot, right? Because the Inuits have to live in the cold a lot, right? That's right. So there are lots of terms, but let's go back again. We don't know. Can we measure it? We don't have a measurement. We know what temperature it is. So, so we've got an idea that these AI tools are able to do very important, very interesting things. They can automate a lot of tasks, but I'm going to make one again, one statement and I'll keep coming back to it. The AI only knows what it's trained on and it can, and if something is embedded in that data and you query it correctly, you can get that out, which is great, right? And the AIs are showing, there's a huge amount of knowledge that we haven't explicitly found in the internet, in human knowledge, and we can pull out, but it can't use the way the humans query and reason over time to put together patterns to be able to predict how they're going to query in the future and potentially outrun humans before we can get there. Um, I don't, I don't understand that statement until I understood all of it until they're outrun humans, till we get there. Because if they can patternize everything, like let's, let me give an example because it's a little bit complex. If I'm talking to an AI and a hundred other people in a population are talking to an AI and we're all trying to figure out something about paint and we're asking all these questions about different colors and how we paint a room and whatever. The way I ask questions in a series of queries, even if it's very similar to the next person, the wording might be different or a specific thing I may be interested might be different. But the AI is collecting all of these different at, at, at volume and at scale, all of these different ways of querying things so they can try to pat the AI, can use all of its computational technology to be able to patternize what everyone's saying and be able to predict what the optimal questions might actually be. So in that way, the way I see it is it could be outrunning humans. Front running them is maybe a better way to put it. So yeah, this is really interesting, so no, no, no, that's not what's happening. Thanks for clearing that up. So I think there's a fallacy here, which is super interesting. And again, look, I'm very happy to be wrong on this, but I want to come to it with data. So AI's are prediction machines, humans are creation machines. And so, AI's, when the creations are kind of shallow, the AI is great at finding them. If you take all these air dish problems, all these ways that you find, you've, in mathematics so the AI's are solving, I mean, that, and you find all these, disproving certain conjectures. AI is a very good at disproving things by finding a counter example, but AI's, an AI has not come up with a single creative mathematical act, right? Hasn't come up with a single creative scientific act. Now, people will argue about this. Now, I want to actually stop this argument because I actually obscure what the AI is a good at. And I tell you, in the last two years, last one half years, the AI is of supercharged my ability to be creative, because I don't have to spend my time doing things that, that, um, right in code and collecting data and processing things. The AI's are splendid at it. They're really, really great, but what they can't do is they, they can't create. They can appear. to create, right? And those creations are relatively shallow. But I would say there were, and this is where I'm really interested in how, where intelligence is. Intelligence solved. Intelligence I would measure is by the ability to solve problems that are unseen. AI's are able to solve problems that are seen or embedded in their data. That's probably called, I don't know, we should call it a different word. Like, it's not not a problem. So, AI's are very good at doing new, can do new things, right? But they can't do novel things. And so the question is, what is the difference between novel and new? And here it is, right? New is something that you can get from your data set. It's just a combination of things. It's fairly shallow. Novel, so you can search for new something new in your data set and find it in the end. So when the AI does something and you've got all that's exciting, you can see that it is just new and it's entirely embedded in that data. Humans are able to do something which I call novel. And novelty is defined by that thing that is in principle not predictable and not embedded in the data. But when I then do it, it's a bit like fashion. Fashion is a music. That's right. You're not going to know when humans are going to do. You might have me back on your podcast one day. I might come back wearing a color you weren't anticipating. Is that new or novel? Well, it's novel because you're in the prior data. I'm just wearing pink, right? The AI is just going to make in the suit. So when I come in wearing, I don't know, turquoise or something or whatever else, you'd be like, that's novel. And so, and it's very interesting that humans are kind of misunderstanding here. But there is a certain amount of misscelling here because we're trying to sell AI's intelligent going to replace PhDs, replace this. And human beings are going to use the AI's to massively flex their cognitive capabilities, right? Not replace them. This is where everyone's going wrong. AI's going to replace science. No, they're not. I mean, because science is about problem solving. Well, you actually just made a point a couple of minutes ago and I can't believe I've never looked at it this simply. It's like the most obvious 30,000 foot in the air point. We've had this AI crisis now and putting that in air quotes, being talked about extensively over, I'd say the last decade, but particularly the last three to where everyone's using it on LLM's and God knows what they have behind the scenes at DARPA, maybe you know, but like they haven't created their own E equals MC squared yet or anything like that, which is making kind of making your point for you. Like there's not something they have not figured out the laws of quantum or I'm just making up things right now, but these things that scientists are constantly trying to test every day. Therefore, not only have they not replaced the highest level guys like you, they haven't replaced science in any way at all technically does point. Well, I would say people have used AI to come up with very good new ideas for making new drugs, folding proteins, solving some physics problems, but these problems are already well defined by the human. That's right. For data processing. So the AI is a great tool. They're a tool. I don't understand. I think there's something really interesting, sociologically happening and culturally and technically happening. Now, now I'm not going to geolocate it. You can all guess, but we have this problem right now where people saying, the AI labs need a lot of money. They need a lot of infrastructure. They need a lot of energy, right? And we're saying in academia, there is this political push against academia. Academia is full of left wing people who basically, you know, don't critically think and all work and so on. Therefore, let's just replace all academia with AI as do science, right? Now, that's just such a fail on all levels. Sure, academia has problems, right? There's but ideologies and things everywhere. But given the AI providers, builders don't know what science is, right? They simply don't. They're populated by computer scientists and mathematicians. Now, and they don't and that's not me saying being a leotist is just like solving scientific problems in a wet laboratory, whether some biological lab or chemistry lab or a physics lab or you're trying to do some mechanical engineering. Problems are very visceral things, right? And it requires creation, like the governor in a steam engine. You know, someone's like, oh, yeah, I need a why do you make sure I can the boy that doesn't explode? And I need to do this thing and you know, build stuff. We interact with the world. Now, we will be able to build models and whether we build I'm building a world model for chemistry. What does that mean? Well, I'm just learning all the chemical reactions by doing them. And then once I've learned them, I'll then make any molecule. That's pretty cool. That's very cool. But there's not magic is just like I have to collect the data. I have to have a verifiable process and then I have to be able to then run it and get it. And so I think the science has never been in a more exciting place, right? And it's just a shame that we have this culture war that's going on at the same time because we have to make a decision. Do we want to spend all this money on CPU on GPU CPUs, water, energy, whatever to basically crunch the data and give back products to people or do you want to spend the money on humans doing that as well, right? So it's kind of like a choice. You could spend money on how much does it cost to produce a human that can solve a problem that you can use chat GPT for. There's not magic, right? You just say, well, if I want to replace a human with chat GPT, I can. But what is the cost? And I've got the human out there who's now got not got a job to do. Well, maybe the human will do something more interesting. But there are so many different things. There's many layers. Maybe we can to go back. I go back to one fundamental thing. We don't yet know what intelligence is. We haven't measured it. We're using AIs to do very important, very interesting things at scale and at speed. But it's not diminishing a human. So you said at the beginning of of when you were explaining this with the intelligence loop, you said there's three layers. We often talk about intelligence, general intelligence, and then super intelligence. So if I'm understanding you correctly, because we can't even measure the bottom layer, you're saying it's like 10 leaps and a fucking skip to get to the next layers, obviously. Yeah. I mean, AG, so I think that the providers of AI, right? And I actually like the products. They're great, right? I don't know about your glasses though. They were doing some shit math before. I might want to check those things. Oh, yeah. I'm not doing any product promotion on your products. But they're not metaclosses. Yeah, but metad might have actually gotten that right now that I'm trying to show for them. No, no, they worked. Anyway, a little after me. Okay. But so let's go back. So we say AI, so AGI, lots of people will say that an AGI is what the term we give to something that can automate at most tasks. Now, is that an AGI in artificial general intelligence? Or is it just an artificial general, I don't know, automata? So like, you know, it would be great if I have an AI that can clean up my emails. But actually, I mean, I do use an AI to clean up my emails. I use one that's about maybe eight years old. And it kind of, and what I do is I'm obsessive email cleaner. I answer all my emails that I want to answer. And I basically file everything and I keep them all. And so I just file them. And it's a, it's a, it's a filing system. So I, I just, because then when new stuff comes in, I don't get lost in the noise, right? I look at my emails. I get so many emails every day. It just, I find it incredibly depressing because I lose stuff in the noise I need to answer or get, you know, I've only got so much cognitive bandwidth. So an artificial general intelligence that the frontier labs are discussing is literally an entity that could just solve a lot of problems, right? Do, sorry, do a load of tasks. Does it solve general problems? No. Humans are good at this. And I think, you know, maybe there are some jobs that humans have that are filing. But we are failure, maybe in, in human society right now, as people go from school to university or wherever they go into a job is maybe we need to generate jobs that have more value and meaning. Right? We can do that. And then I provide a great chance that now let's go to super intelligence. Well, Nick Boster made up this term. Yeah, we got off this last time. We didn't go deep on this. I've read that whole book. I'd really like to, it's just making shit up, right? It's like, it's like, if, look, if you can't define a thing and you say it's super, I mean, I could be a super chemist. What does that mean? I mean, so a super intelligence for me would be to say be somehow that's beyond our current understanding of physics and mathematics, like magic. Again, super intelligence plays on this kind of all. There could be this entity that's way more intelligent than me knows more. Right? Human beings are able to discern how the universe works. We're able to understand far more than maybe evolution as equipped as with it would seem. Now, what the douma says, well, this super intelligence could understand how to, you know, do things that could basically end humanity or build a super duper weapon or something. Well, I don't understand the evidence for that. So what I try to do is deflate the term super intelligence because I don't know what that means. Now, It could mean, let me make it, take a stab at three different things. It could be it's just a faster way of thinking. Well, sure, chess computers can do chess really fast. So I could buy that, but why call it super just fast? It could be that it is able to play through some kind of chain of reasoning, how I can basically look at decision theory. And if I plant these seeds and these decisions, I can basically create disruption. Thank you guys for checking out this clip. If you haven't already subscribed, please subscribe and hit the like button on this video. It is a huge, huge help. And if you'd like to check out this clip, full podcast episode, that link is in the description below or right here. And finally, you can follow me on Instagram and X by using the links in my description below.