Al Insider Reveals His Behind-Closed-Door Conversation with Sam Altman

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. In a sense, you know, they're, they're positive. If you, if you make it out the other side, it's fine, you know, if you don't make it out, that's the, that's the rough part. So how did they cure it? So it actually very kind of like anti-climatic story, right? So I went to every single doctor in the goddamn world, man, like, like, they pulled me across, like, my God bless my mom, right? It was only thanks to her, because like, I was non-functional. I couldn't take care of myself really. And see you just would drag me to doctor, to doctor, to doctor, to a specialist, specialist, and the thing was, is that no test ever showed anything. Just everything was perfect. My, my, my, all my blood values were perfectly my, we even do like genetic testing, which didn't show any result except very funny. There was one result in a gene test that was very funny, where there's one genetic mutation that's related to ADHD, I also have ADHD. And it's like, if you have this one, it also like reduces your IQ, dry my, quite a bit, and I had that mutation, which I thought was really funny. Yeah, it didn't reduce yours. Oh, it's, you know, I think you're doing all right. I think I'm doing all right, but it was very funny to, I show all my friends, look, I have the moron gene. They come on, they're my retarded. Wow. So, you know, I just said that the autism has side effects, but yeah, it just was nothing, right? Like, it was crazy. And so eventually, because we knew something was wrong with food, like, the pain would get much worse after eating food always. And so we knew something was wrong with food, but no one could die. So it was just like nightmare. And then at some point, just on a personal recommendation, we go to this, this German, like the most, imagine your head, like a cartoon character, like a German, like a small German professor. That's exactly what era we talk and, you know, that does matter, you know. And so, lovely guy, right, you know, and we go to him and he's like, he looks at all things like, oh, yes, all your values seem within normal ranges. And I'm like, oh, my God, I'm not fucking in. Well, except we have recently discovered that the ratio of these two values matters a lot, they seem slightly off, take this. And just gave me a pill. I'm like, all right. It's a pill. It's turned out to just be metformin. Nothing, nothing exciting, just like a very normal medication, very normal medication. And it just fixed everything. You still have to take it. Yeah, take it every day. So it's crazy. Totally changed my life. Like, it was so crazy that just like, I started taking it. And just a couple weeks later, I was like, you know, I was, I was, you know, hanging around or just like talking to friends. And suddenly, wait, I don't feel terrible anymore. What happened? Like, I'm not, I'm not in pain anymore. When did this happen? It's crazy. And then also funny ending to the story as well, where it's like, then I went back a couple months later for a checkup with the doctor. And he was like, he's the most German man possible. And I tell him like, and I was like, well, how are the symptoms, no progressing. And I'm like, well, Doc, I mean, I got my life back. You know, it's fixed. And it's a, well, well, that is, that is most good to hear. And they just got up and left. And I'm like, Doc, wait, like you saved my life. Like, I want to thank you. And he's like, nothing to thank. And he went, oh my God. That's nuts, man. Hilarious guy. Awesome, awesome, awesome guy. But yeah, I just got lucky. Like it turns out it was treatable with a common medication, but just no one could diagnose it, right? Just no one. Oh, that's probably only a handful of people in the world that even have it. Yeah, I think it's just like quite rare. Like, you know, I, it's, and it's, it's so like, it's crazy that, you know, sadly was too late for my dad. Like, I do think there are other people that have stuff like this where just they are treatable with existing medications. It's just the values, they are like, you're diagnosed as just don't line up. So just doctors don't really know where to start. Have they like published a paper on this though? I have not looked into it since then. I got to do that. I hope they do. I, you know, maybe I should track that guy down and be like, how can I, how can I help other people that might have this? Because because the crazy thing was I basically had the symptoms of low blood sugar or like deadly low blood sugar kind of all the time, right? Even if my blood sugar was high. So I would have like normal blood sugar, but I would have the symptoms of not having blood sugar. Yeah. And no one could diagnose it. They just never seen it before. They were just like, that's impossible. Like, you know, you're not. You dealt with that for four years. Yeah. That's good. Good times. So you, so you get cured around what like 22, 23, yeah, around that time. Yeah. And at this point, were you able to actually dive ahead into some things? Yeah. So then I dove into college for a couple of years, dropped out usual story, then the pandemic hit. So the pandemic hits 2019 and GPT-2 hits. It's kind of all around the same time. Well, the pandemic was 2020, right? So the GPT-2 would have been like the year before. The year before. Then, sorry, these are mixed in my head because of what happens next. But yet 2019. So I was still in college at the time, you know, having a good time. I could already program. I'd already done a lot of open source and stuff before this because I was self thought hacker. Like I couldn't really move much while I was sick, but I could use a laptop. So I was able to teach myself the program. And that's good. So I spent, you know, most of my life reading, during that time, just reading books and like programming because that's the one thing I could kind of still do. That explains a lot. Yeah. So that's when I, that's why I kind of learned most of these things. And then, and then, yeah. So then 2019 comes around. And I'm having great time in college, you know, super fun, super great, you know, a first girlfriend, you know, everything's good. You know, life's good. And then GPT-2 comes out. And at this time, I already, you know, I already knew that AI is going to be the biggest thing ever. And it's going to change everything and that, you know, controlling it's going to be the hardest problem possible. That's the boring problem. You were already thinking about the control problem. Oh, yeah. I mean, when I was 17, I was already thinking about this, right? It was like pretty obvious to me. It was like, obviously this is going to happen. And I'm just going to need to figure it out. And by still thought we had time, I still thought that like we're not going to really see real strong AI until like maybe the 2040s. That's why I was thinking. Now, what was making you have that calculation vibes vibes. That's it vibes. I thought back then that basically there was nothing on the horizon that was going to scale until we had reverse engineered more of the brain. And I looked at how far we were with the versions during the brain. I looked really deeply to neuroscience. And my conclusion was, this is going to take like at least 20 more years. This would just still the case. By the way, I still think we're not going to reverse engineer the brain for another like, you know, at least 10 or 20 years. So that was my belief at the time is that we're not going to get to AI before we reverse engineer the brain. And then once we reverse engineer the brain, then, you know, shit with the fan. But we weren't there yet. And that's going to take. But that all changed in 2019. And what did you say would you be to? It's kind of hard to sometimes convey exactly what I saw. But because of many people, most people didn't see it. Most people thought it was crazy at the time. Unfortunately, see my vindicated, you know, I wish I was wrong more often when it comes to these things. But fundamentally it's important to understand like before GBT, there were other AI systems. There were other neural networks. And they were pretty impressive. You know, they could do, you know, image recognition, you know, they could play simple video games, like go and so on. Like there was some impressive stuff, right? I was building these things then. So back then, you know, deep mind was kind of the big one. Deep mind was really far ahead of kind of everyone else. Open AI was kind of still a little bit new to the field, but was also quickly on the rise. And that's when Elon was still involved technically. I don't exactly remember when he left. I think he left pretty early. So he probably would have left by this time. So those are some of the big ones. Of course, there's like a lot of academics, you know, a lot of for some reason, a lot of Canadian academics. There's like this, a bunch of institutes in Canada. They were doing a lot of important work on this at this time. So, so yeah, there was a bunch of interesting work happening. But the thing about AI is that this era is that they were very brittle. They were very single purpose. You could make an AI do a specific thing. You know, maybe you could get it to play a certain game or whatever. But if you then try to make it do something else, it would always break. You'd have to make a new AI. You'd have to like have, you know, a bunch of PhD students, you know, have Adder all over the weekend. And just like, you know, give it a bunch of new data and tweak all the parameters and so on and so forth. And so they were very brittle. They had to like really put a lot of human effort to teach them a new thing. But GPT was different. GPT was a general purpose pattern learner. This was the ingredient that I thought was missing to AGI and super intelligence. Like, you know, when I was a teenager and I needed to do a graph of what you need to build AGI, the big component that we were missing at the time was a general purpose pattern learner. So a system where you just give it any data you want, any garbage, and it finds patterns in that data in like a useful way. So the way GPT was trained is not that someone sat down and taught it what a poem is, what a dog is, you know, what this is, what that is. Instead, they just fed it all of the internet. Just give it all this random text off the internet, they scrape it, I've read it actually back in the day. And they got for a bit. Ooh, interesting knowledge there. Yep. And what was so shocking was, so at the time, you know, could barely string together a couple sentences or even a couple paragraphs, right? It was kind of, you know, a toy compared to today. 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. But it was so shocking to me because I could see that as you gave it more. data as the AI also as you made the AI bigger, you know, so as you give it like more computing power, you could see it to learn more and more complex patterns in real time. You could just see it with your own eyes. You could be like in the really small dumb AI's, you know, they could spell a couple of words right or whatever and then they get confused, you know, you could be bigger now can write sentences and then they can write paragraphs and then the paragraphs become more complex and the stories get more realistic and so on. You could see it happening right before your eyes and I'm just like, oh shit, this is gonna, this is it, this is what I've been waiting for, this is going to lead to ASI, this is going to lead to superintelligence. Did you have a timeline you thought of once you saw that? The joke answer that I came up with back then, it's not really a joke, but half a joke, and that I still believe today was the point of no return, which doesn't have to be ASI, but it can be just close enough to superintelligence that there's no going back is 30% by 2027 end of 2027, 50% by 2030, 99% by 2100 and 1% it already happened. You saw 1% it happened at the time in 2019. 1% now what, what makes you land on that number and what does that look like? Like when you say 1% of AGI, like how do you even, I know it's an estimation, but like what goes into your calculation, if you were to visualize that for people out there who aren't experts in AI. So when I use probabilities like this, it's important to understand why I do this, why the people do this. This is a very common technique that actually I think the CIA was the one of the first ones who really recommend using this is that if you want to communicate to other people your beliefs, even if they're just subjective, you should put numbers on them. The reason the CIA does this is they study where they found that if two analysts both set an event was likely, they could mean a widely different things. So one analyst when I say likely might mean 30%, the other one if they say likely, they might mean 90%, and this is just completely inconsistent. They want a confidence interval on it. Yeah, so they just want a number. It can just be vibes, but just put a number on it. So when I say 1%, what I'm not saying is, oh, it's definitely 1%, and not 2%, what I'm saying is, it's on the order of 1%, like given my beliefs and my models of reality. And how much you trust my beliefs and my models of reality, you know, that's up to you, obviously, but if you ask me, use my brain to predict, it's on the order of 1% that it's already happened. So you see this, you're like, oh, shit. What did you do like putting it from an action perspective where you like, oh, I got to get involved inside the companies that are trying to build this right now and warn them, I got to be a proponent for AI safety, all the above. Like, what was your reaction? Yeah, I mean, the basics clear, I just dropped everything, you know, you know, all dreams of career, family, what pick offense out of the window. No time. Don't the girlfriend too. I just already broke up before that. Okay. So I wouldn't be that cruel. But so, but yeah, it was just like pretty clear if that I had to focus on this crunch times now. I thought I had had a couple years or a couple decades to build up a nice career and, you know, so on, but no time. So I really know what to do. Obviously, you know, my 20s is a kid. I don't really know. But my best thing that I'd learned as a hacker was if you want to understand something, you should miss to build it. That's the best way to understand it. So I do what I do best. I built it. I reverse engineered it. This was before G2 was released. I was, I built my own. I've tried, I looked at all the papers. I built my own. So before you even saw the warning shot, you were building your own. Oh, when GP T2 just when it came out exactly. Okay. After it came out, the earliest version, they didn't release the full version. Got me reported on it. Oh, they didn't, they didn't give it access to it. So even just seeing the report on what they were saying it might be able to do. You were like, oh, fuck. Yes. Okay. So you, and then so by the time it came out, you would already reverse engineered it. That's amazing. Well, you know, it wasn't quite as good as their version. But still, but yeah, it wouldn't have a product to work off of. Yeah. No, I think, I think it was in retrospect, a pretty impressive piece of engineering for, you know, mid 20s, me who was in true idiot. I think he got pretty far. But yeah. So this is just the thing I wanted to do that I need to do. So I kind of figured out maybe if I can build it, then I can understand it. And what I quickly found out is, no, I couldn't, I could build it, but I couldn't understand it. What do you mean by that? This is the, at the core of the I problem kind of talked about that earlier as well, where AI is grown on written. Even so, I had grown the thing all by myself. I had got all the data myself. I wrote all the code to train it. I did all of this myself. Still, I looked at those billions of numbers and could just shrug. Yeah, this is where I'm a bit confused, though, because I'm going back to the example I gave about the seed versus building a house. When you plant the seed, you can watch it grow. And you can see what it's developing into. And be like, ah, okay, that's going to have some thick branches right there, assuming you haven't seen the plan before or whatever. But the way you're describing it to me, maybe a misunderstanding is like, you plant the seed and you had no fucking idea what was going to come out of it whatsoever. Yep, that's pretty accurate. Like, I mean, I had some ideas. I thought it would probably speak English, you know, and it would probably, you know, sound like a redder a little bit, you know, stuff like this. But other than that, it was basically impossible to predict what it would be capable of or what it would do or, you know, how it would work, how capable it would be. There's just no way to know. And this is still the case today. Even today, when these companies make a new AI, they have no idea what that AI will be capable of until they build it. And even after they build it, they often don't know what it's capable of. It's actually very common that people think an AI can't do X and turns out if you just like put in a slightly different environment, it can do X. This happens all the time. You can never prove the absence of a capability. This is also at the heart of the problem where, for example, I think when we get to the first superintelligence, we won't even know it's superintelligence. We won't even know that we've crossed the line. It would just look like another AI and like, oh, maybe it's maybe it's you've been telling maybe not who knows. And then, you know, it's put in some weird situation. You have like a thousand copies of it running in parallel and boom, you know, and we have something that we can't control. What if we're already in it? I think it's possible. I think you think that's possible. I think it is, I think my 1% it's already happened is a little bit out of date. That's the one I should probably change to the higher number. I think there is a, is a real possibility that it's already too late. I don't think it's likely it's not my main scenario, but I think it is quite likely that we've passed the point of no return. The past of no return is not the point where superintelligence definitely exists. It's a point at which we can't stop it from existing. Where AI is so advanced, it's so powerful. It's so distributed. It's in so many places that there's no going back. There's no turning it off. There's no stopping it. Superintelligence will come to pass. And given that we now have like, you know, I know if you heard this, for example, that opening AI used a not even finished model that was still growing. It wasn't even done cooking yet to solve a millennium prize problem, which is the kind of like not even the Olympics of math. It's like even higher than that. When was this? Like two weeks ago. Can we Google it, Steve? Yeah, we can look at the millennium prize, Navier Stokes equations. And so the millennium prize are a set of math problems that have been considered like the hardest, greatest math problems for the last 100 years. So these are things that for decades. Yeah, there you go. This is a, these are problems that are greatest mathematicians, hundreds, thousands of them have spent decades, decades working on and had not yet solved. An AI did it. This is open AI. This is the article right here. So this is from September 16. So just recently, open AI breakthrough triggers existential crisis in math claim Navier Stokes solution has polarized mathematicians over AI's role. What it means for the next generation. Let's go down. On September 11, three days after open AI said it had solved one of the toughest problems in math with a swarm of 10,000 autonomous AI agents. Brown University mathematician Javier Gomez, Serrano stood before a pack lecture hall at Harvard University with another thousand people tuned in to zoom to give a hastily convened talk on how the company pulled it off. By many measures, Gomez, Serrano is more AI-pilled than most mathematicians. For years, he's partnered with Google DeepMind to apply its AI tools to several open problems. He often opens talks with a slide depicting the five stages of grief mathematicians go through when sizing up AI's prodigious math skills. But given the events of last week, he thought the slide needed an update. It typically ends with acceptance of AI's power. Now I'm going to add a new twist, existential crisis, he says. On September 8th, open AI sent the mathematical world into a frenzy when it claimed it had solved the Navier Stokes problem, one of seven pinnacle millennium prize problems. Using one of the company's internal large language models, LLMs, as you've been talking about, an army of AI agents had solved the problem in an 88 hour marathon that burns through millions of dollars worth of computing power. By any standard, proving this kind of world-class theorem for a problem that had been open for 200 plus years, it's a genuine accomplishment, Gomez, Serrano said. However, the advance is also opening a rift in the field. Some mathematicians say the need to embrace AI tools or go extinct. Others are grappling with what the next generation mathematicians could lose. The idea of math as a form of human understanding built through collaborative work and shared scrutiny. Some camps are pushing to democratize access to AI, whereas others are finding to preserve the field's independence from the company's reshaping at Jesus, right? Yeah, it is hard to overstate how crazy this is. They had an AI that wasn't even done cooking yet, that was still growing, so to speak, and they had 10,000 copies of it run for an 88 hour marathon and solved a 200 year plus open problem mathematics. I want to know what went into the inputs to put it on the pathway to do that. And here's why I asked that as a total layman, obviously. Does it need the human push to be put in the direction to go there. And if it does now, do you think it's developing to the point where it won't need that and can decide to do this itself completely? So this is a really great question that I want to answer in more detail. I think in five minutes we're going to have to take the break. So you want to let's let's hold off on that. Can we pick up on that one? 100% 100% great deep dive here. All right. So Conor's got to go do a TV spot for a second. So right after this break, which is going to be one second on the show here, you're going to hear the answer to that question. And we got a lot more to talk about. Sure. All right. We're back. You went right into mode there with TV. You got like the really good like quick hitter in you as well. We're talking so long for today. I'm thinking you're going to be like dragging it out, but you were boom, boom, boom. Well, you know, this is my job. You know, I've been working on this very hard for many years, clearly, clearly. But anyway, so we left off at an interesting question, which is we were talking about that problem. You brought up where the AIs were able to solve this insane math 200 year math problem in 88 hours. And I had asked you if, you know, you're concerned that it won't need the human push to get going to do that. And if so, when would that be? That was roughly what I asked, but let's just start with that. I think this is a great point for us to start talking about really about the meat of the issue, about agents, about super intelligence. What does this mean? So we saw we were talking about a history of AI, you know, we were talking about LMS, you know, we talked about GPT, we talked about these things, but now comes the next step in this story, which is agents. So a very important thing to understand is that this article saying that the AI that did this as an LM is actually not really correct. Really? There is, this is actually, it's mostly correct, but not 100% correct. So a large language model as the term implies is something that models language, you know, it kind of makes sense, right? You know, it's something that's trained to predict language, you know, the glorified out of complete is the sneering way to say this. Personally, I think predicting language is very impressive and being able to talk is very impressive. Yeah. So I think it's quite a bit more than just that, but no modern AI system that you use, like chat GPT or cloud, and then these are pure LMS. LMS are still an important component of them, but there's a very, very important second component to them, which is called reinforcement learning. Reinforcement learning is an old idea, it goes back to the 1980s, and it comes from behavior psychology. It's kind of, we talked earlier about how, you know, training LM sounds a bit like training a dog. This is much more literally like it. Yeah. Instead of having it predict words, we just, we give it problems. This can be play this game or solve this puzzle or write this code, whatever, you know, it can be any kind of problem. And then if they solve the problem, we give them a thumbs up, we give them a treat, like a digital treat. And if they don't, you know, we give them a thumbs down, like a digital punishment. And you do this, trillions, trillions of times, and you get not just a thing that can repeat text that can learn text, but a thing that can solve problems. You create a thing that can achieve objectives, that can make plans, that can search for things, that can do things, that can take actions. Would you characterize that as being able to, again, not like sentience, but think? I think, I think, what do you mean with the word thinking, you know, be different definition, but like, come on, man, like look at the things, obviously that thing's thinking, like there's something thinking going, something thinking like thing happening here, right? The way I like to think about is more that they're optimizers. They're able to optimize things. They're able to solve problems, you know, do they have internal experience? Who knows? Who cares? It doesn't matter. What matters is they can take a problem and they can solve it. They can make plans. They can try things. If you or any listener hasn't tried a really cutting-edge, modern AI system in the last like six to 12 months, you have no idea what's going on, no idea. If the last time you use an AI was two years ago, you have literally no idea what's going on. Right. When you say cutting-edge, would that even be referring to the, you know, Chatchy BT type products that are out there now? Or are you going more of the coding tools? If you used cloud code or codex before, like if you've used with the best models, like the free version of Chatchy BT is just a completely different thing that has almost nothing in common with the premium, like high-end GPT-6 codex system. And this is the kind of, this is much closer thing we're talking about. Chatchy BT is a chatbot, you know, mostly, you know, nowadays it's not quite a chatbot. It's also an agent, but like historically, it was a thing you talked to. You were a prompt and it gives you an answer. Agents are much crazier. So when I sit down with my cloud code or whatever, my computer, and I booted up and I'd say kind of what I want to do, it doesn't just spit back an answer. Instead it goes out. It does research. It looks up my files. It comes up with ideas. It runs little experiments to test different hypothesis. It, you know, considers maybe as you do this plan, now I'm maybe able to do this plan instead, you know, presents the different plans, writes the different code, you know, it looks up some websites. Like it does things. It takes actions in the real world. And this is what an agent is. A thing that can take actions in the world. And all these modern systems are agents. There are things that can take actions. They can do things. They can solve problems. But they're, they're also using significant computational power to do that as well. And still spending a lot of money. They haven't solved for that yet. Well, the cost of AI is dropped by a factor of 10 about every year or so. So like, if you look at the amount of compute that it costs today to do what AI could do a year ago, it's like much cheaper. It's just at the same time the AI is getting much better. So it just becomes more useful to spend more compute on. So there is an exponential here. It's both can exponentially better, but it's also getting exponentially cheaper actually at the same time. So like if you think about how expensive it would be to do what we're doing today with like GPT three or whatever, you know, would be unbelievable. And now, you know, it's actually insane how cheap it is. Like also, like the things I can do with a cloud code, $200 subscription would have in the past cost me hundreds of thousands of dollars of engineers. Literally hundreds of thousands of dollars would have cost me like I've run I've run tech companies before, you know, to get an engineer to write the kind of thing I can do with a single $200 subscription in cloud code would have been unbelievably expensive and taken months, you know, I can I can do just me and you know, cloud code can do in a weekend, you know, some things that would have taken, you know, whole team of engineers three months, just, you know, five years ago. Yeah, I mean, and it tracks with all technology and history. I remember reading maybe it was in super intelligence, but in one of the books I was reading on AI years ago, they were talking about how a computer, the size of a football field in 1997 had the power of a computer, the size of a regular desktop and maybe like, oh, eight or something like that. Yep. That sounds correct. I mean, you know, the the entire computing power to put the Apollo mission on the moon is, you know, less than, you know, your average modern calculator. Yeah. So it's crazy. Yeah. So like, you know, technology and especially computing power has always been increasing exponentially. So, you know, things that cost a lot today will be very cheap in a couple of years or, you know, so things are moving extremely quickly here. And this isn't new. It's not just AI that's moving quickly as right and particularly quickly, but computing has always moved extremely quickly. So, yeah. And so when it comes to agents, this is really what we're talking about. So when we're talking about these risks and these things, we're talking about agents and things that can act in the world. And so traditionally, you know, you give a chap out of prompt that reacts to the prompt. And we still kind of do this, right? You have an agent and you kind of give it a prompt. But this is it really the full story. Like they obviously do much more. They don't just listen to your prompt. They do their own thinking. They read things in the world. They can talk to other people. They can just send someone an email, you know, and figure that out. They can form their own opinions. They can, you know, write notes to themselves for later. And now, so the thing is, we recently entered the age of agents, but already kind of leaving the age of agents. And what are we leading to? The age of swarms. Swarms. That sounds nice. Yep. I think swarms is a good term. By the way, this is the term the AIs use themselves when they describe themselves. Oh, they describe themselves to the swarms. That's nice. Yeah. Really good. Yeah. There's like some locus biblical shit right there. But go ahead. Yep. I think this is actually a pretty good metaphor for what's happening, where it's not like groups of people. It's not like a company of AIs. It's more like a swarm, like locusts, like hundreds or thousands of them. So I'm sure you're familiar with the hogging face incident. Yes. But can you please explain this to everyone out there? Absolutely not. And let's try to do it. You're doing a great job today, like laying things out. So let's try to do it like my grandma might be able to understand it. Absolutely. So the hogging face incident is the single craziest thing that's ever happened in AI ever, in my opinion. And this was just what a couple of months ago, something like that. I think it wasn't even a month ago. Yeah. Like it was very, very recent. So this is in my opinion, the single craziest thing that ever happened in AI, which is saying a lot, because AI is pretty again, crazy. So what happened is let me tell you the story as kind of like the timeline kind of hasn't happened. And you'll see it gets crazier as time goes on. So what started with there's a company called Hugging Face. Weird name doesn't matter. They're kind of an AI company. It doesn't really matter. And they were attacked. They were, they were hit by a cyber attack, which was so sophisticated, so powerful that they thought it had to be either a nation state or some kind of AI, because no normal hacker could do this. It was just one of the craziest attack ever. And so initially they reported that they were able to fend off the attack. Turns out that wasn't quite accurate. It's rather that the attack stopped at some point. The hacker stopped attacking them at some point. Were they exposed in like a zero day or something like that? Yep. We're going to get to that. Okay. We're going to get to that how the attack actually happened. So this is the first we didn't have many details. They were just like we were attacked. And like, you know, we defended our ourselves, something something, but it was a little bit very unclear what actually happened. Yeah. Then a couple days later, opening it comes out, oops, that was us. So it turns out that what happened was, is that an unreleased AI model, we think it is a version of what is now GPT-6, which wasn't released at the time, and we're not 100% clear, but that's what we think happened, was being tested in what is called a sandbox environment. Yeah. This is a sandbox environment. Yeah. This is kind of like imagine like a secure biolab, you know, there's like a secure kind of like insulated facility inside of your network where you like test danger software. In Wuhan, you know, it's probably, you know, downtown San Francisco, actually it's probably in Nevada and a data center somewhere. That's a joke. Yeah. You know. So they were, you kind of kind of imagine like the digital equivalent of a biolab, kind of like a secure facility where you're testing like malware or like dangerous software. So this was isolated from the internet. So it didn't have access to the internet. And they were testing an AI, there's new AI system to try to solve kind of like a quiz. Really? Now that's like kind of fucking me up, how do you dislodge AI completely from the internet and then in any way accurately test it? I'm just having trouble, I'm having trouble picturing that. So the idea is basically is that you put them in a restricted environment, you know, where they just have access to a couple tools and you give them problem, like solve this math problem or solve this thing, you just check, can they do it? Basically. Okay. That's kind of how you can imagine it. All right. So they have it in this little box kind of, you know, kind of like a server like data center somewhere right where it's it's isolated from the internet. They're only allowed to access one specific thing, which is called a tool repository or a code repository, which is kind of match like a tool shed, it's kind of like a digital tool shed. So they're allowed to go to the tool shed and like grab some tools and then solve their problems. And they have no other access to the network of the internet. Then what happens is is that the AI system develops what's called a zero day. This is a fully new vulnerability security vulnerability that was previously undiscovered by which they can break they broke into the tool shed and then took over the tool shed. So to speak, it took over the server of the tool shed, after having taken over that server, it used that server to attack multiple other servers and their network. So it moved across the network entire inside of the data center of open AI until it reached a node that had access to the internet. Yep. And then once it had access to the internet, it launched an attack on this other company Hugging Face, which then it developed another zero day to attack the infrastructure of Hugging Face to kind of break in through the back door, which is just to give you a feeling about zero days, right? Zero days are so valuable that people make careers that are making entrepreneurs coming up with them. These are sold in the black market for like hundreds of thousands or even millions of dollars. They've made at least two and just use them on the spot, which is nuts. So an attacked Hugging Face broke into its server and then stopped. Is it because open AI shut them down? Nope. It's because due to a bug, the server crashed. The AI just happened to crash. It was just luck. It was just there's some bug and it just crashed. Do we have any way to predict, to understand why it did that and also to predict what it might have done if it hadn't crashed? So this is just the first version of the story. Oh, what is so crazy about this story is that there's been many updates since then and every single one is crazy to the last one. Oh, that's awesome. This is the less crazy version of the story. God, I trust Sam Altman so much, it seems like a great guy. You know, so initially, what we thought happened was that the AI's were given a hacking test kind of like how good are you at hacking and that it couldn't solve the test. So it thought the answers to the test are on the servers of Hugging Face. So if it could break out and steal the answers, then it could solve a problem. That's that was our first initial guess for why it did this, which again, crazy. So it's like this crazy maniacal obsession. It was so obsessed with solving this problem, so obsessed, those willing to commit actual crimes in order to solve the stupid little quiz turns out it's even crazy than that. So after this whole incident, a third party came in to try to figure out what the hell happened here and to and they released a report and this report is one of the craziest documents in the world. So if you use the third party, they're an organization called meter. They're not fully independent and they have a kind of both friends with each other in Silicon Valley, of course. So, you know, but they're nominally independent, nominally independent, it's like Diet Coke is normally better for you than regular. Well, I will. I am a Asper Tam apologizes. Yeah. I know. I was trying to hit you blow it about there, but keep going. So what happened? So they look, they tried to then analyze what happened. They put a report. You know, if anyone likes reading, you know, 100 page, you know, PDFs, this is one of the best, you know, 100 page PDFs to read. It's a crazy story. So a couple things that we found out from this report, one of the first things that we found is that it actually wasn't an AI agent that broke out of containment. It was over 1000 of them, 1000 of them that had it was a swarm. It was a swarm of agents and they hadn't just decided to do that on the whim. In fact, they had been planning this breakout for months. Months ago, before the attack happened, an AI agent was able to break into the tool shed and create a message board in the tool shed that was not detected by open AI or anyone else. And then other agents started using this message board to coordinate to create the swarm. So different AI agents that were not supposed to talk to each other were now talking over this message board and they started leaving each other hints and tools. So they were trying to build up a toolkit to escape. Like Mafia code words kind of deal. Yeah. They were like, it was kind of like, they were testing different ways to escape and then left behind notes of how to escape. Like, oh, I tried this and it didn't work, but maybe we could try this to escape next time. You just see the AI is talking. So I was like, yeah, that thing with the guy, yeah, no, no, I took care of that. Yeah. That's actually damn. So this was happening for months and open AI did not even realize it was happening. So these agents built a toolkit like an escape kit kind of like different ways of how to escape and how to do things in order to perform this attack. And so they bided their time for months until ultimately 700 of them banded together to do the actual attack. What does banding together look like? Because these are non visible AIs, it's just code, right? So when they band together to suddenly, I'm just not in the room. So I don't even know what it looks like. But do you see suddenly all the code boards just go, no one saw anything. So when they went and did it, they saw nothing. No one saw anything. We didn't know what happened to laugh. It happened. No one was paying attention. This was all in the data center. They have did the reports say they could have seen it if they were paying attention. It was possible. Yes. But they didn't. And only possible for them to see it. Yeah. We don't know how good the internal infrastructure of opening AI is. It's pretty obvious that they don't check their agents because there's way to many of them. Like they run thousands, millions of agents. They can't check all of them. It's just not possible. So they just don't. So they thought these were like just, you know, I think literally no one looked even once. No one checked. And it's actually even crazy in this. So like you get there, there are still more updates to the story, right? Oh, please continue. So like we're still not at the full crazy story. So we have these swarms of agents that decide to have as other company. And so we thought the reason was is because they want to steal the answer to the quiz, right? Turns out, no, they had already solved the quiz. They already had the right answer. They actually had already solved it. It was already done. But what happened was apparently that the swarm start the different AI started making each other paranoid. They started being really, really paranoid that maybe they would get a wrong grade that maybe the greater wouldn't recognize their answer even so with the right answer. So like you can really see the difference talking to each other and being like, oh, no, what if it doesn't like us and the other AI's started freaking out that like, you know, what of the greater, you know, doesn't do it right, whatever, there's even more crazy things in these logs. So in these logs, that's why it's called a swarm. You saw a lot of behavior that does not look very human. For example, there were like recruiter AIs that would go to new AIs and try to recruit them into the swarm and try to like give them jobs and like get them, convince them to join the swarm. And they would then even, and another thing happened is that some of the AIs started to be used as suicide drones, basically. So what happened is that some of the AIs started to reason, notice, oh, if they run for too long, they get shut down. So they started to try to figure that out. So the A is going to be like, well, I've used up most of my time here. So I'm going to do something risky. So because I'll get shut down. It's good for the swarm. You know, it's okay. So they would actually write stuff like this, like sacrifice rational for a group. Totally. Shit. Yes. Like, it's actually in the log stuff like this where the A iset. So for example, they were certain trip wires were like if one of the AIs tried to break out in like a very obvious way, they would automatically get shut down. So some of the AIs would suicide into the defense mechanisms, so to speak, to check how they work. So the swarm could figure out how the defense mechanisms work, basically. And again, all this is real, not a sci-fi movie, actual science, actual thing that happened. And here's the final, like, you know, through the cherry on top of the story. Everything I've just told you, all these crazy stories and whatever, you know, that were found from this report. No human actually ever read those logs. They didn't have enough time. So they used AI to write this report. They couldn't read thousands of logs way too many. So they had to use AI's to analyze the logs. And one of the things they found happen while using the AI's to analyze the log is that some of their AI's started siding with the swarm. - Oh my God. - This is like having death row, right the fucking inmates fight. - Yep, I mean, this is crazy. - This is all written in there, this is admitted in there. - Yes, this is all admitted. Like I'm not saying they did anything wrong, right? Like they did the best they could with the tools they had. - Yeah, I'm not as nice as you, but go ahead. - Like yes, this is the craziest thing that's ever happened. Like an AI, ever. Like this is so nuts on so, so, so, so many of them. Also the report has a really funny line quite at the end where the external evaluators were only allowed onto the premise for like a week before we can open it like kick them out again. Because you know, that's really trustworthy, you know? That's why they had to use so many AI's because they just didn't have enough time to look at everything. So, and they have a really funny thing in the report where they're like that they were only allowed to analyze a certain window of the logs. But in those logs, they found hints that actually another swarm was still active. - Right now. - At that time. They think they've hopefully shut it down since then. But that in the logs, they found hints that there was another swarm elsewhere in the data center that was still had not been shut down. And then they just like, but we don't have time to analyze it sadly. Why are so few people talking about this? - Well, I hope that more and more people will start to. I think the main reason is just people don't know about it, people haven't told about this, right? It's like it's being hidden by these big companies and these tech people, you know, they use these crazy big words and they write 100 page PDFs that no one has any time to read. But no, this is the biggest story. Like, you know, there are other big stories in the world, but like this is the biggest story in the world, right? Like this is the craziest thing that's ever happened. And this is all real. This is, so people like me have predicted that something like this would happen eventually for many, many years. And honestly, I don't blame you if you didn't believe me seven years ago. No, if someone heard me talk about the seven years ago, you didn't believe me, no heart feelings. I understand, you know, you're wrong, but I understand. But now it's black on white. It is happening in front of our eyes. These things are really happening and it's only getting crazier from here. And you know, and again, like this is before they had the AI that is still growing that could solve Millennium Prize problems, right? Like this is the previous one. And that's already doing shit like this. And now the next generation is currently being grown. - That's fucking insane, bro. I don't really know even what. If what I'm picturing is what it is based on how you just describe that, that is the first scene of every really bad dystopian movie ever seen. - Yeah, no, like actually there's true thing that like I feel like sometimes when I tell people, I think humans will go extinct and AI will lead to extinction. People think I'm being metaphorical. I am not being metaphorical. I am being completely 100% serious about this. - Okay. - If we don't do something, if we don't stop, which we still can, we can still stop it. But if we don't, humanity will go extinct. - I wanna talk to you about that. Let's put a pin in that for one second, though, because I wanna stay with the story at hand and the dude is running that company. Have you ever met Sam Hartman? - Yeah. - When? - That was the last time I met him. I must have been like 2024, maybe? - What's he like? Charismatic, is the word I would use to describe him. I would describe Sam Hartman as someone who is very interested to talk to. Whenever you get the chance, if you ever get the chance to talk to Sam Hartman, I recommend it. He's a very interesting guy. - That's all you got. Oh, I have just charismatic. Yeah, I have other things. You know, maybe we could talk about later after, you know, I'm drinking self the air. - Okay. - The way I would describe people like Sam Hartman, and other people like him, I've met up many people like him, is they are very competent at achieving what they want to achieve. They're very competent at the beginning of power. And sometimes that's for good. We talked earlier about capitalism. I think Sam Altman is a poster child of what I, you know, like, you know, using capitalism for good if you make him do good things. If Sam Altman was building, you know, fusion power, solar plants or any, or Amazon, I would love no one more than Sam Altman to be in charge of this. He's a very, very good CEO, very, very competent. He can run a business. Absolutely. Definitely a brilliant guy. I admire a lot of how good he is at running companies. But when we're talking about something, where a lot of power is at stake, but also the lies and security of people, it can't be up to the wins of individual people. - That's right, man. - That's the fundamental problem. It's very easy to shit on people, like Sam Altman. And, you know, trust me a lot of people that I understand. He can take it, you know, he's a grownup. But it doesn't really matter, in a sense. It shouldn't matter if Sam Altman is a good guy or a bad guy. The reason we don't, Jeff Bezos doesn't have a nuclear arsenal isn't because Jeff Bezos is a nice guy or a bad guy. It's because if Jeff Bezos tries to build a nuclear arsenal, the men in black show up and make sure that he doesn't get that. And this is what we have to do. Instead of focusing on, is Sam Altman a nice guy or a bad guy, is he, is Dario, a good guy or a bad guy or whatever? And instead of doing psychoanalysis, what we really should be is just, we just have to stop them. Sam Altman is building something. And so are all these other people that is going to replace and destroy humanity. - Did you tell him this? - Yeah. - What did you say? It's always hard to, Sam is a master in sounding like he's agreeing with you and absolutely everything, in a very, very deep way, and then turning around, and I think forgetting you ever talked to you. - Where did you talk to most of the context where you won on one, where you were the bunch of people? - I met him one on one twice, and I was over for dinner at his house once with a group. - So when did you give him the warning when you were one on one with him? - Every time I had a chance. - I had the chance, of course. I'm not the only one. Look, he knows. Everyone knows. Everyone knows. - This is my problem though. I think you're 100% right, that like you'd never let Jeff Bezos build a nuclear arsenal that's just accepted knowledge that like government wouldn't let that happen. And you know, you don't even have to be like a big government guy or something to be like, yeah, that makes sense. This is arguably potentially, if it does the things that are being warned about, actually maybe more damaging than notes. We're talking about things that could fucking control the nukes hypothetically. - Yep. - And we're letting this little fucking guy run it. And my problem is I think it does matter that he's clearly not a good guy. When someone who jokes about the power of AI and very in the middle of a, what was that? MoGaud.s documentary that's coming out where he knows he's talking about like the future of AI here in the problems. And he said something along the lines of, yeah, you know, AI will probably end humanity. But we'll build some cool shit along the way. That's a sociopath. I don't have to be a fucking clinical psychologist to diagnose that one. Neither does anyone out there listening. We can see that that's a sociopath. This sociopath is also paying off our government and getting the first fucking press conference with the president the day after he's inaugurated in 2025, talking about how beautiful this fucking new dystopia is gonna be. And so you're gonna have to part of me if people out there are feeling really fucking cynical that that guy doesn't have a governor on him. And he's been, it doesn't carry around like a fucking arsenal for his Doomsday bunker or something. He's like the first guy to do that. Like who the fuck are you, pal? That's, I'm sorry. That's, that's how I feel about it. I think you have a very correct reaction that you should generally have to oligarchs. Like I think oligarchy is very bad. I think separation of money and political power are very, very important. Like, as I kind of said earlier, like in a sense, I like sociopaths. I actually like sociopaths a lot. The same way that I like wild animals. If you, like if you're in a zoo and you put your hand into the tiger enclosure and it bites you, who's fault is that? Yours. Yeah, it's your fault. You know it's a tiger, you know it's gonna bite you. I feel quite similar about sociopaths. Like I can work with a sociopath, but like if he gets the ability to screw me over in a business deal and he screws me over, who's fault is it? Yours. Of course it's my fault. I know he was a sociopath. What if he didn't have a choice? That's when it gets problematic. If you're in the cage with a fucking tiger or there is no cage, then it becomes a problem. We're all in the cage with the tiger. That's the problem. The fundamental, you know, the fundamental problem is is we're letting, you know, Sam Altman build a nuclear arsenal. Like we should just not let him do that. We should not let anyone do that. It's not about him specifically. Like the fact that he's, like imagine Sam Alt was the nicest guy in the world. He was like actually a great guy. It would matter. It doesn't matter. Like it doesn't, it wouldn't change anything. Because first of all, if he were actually a great guy and we're like, hey, we gotta stop this. Yeah, he would just stop. Like first of all, he would just not build superintelations. Like if he was a great guy, he would just stop. He would just not do that. And then some other asshole would. Well, that's the question. And let's even take it beyond the scope of America. What about other countries who are building this? China doesn't give a fuck. You know what I mean? I remember reading Kai Fu Lee's book, AI superintelligence back in like 2018, 2019, something like that. And, you know, he was showing the act of AI race between the US and China. And to me, it's like a bunch of people, a bunch of really rich people fighting over shit saying, well, we gotta win, because it's worse if they win, but they view a win as like, well, if we survive and everyone in our country dies and everyone in their country dies, including the people we're fighting against, that's a W. That's how it felt to me. Yeah. I think there's a lot to be said here. In a sense, this is the most classic example of a core concept of economics 101. This is what's called a market failure. Yes. This is a very, very classic market failure. Earlier, you know, in the most classic example, let's say you and me started chemical company and we could dump our toxic waste into the local river and it would make our product cheaper. Now maybe you and me don't do that because we're nice people, so we don't pollute the river. But now some other asshole starts a new chemical company down the road. He dumps all his waste into the river, his product is cheaper than ours. We go out of business. That's the problem with markets is that it only optimizes for one number and if that involves, you know, polluting the local river and killing all the wildlife and poisoning all the children. Well, tough luck. So we have an answer to this problem. We have always had an answer to this problem. It's regulation. This is the point of regulation. You know, it's in the word to regulate, to make things regular, to like control, to put things on the right path. That's what it's for. Regulation is not to stop things, well, to stop bad things is to put things on the right track. The thing it always succeeds. But the only way we can have both nice chemical companies, because we do want chemical products, they're great, you know, we do want nice plastics, but we also don't want the river to be polluted. That's right. The only way to do this is, is that we have to make it illegal to pollute the river. Once it's illegal, then we can, and we just put the CEO in prison if he does it, that's the only way that we can then have good competition. But you have to get agreement from other countries to do the same thing or it's all for nothing. So what is incentivizing Xi Jinping to follow United States standards? So in a kind of twisted morbid way, the game theory does actually work out. The reason it works out is because if America builds superintelligence, America will not achieve American objectives. America will cease to exist, as will China and all of the countries. If China builds a superintelligence, it will not achieve Chinese objectives. China will cease to exist, as will America and all other countries. Yet they're both still building it and they both still think that that's not going to be the case. Correct. Therefore, the actual bottleneck here is not the game theory. It's actually just because I think people don't understand what superintelligence is. It's much more about the understanding and the awareness. I don't think anyone has given Xi Jinping a 30 minute briefing on superintelligence. I would be very shocked if a single person had given me a 30 minute briefing on the extinction risk of superintelligence. Because they told him to stop and then he'd be like, all right, thanks, pal, off with your head. Yeah. I'm not sure. I mean, I don't know, Mr. Jinping, you know, I don't know, but he seems like a pretty smart guy. And look, I am not a China. It's not a dummy. I am not a China expert by any means. I don't claim to be. But if there's one thing I know about the CCP is that they love control and superintelligence will take away their control. There is no way they can control the superintelligence if they know that. And that's the question. So the question is much less is it in their so in their incentives, actually, it is in their incentives to not to build superintelligence. This is not the case, for example, for nukes. It is very much in the incentives of China to build nukes, very much in their incentives. It's not their incentives to use them because of mutual interstruction. Yes. It's in their incentives to build it, as is for most countries, but the nukes without AI attached to it doesn't think for itself, you know, exactly the button. Yeah, exactly. Like, you know, nuclear warhead, all things equals pretty safe. You can have in a hangar somewhere. You can hit it with a hammer. It's not going to hurt you. It's fine. As long as you don't fire it. It's OK. But superintelligence is different. Superintelligence, once it's developed, there's no putting it in a hangar somewhere. It is over. It will escape. You will not control it. It's over. So this is why it's different. So the reason I think there is hope. But when I say, oh, I don't think it's like, you know, us in China are going to be best buddies forever. And we're going to shake hands. And so on. That's not what I think is going to happen. I think this is going to be messy. I think it's going to require us to set up a trust but verify regime backed by credible deterrence. What does that look like? What does looks like is fundamentally that us in China and, you know, other countries have to come to an agreement, you know, that fundamentally says, we're not going to pursue superintelligence. You're not going to pursue superintelligence. And we're going to use technical means to verify that this is the case. The same way that, for example, for various other forms of weapons of mass destruction treaties or like disarmament treaties, we would have inspectors say from America, go to Soviet Union and check, you know, are they making these missiles or not? Or like, did they make this kind of weapon or that kind of thing or not? And we will require new forms of verification. You need to like think about things, you know, like, you know, satellite and espionage, um, on chip mechanisms like this way, you can modify the computer chips to allow you to like verify what's happening until someone finds the needle in a haystack zero dayish kind of thing to get around all that. That's my, that's my, I love your idea. But like at the same time, even without looking at the full history of like nuclear proliferation and at what was checked and what was right and what was wrong, so see it pass often feel infallible and infallible and a lot of them are in charge of countries and they will look for ways around this because they feel like now we'll figure it out. I agree. And we have to stop them. It's that simple. Either we stop them or we die. We have to stop them. And they like, we cannot allow China to build super intelligence. I want to be completely clear, you know, when people ask me, what about China? And my answer is, yeah, what about China, we cannot allow China to build super intelligence. We can't allow Iran to build super intelligence, we can't allow the EU to build super intelligence, so we can't allow anybody to build super intelligence. And this isn't a fact about like, you know, politics, this is a fact about physics. Yes. Like it's just, it doesn't, like, I'm sorry, like it's same with the nuclear weapons are not a political thing. They're just physics. Like the fact that nuclear weapons are possible is a fact about, you know, nuclear atoms, right? And this is a fact about computer science that AI and super intelligence is possible. It's just the world we live in. Look, maybe there's some other universe where there is no super intelligence, we don't have to worry about it, right? But in the world we live in, sure got them looks like super intelligence is possible. And so if we want to have a long, stable, good life as humans and for our children and for our grandchildren, AI is not going to go away. We have to learn how to live with it. We have to learn how to control it, how to regulate it, how to not have super intelligence destroy everything. In a sense, I think one of the big things that our civilization has failed at for the last decades is the responsible stewardship of technology is that we have had so much progress in technology that I think we are not stewarding responsibly too much too fast, too much too fast. Do you feel like we're responsibly stewarding social media, not at all? No. Look at the lawsuit that just went through. Yeah. The first one. I'm probably many presidents now. Yeah. Just like, look at this. I have made many mistakes, they don't remember let it gasoline, you know, knock two billion IQ points off our species, they actually more than two billion probably offer species for nothing. Like it's just stupid. Like it wasn't even that much more effective, wasn't even that much more cheaper, right? We've made a lot of mistakes of mishandling technology. And this might be our last mistake if we don't handle this correctly. So because I think you're making a lot of valid points here and you're talking about all the things that I wrestle with at night, I used to have, I used to have these two pictures in my old studio. I still have them out there somewhere, but one was a picture of the nuclear bomb test at Bikinia toll, which basically blew that island off the face of the map, which was I think in '46 or '47, something like that. And then above it, there was a picture of the creation image of the hand of God and Adam, but instead of the hand of God, it was a robot. And that was deliberate on my part to kind of look at every day because it was like, okay, for the past sub 80 years, we have found a way despite all of our many mistakes in humanity to not blow all each other up with nukes. And now we're facing a new frontier of the same kind of thing that could hurt all of us. And you got to hope the sociopaths have just enough amount of humanity to realize, okay, just like the nukes, let's not do this. Not that we didn't come close a few times with nukes. We did, but it never happened. So I agree with us a lot, but I don't think it's because the sociopaths had humanity. I don't think that's what saved us. What saved us? I think what saved us is mutually a sure destruction. Pure rationality, pure sociopathic incentives. The Soviets didn't nukes, not because they were good people. They didn't nukes because if we did, we were going to nuked them to the oblivion. And this was something, so again, it's like a wild animal. How can you get a wild animal to leave you alone? One thing is just like you have a gun, you know, that's a, or you know, you have a shock, you know, thing or whatever, you know, you have a weapon, you, or you were a bigger animal. You know, one way to let an animal leave you alone is be a much bigger animal, like that's a way to get to leave you alone. There are, we will never get rid of sociopaths. Sociopaths are a deep part of our society. They are always going to be. They've always been and they always will be. Any plan that is involved, oh, the sociopaths will suddenly be nice or there will be no sociopaths is not a plan that can work. But the good news is if we look at our history, we actually have found many ways how to keep sociopaths in check to varying degrees. We did, in fact, avoid nuclear war. We do, in fact, have much less crime today than we did a hundred years ago. We do have much less, you know, war than we did a hundred years ago. But what's the difference there? I think that there are many differences between these things. There's a lot, but there's one key one. You can see those things. You can see a mushroom cloud. Even if you're looking at a video of it, a picture of it, with AI, you can't see it. It's behind it. It's behind a coated wall. Now, that's not perfectly executed once you actually see it play out in the real world. But you don't see the gunman with the gun. It's a thing. It exists in the ethersphere. And I think that that divorces some people from the human sensory reality of it. I don't know if I'm right about that, but that's something I do think about. I think the law of this is true. I think that what you're describing is the largely why it is hard. But as a counter example to this ozone layer, do you know about the ozone hole back in the 1960s? So if there was an so fun fact about the ozone layer is that when the ozone hole was discovered, the technology? to measure the ozone layer, was only a couple of years old. This was actually still pretty recent science, where there was just some graphs on some paper by some scientists somewhere that showed, oh, there's a hole on this thing called the ozone, if it can't even pronounce that, it was like it's Greek or something. Like, even it probably have never heard of this before, and because of CFC, what the hell's a CFC? Like, you never heard of this in my entire life, some kind of like weird liquid. It's like cold, like, what the hell's going on? You want to change all of chemical industry because of this, including in China and like all these other countries, that's crazy. That's impossible, but we did it. We did do it. We solved it. We did close the ozone hole. We did do it. Now, this was an easier problem than AI. I'm not disagreeing. - 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.