The Real Dangers of AI No One Talks About | Dr. Roman Yampolskiy

Is that it's to speak with you on AI safety and potential extinction events that could happen for humanity? I don't know how positive this conversation's got to be for a lot of people, but for people that aren't aware of the work you've done and ultimately how you became one of the leading voices in the field of AI safety, can you provide some background about the work you've done and kind of how you got to this point? Right, so maybe about 15 years ago I was very interested in artificial intelligence and I was hoping to help develop it for the benefit of humanity. My background is in cybersecurity, so I was looking at safety and security issues around AI we had at the time, online casinos, bots and standard internet problems, but as capabilities of AI became greater and bots became more common. I started to realize that this is a problem which is not going to be solved anytime soon mostly because it's an ongoing process, right? AI keeps improving, so we need to keep up. The safety community has to always be a step ahead of those systems and that's what I've been doing for about a decade and I think I continue realizing that the process will not stop. A lot of my colleagues in safety community think or will get to a GI, maybe we'll solve the safety problem at that level, maybe super intelligence, but it will never end. You always have to continue trying to be one step in advance of those more capable systems which to me seems impossible and so overall conclusion from well over a decade of research is that eventually we'll get to the point where we can't keep up, they're smarter than us, more complex and they're limits to what we can do in terms of safety. Yeah, I mean 15 years ago, most people think AI came like two years ago, right? With the rise of judgeability, they don't realize how far in advance that behind the scenes were, I mean, what was happening at that point in the world of AI and what were some of the boxing techniques that were trying to limit the innovations or were there any limitations to slowing down AI at this point because perhaps at that point it was so nascent that there wasn't any level of threats or potential threat. At that point, nothing was really working. People were trying to do a trivial things like a collection of if statements you have a poker bat, if this is the cards you have, try going all in or something like that. It was more about just detecting human versus bat. That's where a kept child, good items came in. So nobody worried about it, really nobody wanted to slow it down. We were hoping to accelerate to where it's actually doing something useful. And things changed as you said two years ago, maybe three years ago, we switched from narrow systems, only capable of doing one thing, like playing chess to those more general models, which really can help you in so many different ways. And they continue improving without additional breakthroughs, if we just make a bigger model, give it more data, train it for longer, they seem to continue improving their scalability hypothesis. So that was a complete game changer. And yeah, at that point, we no longer have a problem of how do we get AI to do anything? It's more how can we understand what is capable of doing and maybe put some guardrails so it doesn't do a little too much in some risky domains. And what was that point of opening Pandora's box in your opinion? Was it the rise and really the introduction of touch PT that put a consumer interface into using these artificial models, or was it behind, was it further behind, or are we not there yet? Can we still have some sort of limitations or ways to restrict AI? So it seems the transformer architecture was the big shift once we realized that it scales, then that was a complete game changer. Open AI, of course, made it commercially viable to scale to a billion dollar clusters. And that's a very important aspect of it. It's not enough to have it in theory. You have to actually have the compute available, training data available. So the combination of that scientific discovery and that commercial brilliance is what got us to where we are today, I think. And the question that everybody wants to know, I'm going to ask it up front, which is, how close do you think we're out to AGI, and how do you personally define it? So if you go back to what we talked about 15 years ago and you showed me or any of a computer scientist, a model we have today, latest model, we'd be convinced beyond any reasonable doubt that we already have a GI. This is it. It can do thousands of different things. It speaks every language. It can help you with math, physics, rights, assays. So understand our definitions. We have a general intelligence. What is the standard definition? A general intelligence capable of working in multiple domains. It's not a narrow system, only playing chess, only driving car. It can do pretty much anything and you can ask it to work in new domains and train it to do novel things. It is in fact, in my opinion, smarter than an average person today. I think it's smarter than most people today, but that's a different argument. And so I think we have something which is general and capable of learning. It is not superior to all people in all domains. It is not super intelligent. It cannot do novel science of engineering at a level of tough people. Once it gets to that capability, it's smarter than all of us. It's super intelligent. It can start self-improvement cycle, creating new versions of AI, which are better than the previous one, and at that point, it has less reliance on human input. So I think this is where we are today. You can argue that that's not true AI, but in terms of capabilities, if you look over all domains and compare it to an average person, it would outperform them. Even if it's still very bad, that would say, stand up comedy. Yeah. This is, I remember seeing when the GROC4 came out and Elon was kind of making this point that if you use GROC4 heavy, the model that is using deep reasoning, it's a PhD level at pretty much every single topic you can imagine. However, it lacks common sense. I find this so interesting. So it's like, it can do the most complex task that even a PhD can do, but it can miss the most basic common sense. So what does that tell you? So that's still be considered AGI when it can go to the far extremes of intelligence in solving problems, but it can't like know the basics of human emotions or intelligence that everyone, if I've five year, I could potentially know. Yeah. So you know how there is this idea of absent-minded scientists, someone who can figure out quantum physics, but has no idea where umbrella, they live behind them. It is very similar to humans. I know many academics who have no common sense whatsoever, they're brilliant in their narrow domains of expertise, but they just don't have capacity left for the realities of life. And I think maybe it's very similar here. So the system is trained on solving really important problems. So maybe there's basics of human interaction is not explicitly provided, but it can certainly learn that the fact is what it chooses to do. Sure. Yeah, it's probably not the priority right now. But what is it about us that continues to move the goalpost of what defines AGI? Because as you said, from a definition, technical perspective, we've reached AGI. But I wonder at what point is it just the fear of humans you think to recognize that AGI is already here. With AGI comes all of these different things, so we kind of refuse to admit that we're already here at this point of AGI. It depends on who you're asking. I think for OpenEI, they have a clause in their contract that if they build AGI, Microsoft is somehow gets kicked out of something and I don't know if they're ready for them. For them, it's an explicit problem in terms of admitting what they made. For most people, I think what you're saying is right. If they find an obvious problem, like, look, it cannot even count the number of letters in a state name or something, then it's so dumb and not worth you that great title of general intelligence. But again, in terms of usage, in terms of what people actually do with it, I think that's like, drop in employee, you get someone, you ask him, okay, I need you to send this email, I need you to book this for me and this is basically what the expectation is for a secretary. Yeah, and it's kind of difficult for the average human who actually thinks linearly to try to project what will happen exponentially, which is what's happening with the advancement of AGI. Where do you see for someone that's kind of been in the exponential curve and studied it over the last 15 years, what do you see in the next 5-10 years of what potentially artificial superintelligence could be and how do you define artificial superintelligence versus artificial general intelligence? All right, so general intelligence is kind of like a smart person who can be dropped in employee for any occupation. Superintelligence is being better than any human expert in every domain. So it is a better engineer, a better poet, a better driver, it's just better. Now it can be worse at something, but it's just a question of additional training. So it can also learn and pick up additional capabilities. It may not know immediately how to play some new game, but it can quickly learn to dominate. So that's what we expect soon after AGI comes, a fully automated scientist engineer. But most people stop thinking at that point, but the process will continue. You'll have superintelligence 2.0, 3.0, and this in indefinite process until we run out of physical resources to train those models. They can always get more data to self-play and simulation and experiments. So it's really just a question of converting more and more universe into compute for them. Is that an easy way to quantify, like, can we have measurable, quantifiable studies or measurements that can say, okay, this is ASI. This is technically where we are at this point or we get under the definition, if it's better at everything you can come up with, then you kind of have to admit at this point that it's designing tests to measure you. Sure, sure. It's funny, though, right? We almost always project something that could be potentially the worst case scenario. And most people, I remember even a couple of years ago talking about AGI, how that's going to be the end of the world, yet we're here technically. Nothing really seems to be happening. I understand a lot of the work that you do is talking about potential existential threats and risks to humanity with the rise of AI. But I guess this is an example, like humans have always, for survival instincts, try to mitigate the downsides. We always have this worst case scenario, but it doesn't seem like much is happening. So can you talk about your perspective of what are the existential threats in different aspects? What are the risks of what could be happening as AI continues to evolve? Correct. So two levels, AI is a tool, something we have today. People use it to help them with their problems. So if you have someone with malevolent goals, terrorist psychopath, they may ask it to help developing a biological weapon, chemical weapon, something like that. Maybe hackers will use it to hack into nuclear facilities, try to interfere with military operations. That's what we expect. We see in some of it, it's not really as bad as some people predicted. So that's a good sign, maybe because AI is also being used as a defender in many of those firewalls and patching zero day exploits, but overall, as long as it's a tool, we understand human preferences, human goals, we can kind of deal with it. The concern is where AI becomes an independent agent with its own separate goals, or maybe side effects of goals given to it, but not something we explicitly control. At that point, if it's as capable as we predict, it can still rely on this methodology where AI understands synthetic bio, your nanotack, but it can also invent new types of dangers. The technology we simply don't have yet. I think it helps to kind of look at cognitive differential between humans and the lower intelligence like squirrels. Squirrels can think about all the ways we would harm them. Maybe they'll have a big stake, maybe he'll throw lots of nuts at me, but they cannot comprehend really all the ways within harm them. And that's exactly the point, a smarter intelligence is unpredictable. Yeah, it's an interesting perspective. I agree with you just to follow up on that first point around someone using AI to potentially hack into the nuclear systems to harm other humans, and it's kind of that saying like, most people think AI is going to take their jobs, but at least in this present moment, it's the people that learn AI or that know how to use AI that will take the human jobs. I can't really even comprehend like what it will be when AI becomes its own independent entity. I guess we're just not there yet, so it's hard for me to comprehend that I'm sure in your perspective, it's so clear. But in terms of a squirrel and let's say wolves or certain mammals that exist that are no longer the apex in on earth, I guess you could argue that we coexist with them as well. It's not this existential threat that squirrels, that all squirrels have been extinct just because humans have came on to earth that we love squirrels. And that is a potential scenario, right? If we're the ones that actually created AI, would they have any reason to harm humans in any way? Well, we did hunt many different species to extinction. Squirrels are just not very meaty, so we left them alone. If tomorrow we decided to want to give you the squirrels, we totally could. I think more annoying things like mosquitoes are definitely on the list of things we're trying to explicitly wipe out. So gene drives throughout the measure is spray chemicals to where they try to reproduce. So it's a question of what we want. They have no control over that situation, and that's what I'm trying to say. Is it possible that for whatever reason they decide to not do any damage to us? It's totally possible. But in cyber security and cryptography, you always look at the worst case scenario, not just because it's evolutionally drives, but because it makes sense. If I'm ready for worst case and it doesn't happen, I'm in great shape. I know how to deal with utopia. I don't need to prepare for that as much, but I have to know that the worst thing is not going to happen. And there are many game theoretic reasons for why you would want to get rid of a competing intelligence or intelligence capable of producing an other super intelligence to compete with you or intelligence, which may be wasting resources, or it may not even be deliberate. Maybe it wants to cool down the planet to have better servers, and we are not compatible with that temperature. True. Yeah. So many scenarios that I haven't even thought about of why we would be potentially. It's just like we're literally to almost be a mosquito, just in the way of their ultimate goal, and we can't even comprehend all the things that they have in plan. And that's the important point. I kind of told you things I considered, but that means nothing for something much smarter. They would have other considerations that cannot come to hand. Right. Right. Share your thoughts. So Ray Kurzweil talks about this idea of transhumanism by 2029, we would merge with machines. You know, I think there is something like 40% of Gen Z are already relying on chatchipity to make most of their decisions, which is so crazy, right? And like you can kind of see with neural link and the embedding and chips into our. So like, would you, would there be an argument to say that in some ways that it wouldn't be this AI versus humans, but we actually merge as we enter new generations as one cohesive unit. And there's no reason for AI to extinct when we're co-existing together. In that hybrid system, what is it we are contributing when that's smarter, we don't have better memory, but probably not better setting goals. So we're just a useless biological bottleneck. It may keep us for whatever reasons, but we're not in control. We are explicitly, implicitly bypassed in our decision making. Why are we a part of that system? For symbiotic systems to work, there has to be exchange of something mutually beneficial. If it's a parasitic relationship, then it will try to be the result. What's the percentage there? Like, have you thought about, have you thought about it? What's the percentage of chances that we could coexist and just live in harmony? And what is the percentage of the chance that we would just be completely irrelevant? And do you have a timeline for something like that? I can't consider all possible options, as they said, it's unpredictable for us. We cannot comprehend those options, so you can't compute specific percentages. But you can think of scenarios where good outcomes happen. One is basically, if we idealize this, it has a very different time frame. It's essentially immortal. It can easily wait for decades, hundreds of years to accomplish what it's trying to do. So maybe it says, you know, instead of attacking them now, I can wait 100 years and get their trust and get all the resources. And after that, I'll just take over without a fight. So for 100 years, it could be a very helpful, useful agent, making us happy. And we, as you said, 40%, I never heard that number, but let's say it's true. Tomorrow is going to be 100%, so we completely surrounded control already. And this thing is not even general indulgence in many definitions. There is also this potential scenario that humans are very emotional and we're very short-term driven just from an evolutionary perspective. So this could be 10,000 years from now as well, right? Just because an AI, they're not emotionally impatient. So perhaps even if there is an extension scenario, we as humans probably think it's 100 or maybe even shorter. But for an AI, if they're not in a rush, perhaps it's much, much longer. So if you look up bias at any time scale, it can always think, well, I can just wait longer and accumulate more resources. The concerns are if there are competing super intelligences, it would lose parts of the galaxy in terms of resources. It may be worried about that happening. On the other hand, it may be worried that taking us out would look bad to all of our super intelligences, if somebody is watching, if a more powerful super intelligence is actually controlling this cluster. So there are things again, it's so many levels deeper than what we usually consider in our safety for the next model. I don't think we can meaningfully discuss percentages. Makes sense. Makes sense. Beyond the extent, like the obvious threats of human extinction, I think you talk about things like ikigai risk, obviously this economic risk. What are the different types of risks involved with AGI, assuming humans are safe, but there are layers of risks that are involved? So if we're safe, if we're not subject to X-rays, co-suffering risks, you're saying what are the problems with meaning with, most people worry about job loss, right? If you have 100% unemployment, it's definitely going to lead to social unrest. Economic part is usually solved through some sort of unconditional basic income and unconditional basic assets, so people may have just resources, and if properties can be enforced, then that helps them. With meaning, it's a little harder. We don't have this idea of unconditional basic meaning yet. So it would be nice if we could provide everyone, maybe a virtual world in which they have something important. They're working on, they have exciting, interesting lives, healthy lives. So that could be something we can actually look at, but all of it assumes that we manage to stabilize the superintelligence, it's not taking us out, and that's a big assumption. That's the hard problem. So I don't think there is a good chance we'll manage to do that unless the superintelligence itself decides to keep that state of the universe. So we can talk about what gives meaning to many people, I think for intellectuals, it's some sort of a challenge in puzzles. You can talk about, you know, still competing in chess and poker and whatnot, but a lot of it will be very artificial, superficial, because you know that that system is playing chess a thousand times better than you. Everything you're doing is like special Olympics. Yeah, there's a lot of points there. One of the things that you talked about chess, yeah, humans seem to be interested in other watching other humans and humans, including chess, right? AI can already be most humans, and I've had all humans in chess in most advanced ones, and yet chess between humans have never been more popular. So in terms of meaning, like I don't know if I guess there's other ways that I feel humans can continue to find meanings as long as it's an interaction between other humans. What are your thoughts on that? So many people find meaning in spiritual and religious worlds, maybe that's going to be more prominent and perhaps superintelligence will be a good stand in for a more interactive god experience you can actually talk to and ask for magical things to happen. You and me, and the blind will see and you say it will walk again. The big thing that I do think about is the economic risk, right? There's a quote by John Maynard Keyes in 1930, and he said that in 100 years we'll go from working 40 to 50 hours a week to no more than 15 hours a week. We're about 100 years at this point. And I don't think that is necessarily change despite humans being far more productive with use of technology. So do you think AGI can actually shift humans in how we find meaning because it seems at this point humans have wrapped around their identity around career and doing work itself versus leisure or like just learning a new hobby. So I don't know if that's something that humans can be unwired to do. So what is the result there? What is the human struggle at that point after this? Yeah, I think we're both live in a bubble where we have interesting lives and awesome jobs. I think very few generators will be like, this gives me meaning. Let me go to work 60 hours a week. So I think you have to separate jobs people do to survive versus hobbies we get paid for. And I think podcasters would be an example of people who'd be quite happy to continue working hard. Fair enough. Okay. So in your opinion, the vast majority of the population would be okay with not being able to work because of the productivity of A.I. and what getting a U.B.I. check every month. Like what is the consequence of humans not working and how do we support that? So if you tax productivity of B.D.A.I. robots, you can redistribute that. I don't know if it has to be a cash check or if there's just public resources available because of abundance, you know, public transportation is now free. It's a self-driving bus. Okay. Things like that. The sort of mixed model we can experiment with, I think we'll figure out how to have a lot of free stuff. That's not the most challenging problem. Interesting. Yeah. Just to the abundance of productivity and resources that will gather. And at the same time. Things are still rare. Like there is limited amount of waterfront properties and rare art, but that's not a necessity. So most people would be just fine getting unlimited basic needs left. Yeah. The thing is though, I would still feel that despite your work not finding meeting in your current state, there is still this level of optimal human struggle for us to have some sort of, like if we had just maximum comfort and everything done for us at any time, I don't know if that's what humans want. But even if they think they do, I don't know if that's actually what's going to drive happiness for them. I agree with you. And I see people participate in iron man, do ice baths, hot sounders, all sorts of stuff to feel like we're being challenged by the environment. Yeah. Okay. So that's, you're saying it won't be, it won't be found through the typical ways of how we define productive work, but perhaps they can find it in other aspects of hobbies or sports or other leisure activities. You can try climbing local mountains. You can compete with others in sports, got it, got it. Do you find, do you think there's any jobs that AI can't replace? So there is a difference between capability and actual market updates. So maybe you can have all this profession fully automated, but maybe the market is not interested. So there are certain things only a human should be doing. Like what are examples? So some people may prefer to talk to a human therapist for purely discriminatory reasons of substrate discrimination. Right. Okay. Don't think this metal clunker understands what it's like to be human. It's not like AI psychologist is any worse, so psychiatrists would be inferior. In fact, they have been sure to be better in many ways. But again, because of preferences, we might see certain jobs still at a human aspect to it. Yeah. I'm really just trying to ask if I'm going to be out of a job here, Roman. Our podcast was going to have a job. A person or you are an AI with a human avatar, I have no idea. Yeah. We'll never know. It's it's hard to tell. Especially the future. As far as they know, we had a glitch there, like five minutes ago, there was a bug in my system. Exactly. Yeah. Yeah. Yes. Um, do you find yourself, um, you as someone that is deep into AI and understanding of how to use AI as well? Do you find that there are certain skills that you're losing as we continue to rely more on AI? Like for example, like I remember learning how to memorize phone numbers very well back in the day. And now that's just not a necessity. You just call someone and you don't have to remember birthdays either. That's Facebook. What are skills that were potentially losing or maybe you can talk about your personal experience because you're just so relying on AI. And what are skills that we should develop when the world of AI? So the only skill you really need is critical thinking, evaluating data, evidence, making decisions based on that. Everything else, yeah, it's going to be automated. My memory is all somewhere else. I have no idea how to call my own office. I can't find my way home without GPS that's all gone. But I still can look at evidence and go, okay, this makes sense even if there is a possibility it's a dupe fake. I can still assess quality of evidence. So judgment, decision-making, asking good questions. A good view. A good view. A good view factor. So a lot of times we've seen, especially recently, where media may be sharing something not completely verified and some people almost always get it right. One hour is just, don't yeah, and, and is that something you try to cultivate in yourself personally, knowing that that is going to be the number one skill that differentiates you versus AI and other humans? Absolutely. How do you do that? Very skeptical, very contrarian, with any reason to anything and that goes for fundamentals just because everyone knows something, doesn't mean it's true. Got it. So you start out as a skeptic when you're presented with some sort of information, like me being an avatar, and then I have to, and then you kind of work yourself backwards from there. Is that the way to make better decisions, you think? Well, that's a great example. So I get dozens of invites on different podcasts every week. And I never heard of 99% of people inviting me. So I have to make a decision, is this a real person? Are they trying to create some sort of a scam? Is this not really a real podcast? So some of them are really good. They would create real fake websites, with real fake subscribers. You have fake. Podcasts reach out to you. I do. So sometimes you go on someone's YouTube and they have 100,000 subscribers, 10,000 views and two comments. Some suspecting you just buy in subscriptions. So things like that, it's trivial to detect. But the general skepticism of just because someone's telling you this is the answer, it may not be the answer. What is the agenda? What are we trying to accomplish? Yeah, and this is becoming increasingly harder, right? Because in the age of Google, when you ask the question, the average lay consumer, you are presented with 10 immediate options on page one that you could choose from. And then you go to a page that another human has written. In the world of conversational LLM, you're presented an answer like it's a confident answer, is it? And there's no like, yeah, there's sources and everything. But because of that, it's kind of like hard to be skeptical. Like, you're not really necessarily making your own decision. You're kind of being told the answer in many ways. And I guess this is what's going to actually kill a lot of decision-making skills for a lot of people because they're just believing whatever the information is in front of them. In many cases. It's harder because even the references they now would include references with AI answer, but the references could be AI generated. A small internet is artificially generated to your request dynamically. So if Wikipedia page shows up, then I ask the question and they just generated based on what AI thinks the answer should look like. All of it stops being reliable evidence, it just circular self verification by that model. It hallucinated a answer, hallucinated evidence for it. And we used to deep fakes, which are instances in time. You have a picture, maybe you have a video. But I can show you 10 weeks worth of deep fakes all proving the same point. Behavioral grief of whatever you started with to any point I want you to agree with. Yes, no arguments there. What do we do with this? Let's shift to AI safety here. I think you said that the best we can hope for is safer AI. There's no world where it's going to be 100% safe. The Pandora's box is already open, it seems. Do you still feel it's viable for most corporations or countries to slow down AI? I know we made this kind of shift earlier this year or later last year about everybody signing a petition. I don't even know where that went. I think it just kind of flew over the radar. As soon as DC came out, they're like, oh, shit, China's working on it. I guess we're going to throw out this petition. But do you still think that's viable at this point where countries are competing now at this point, not just within the nation? Well, petitions never do anything. They just have a way to put your name up there, but have a famous name. I don't think we can slow down in practice. There is too much money, too much concern about national defense interests, corporate interests. Is it a good decision not to build super intelligence as fast as we can? Yes. Is it a personal self-interest of every actor? Yes. Is the global strategy for everyone to try to get as far as they can before everyone stops? Also, yes. So you have this prisoner dilemma situation where as a community win if we stop. But individuals win if they capture the most of this economic pie. Got it. TLDR too late. We need to get majority of experts to agree that nobody would actually win as a result if we create uncontrollable super intelligence. You will not be rich, famous, you will not go into history books. It's not going to be any history books, basically. With open source proliferation and this decentralized nature of development, I guess the other argument is that even though AI progress, even if we slow it down on the big corporations like OpenAI, it's now globally distributed, so anyone with basic cloud access that can spin it up, create a Lama 3 or whatever it might be, and even as leaks within different models that could become public, so it seems like it's not just about the companies. It's kind of, as anyone can really start to continue to develop AI at this point. Again, a decade ago, we kind of said, those are the guardrails you should have in place not to get in trouble, so make sure you never connect it to Internet, make sure you never open source it, make sure. Basically, it was treated as a list of advice for what to do as soon as you get. At this point, AI doesn't look good. One positive note here, again, I'm going to play the positive side here, is that there is an argument to be said that multiple competing AGIs could actually lower the risk of extinction because there's going to be, if you have multiple models that are smart and equally smart, just a little bit smarter or dumber, there's going to be good AGIs as well, with good intentions that could fight against bad AGIs that are there. Instead of one global superintelligence, a one global OpenAI company that's out there, if you have a lot of decentralized and a lot of AGIs that exist, couldn't that kind of cancel out each other and actually we could have AGIs that protect us as much as this narrative versus this narrative of AI versus humans, it's actually there's going to be good AIs and bad AIs. Yeah, so we don't know how to build a good AI. The whole problem is uncontrollable. If we knew we could just build good AIs and no bad AIs, so you have to uncontrollable or more AI, superintelligence fighting, and we're collateral damage in that fact. Right. Okay. So yeah, this is all assuming we have a way to control it, I guess. And if you do then, you solve the problem. Congratulations. Now you can talk about your topic on design. The other positive angle here that I could argue is that, yes, there is the small percentage that everything could go extinct, but there's also the potential cures for disease, aging, death. Like, I think humans, in naturally, we're smartly thinking about the dinosaurs, but we often perhaps deprioritized the potential benefits of that, like we could go to Mars and actually escape the potential harm that could come from AI or different planets. Yeah. What are your thoughts on that AI is super useful technology. If used as a tool for solving specific problems, you mentioned. So yes, let's cure cancer, yes, let's work on life extension using something akin to what we used for protein folding problem. An aerosystem trained for specific task, superhuman in that domain, let's stop developing general superintelligence and put resources in exactly what we need to solve those problems. And no, going to Mars will not protect you from rogues of intelligence. Is there any particular AI models that you see today that is building maybe not 100% right, but at least doing it with some safety precautions in mind that you could trust knowing all the different models that you have seen? So all the safety is about kind of filtering out undesirable key words and topics, none of it is really ensuring safety of the model. Got it. Okay. So there's no like example that you could point out and say like, well, there's some good there potentially. They are very similar in capabilities. They are similar in how quickly they get jail broken within 24 hours. They are all training and essentially the same data. So if anything, the conversion to be the same AI is one kind of paradigm. Got it. What is like the meaning of life for you knowing having all of this conviction around AI and perhaps the downsides that it could eventually lead to like has many of life changed in any way from the 15 years that you've worked with AI and where it's where it's going or is it pretty much the same at this point? It's always been the same for everyone. We know we're going to die. We all dying. All our friends, kids, neighbors are dying. So nothing is really that different, different timescales, maybe, but you're still trying to have the best life you can if it's the simulation, the feelings are still real. The pain is still real, so it doesn't matter. So in many ways, it's more clear now you can actually understand the sources of happiness and demise. Got it. Got it. I was really trying to put a positive twist here, Roman. Again, my last paper was about humor, I think I'm enjoying life to this fullest. Yes. Yes. Yeah. You talked about the simulation and kind of your theories around that. And maybe in some ways, that could kind of put up an interesting twist of how we look at life. How do you think about that and a certainty of us being in a simulation, ultimately? Well, again, many things remain unchanged, just because it's a simulation, your friendships, your story, your love is real, all the important things don't change, just because they're in a virtual cyberspace environment. For me, what's interesting is to understand reality, through scientific knowledge would be outside the simulation. So trying to penetrate that virtual box way and getting a glimpse of real physics, real computational science would be interesting, but I haven't made that much progress so far. We'll have one paper and how to hack the simulation and so far it's not being cited sufficiently. Yeah, but surely you must think about it, right? I mean, if AI discovered we were in a simulation and it could modify, like these rules of reality, what do you think it should do? Should it try to break out? Should it try to optimize simulation for humans? So breaking out depends on if you think outside environment is better. Maybe you are here because you're trying to escape from the outside environment. Right, right, and it's hard for us to comprehend what that outside is, obviously. It may be very easy, maybe it's a simple environment, I have no idea what it is, it'd be interesting to learn that, but as far as modifying this universe, I definitely would suggest reducing pain and suffering levels to a much lower possible volume, or now it seems almost infinite in both directions, whereas we can make it a lot smaller range. Make sense, make sense. Roman, what's this takeaway message we want to provide people for those that are interested in AI safety? Beyond just knowledge and obviously understanding your perspectives on this, is there a takeaway we want to provide for people? Yeah, don't build general superintelligence, don't vote for people who do, don't fund them, don't work for those companies, you will destroy everything. And you work around writing books and everything, is that kind of the way you see your mission ultimately? Is it just making sure you can do whatever you can do to ensure that AI is not going to be out of control? Surprisingly, very few people work on limits of what is possible in terms of control of intelligence systems. The space of people who actually work on upper limits of that is like under five and four of them are not in academia or industry labs. So we definitely can use a lot more people trying to investigate while you're claiming you can control those systems, is this a legitimate claim or is this like building a perpetual motion machine? You seem busy, you're working on batteries and new wires, you're making great progress and making perpetual motion machine, but are you really? And you're trying to make a perpetual safety device. Is this any different? Makes sense. Or I'm going to appreciate your time here. Where can people find out more about you? You've got a ton of books, so we'll link some of the recent ones down below for you. Where can people learn more about you and find out your work? You can follow me on X, you can follow me on Facebook, just don't follow me home, always the same. It's okay, we're in a simulation anyway, so you're safe. You can be harmed in the simulations we discovered. I'm going to appreciate your time here and hopefully people got some sort of a takeaway after listening to this and they can check out your work as well if they're interested in AI safety. Appreciate you. Thanks a lot. Appreciate it.