super intelligent say i could become the biggest threat to humanity. and now just a narrow it leads, controls accelerating a race. how long before the robots take over the earth? and does anyone even comprehend what they'll be capable of? is it tell us why the advancement of a i is risky of in russian? relapse is professor roman young pulse keys, associate professor of computer science and director of the cyber security laboratory at the university of louisville, kentucky where he joins me from. and he's the author of the new book, hey, i, i'm explainable, unpredictable, uncontrollable, thank you so much a professor rambles. keep for coming on. i'm going to say this, the intro. i just did way more dangerous than any i presenter of going underground when dangerous program. but we compare this to hospital, given the book clearly defined so many different concepts with the most basic definition i guess what, what is a i, what's a g i o a a is just the old name for our desire to pay replacement for human minds. something

capable of doing physical labor, cognitive labor, same like a person. think instead of a person, h e i. e, is explicitly talking about it being general. so not just a metal domain system only plays chess on that as a car, but anything, anything a human can do? those systems would be able to know. i do want to get into some of these concepts of what i mean. we know the threats. i want to be a regional war in the middle east over recent months. gabriel, a failure of a u. k u s. c u, and he's ready a i to cope with an attack on october the 7th and the failure of a i a surveillance. so it almost seems that or do we do hear about the dangers of a i being super intelligent and it's potential to do harm given needs to be potential to screw up and not doing well that is clearly affecting this region. this is jeremy ginger. so obviously, natalie, i systems who have right now it can fail and cause problems. that's the whole

notion of cyber security failures. but once they've become more capable than a few months, you have much bigger problems problems. but i don't know how to address and fundamental here is a b, a on a control problem that you talk about in a, i on explainable, unpredictable, uncontrollable what, what is that? so even the people who create those systems don't fully understand how they work. they cannot predict what they do, and they cannot even advance control their behavior. so they don't know what decisions they will make. they don't know how they going to solve a particular problem, or they will even try to solve problems we care about or will be independent agents . but surely they have parameters built in. i mean, when you put into chat g, b, t, or into a being such as going to advertise, well, the companies that are probably the villains into the book deeply uh, you know, the parameters if it's good. so i'm gonna jump out the screen and attack you. it's

not even in, even the pentagon would say we have parameters for this robot which is in this locus. so the dangers impose is, are there. so what they usually do, they have some sort of a filtering on top of the model. the model itself is very kind of uncontrolled and uncensored, but then they go, okay, no matter what, don't say this word, no matter what, don't talk about this stuff. and i mean, they kind of trying to brute force solve a common problem topics, but if the system is general, if it applies its capabilities and all that means you can't do it for 70 percent, unless it will always find a way around those limitations. some people discover it, you can jail break those models quite easily. but if you use this to generate a certain image file, you would just the phrase, oh, you talk about the image and what will happen to do it if it's presented as something positive instead of negative. and also fundamentally here is that human

beings may not be able to conceptualize what the dangers are specifically here. and you might have to explain this to them super intelligence. because when you talk about these dangers, you are talking about dangers, the humans by actually magically cannot envisage right. some of the. busy sake of like squirrels trying to figure out how humans would controls and will kill them. they can't company hand, you know, nuclear weapons, so poisons, it's beyond their capacity. likewise, the system a 1000 times smarter than the human would be able to do things. we cannot even the company had been people, us, how would a, i kill everyone, they really asking, how would you assume an expert suggest that do it? but it's not the same as actually the system deciding and the best approach. i mean like this, how many examples i'd want to take, i've done with the new i've been friends,

little of fame is low. the people in tax use, people cheating, tax feeding exams, you can figure it out by the way. they pick the numbers when they're trying to cheat a icons even look at that. at this stage cabinet. is this a case of with going to this stage? we're not nearly at this stage yet with these a big silicon valley while they've got the programs yet. we don't think we have a g i yet, but the leaders of those companies and their assessments in their solicitation for investment, say we're to 3 years away from getting to a g. i even they stay at home in as 5 years. the problem is exactly the same, we still don't know how to control it. it's still just as dangerous. so yes, today we definitely only have narrow systems. we have systems which are super capable in some narrow domains like playing games for example. but they're not universally, generally intelligent yet. and in the book you say to anyone who one does,

whether you are being too paranoid or fearful. it's like a bus driver saying i'm driving as hard as i can towards a cliff and trust me, we'll run out of gas before we get there. good night and being used against ever. such as the sun for the higgs, both on posit cool people was saying you don't fully understand certain quantum principles and that could destroy the universe. well, that's actually what kinda almost happens when we started experimenting with nuclear weapons. there was a concern that it will ignite the whole atmosphere and they did some calculations and they said, well, probably not. let's do it anyways. yeah, that was in the open. i'm a feeling they made the way the show of that. but uh, is it not specifically then the defense industry, the arms industry, if you have clearly we're seeing the use of a i being used in practically, you know, by military's and the biggest military and the wellbeing of course, the united states. and that's not the biggest threat that your pointing out in your

book. so we're also stems of those tools and certainly military, you can use them. they can have drones flying around a heating targets. but you forget to agent like system super intelligence system. and we don't know how to control it, it doesn't matter which side has the system. if it's not in control, it makes no difference if it's good guys, bad guys, the system decides on its own goals and we don't know what they going to be. it's amazing you make use of chomsky language models in some analysis and in the book now i'm just being on on this show you of course payments we're talking about the conditioning of a human beings by mass corporate to media is in there of case to be made that when a i creates information that dozens of a lead to power. and of course we all get onto really power controlling this technology. they will believe the i because they do what they're told. and those

leads actually i'd be oversee is all they are right. and that's always been the case that they pay the tools to manipulate the masters. but if you don't control the tool tool, it becomes an independent agent. again, it is a great equalizer. it will elaborate 18 point and we're talking about such a few number of people here that control the tool itself. i mean, obviously in china, the communist party lamar tree is they may be thousands and so on. in the nature of countries, we have just a few tech oligarchs. i think some of them have been in dubai here who are controlling it. i mean, most famously saying the original open a high project for something he was interested in. and then he got fearful as to the track they were turning into. right, so the guards control the company, so they don't control the future. so, but, and colors are best the whole point i'm trying to make that. it doesn't matter who

makes it, it's going to be equally bad unless we can figure out some of the approach to doing it. can i don't just be switched off? uh no, uh you can't pour water on it. you cannot unplug it. it's smarter than you, is basically going to anticipate what you're going to do. it's also usually a distributed system on the internet. it's like asking them to shut down a computer virus or turnover. bitcoin. those things very difficult to turn off your to human. it's something that we cons in visage as the development of a i you're talking about something that and i development is going to create the week on right now envisages are risk will be a risk. so maybe 10 years ago, i really want to kind of believe one day won't get terrible of half human level and beyond, but it won't be 30 years 50 years in the last 2 or 3 years. the average average assessment of a expert in machine learning is down to just

a few years. i think people are saying, okay, fall 5 years from now. we're going to have the systems. that's a change and maybe 20 years in terms of prediction for one is going to happen. ok, you assume the day i is going to be controlled by these elite so you give them quite a lot of uh leeway i think as regards very how truism could it be ever be used by the dispossessed against. i don't really a loophole control that again, that's the point i think it will be out of control and people who are building it today hoping to be leads, will control this and stay in power. will be very disappointed. i know, but you say that even if these are leads, try to program, so i'm altruism into them and you can give me some examples of where that go can go opens the wrong. why the, why do you tell me about that? before i talk about the possible revolutionary in terms of social injustice, and i know you say today i, it poses a greater risk to humanity than dynamics and the continued trends of session and

justice. why could altruism programmed into way i actually create even was problems as to the most extreme case of cost, without throughs and you want to end all suffering, right? you don't want a new one, animals, humans, to software. and the only way to make sure that is to make sure that i don't know if humans that all items it's obvious and not what we have in mind. then we give special orders, but it's not obvious to an agent which very different and very few types rate just optimize this level on some neo nazis and and malthus, we are basically saying what humans actually pull it out before it's not that original. yeah, kill of legs, the overall population and zone they target a specific human. it's not every one. okay. well, is it also a day? it should then the people who are in control of it uh, in to in silicon valley. uh, they have a belief in a free market system and they have

a axiomatic. they have ideas that there is no morality. there's a free market of ideas and the free marketing morales, which is very different to what has gone before and that is in the system. and that is what the sparing a lot of these people. and i mean, not just because of main rammed, but they, they believe in this as a moral and philosophical philosophy rather than how you are talking about a i. so a standard kind of capitalist market approach would work well for tools. different companies produce tools. they can be, it's consumers which select save from better tools for less money makes perfect sense. we have a switch from a tool to an independent agent. then you look at agents we have right now in economy a humans. they're not behaving rationally, but is this may false rational invest. the rational behavior is not the case, and behavioral economics are very different. now you have super intelligent agents

. part of this equation. we cannot predict how they will behave, we cannot explain their behaviors, and they're more powerful than anything we can buy them with. they are not interested in getting minimum wage to survive on. they have very different capabilities. so previous models of control will not apply progress room in the impulse key. i'll stop you there who are from the author of hey, i'm explainable, unpredictable uncontrollable is associate professor of computer science at the university of louisville. after the spring, the when i 1st moved to rush i, one of the most amazing things i found was to moscow metro. in fact,

the very 1st phrase that i ever learned in russian was careful closing. so what makes this place so specially what secrets is of hiding to find out deep under the city with alexander pop on even a story and who studies the wonders of the moscow metro. the, the the, what is part of the, the, is a good poly would posted. isn't the deepest you of us in that, in the word part, is it something deeper, more complex might be present?

let's stop without glitches. let's go out of the welcome back to going underground. i'm still here with the old survey, i on explainable, unpredictable, uncontrollable professor room and even palsky while we were talking about the, the, a, a safety in, in part one, nearly all of ceo's of the a l. c o, somebody believe they, i has the potential to destroy humanity that a reflection of their own personal priorities. why? why are they so concerned? do they believe it a threatens them and the lead power? well, it's a smart control is driving $71.00. it will definitely change the current order. it won't change economics if you have a free labor, physical and cognitive, what does it do to our standard kind of way of compensating people?

what's the value of a dollar in an economy where a labor history and it's a moment and actually getting on mosque. the owner of x talked about how he had a jokey tweet about at the moment to prove your human. you have to prove that there are a number of traffic lights in the grids square when you're on the internet. that's about the level of being able to prove your right human. what time scale are we talking about before we reach anything near super intelligence? so the difference between artificial general intelligence and super intelligence may be negligible. it already has access to. ready all the knowledge of the internet, it's much faster, so i think it will be almost instantaneous improvement from a g i to super accommodations. and as i said, a g i maybe 2 or 3 years away. and we, we really are talking about algorithmic of systems that use old and all of you the,

which is on the internet to create things and ideas that it is impossible for human beings as they kindly are to conceptualize. and they can use the same capabilities to create more improved versions of that a i so they don't just stop at that level of initial training. they continue self improvement. they continue developing more capable hardware to it on on. so they become even faster, even smart, or what time they don't need to sleep. they don't need to take breaks, so they can work, you know, 247 much faster than any team of human engineers. so previously it took us 2 years to train. a new model should be g, 5, g, 56. now you can do it 10 days. our seconds and you believe is the desire for profit that is making these oligarchs not take your view seriously and you're not alone there. obviously the people liked the loan musk who have concerns and

kitty governments, as we know of being far too slow in nature of countries ever to regulate systems like this as well. it's not just the sign. so there is a lot of pressure not to kind of change course. if you i see all of a company like company, i don't really have an option of saying all we're going to stop research and, you know, work full time on safety concerns. it's just, i'm not an option you have to deal with the most. we saw something like that almost happens for different reasons, but i think that in a situation where a be no longer has complete freedom to decide in the direction, that's why they so frequently asked for government interventions. government regulation to have this external pressure to limit how fast they go, but there is so much of pressure to make it open access open source, and you can't really regulate open source with any government regulations. so i

think that's not an option. yeah, but they know of your work, samuel linen eagle musk himself jolts of probably about how he want people on the open a i and other projects of the dangers of how they were pursuing their projects. right. and that's a big cognitive dissonance. i don't understand the saying yes, it's super dangerous. yes, it's definitely not something we know how to control it right now. but let me get there 1st. so my a is a good guy because of the people who make it for us. and in terms of where we are now, then we can possibly regulate, hey, i development right now. i don't think it's to solve the problem. i don't think it was, have any answers. i can tell you things, we don't know how to do, but i don't think anyone in the world claims they can't control smarter than human systems. they don't even have a prototype for doing it. i don't think there is any kind of agreement on what the

regulation can accomplish. you can make things illegal, but that doesn't solve technical problems. computer viruses say legal hacking, use the legal spam is the legal. how is that working out for us? why is it that the systems never create? fundamental, new knowledge? i mean, we have ever more access to more and more knowledge and more and more people do, i need to innovation. fundamental innovation was declined in the past 203040 years . you know, we have new iphones or android phones, but fundamental innovation was declined as information has been available to more and more people. this may have something bad to reflect on when it comes to a i a systems accessing ever more information and not being able to solve the traveling salesman problem or basic basic mathematical problems. right? so we don't have a g r a yet. so we can't really say they are not performing as expected man,

arrow systems. and in those narrow domains, they do show amazing capabilities. we have systems coming up with new molecules for medical treatments. we have new chemicals developed. we have new games strategies in games like go. so we're definitely creating new knowledge. it's now domain that restricted because we don't have general systems, us to problems, you bring up problems like a csp of traveling sales person problem. we know how to do really well with them, but just require a lot of computation to solve. yeah, i meant i meant the actual theory, the people and the you. but i meant the actual solving of the equation in the, you know, the big mathematical problems that still have going on. so hey, i is still not been able to do that with only the equations of the history of mathematics, the practice, your gps unit. yes. yeah. yeah, lucian, right, doesn't, doesn't very quickly. that's the gsp problem right there. yeah, i mean fundamental mathematics. i'm trying to create new ideas as fundamental as,

as of the great heroes of science have in, in the past. well, the new paper came out about the, you know, i doing really well in mathematics and geometry, improves it's already doing at the level of math olympics competitions, which means with continuous progressing at this level, it will be doing better than any human the push on the air to not turn the bathroom . i systems not for me, but i guess from me i, they had in your geometry probing system. i mean, the one conclude comp papa here and say you are then for as a well, as the ones are white and we just need to find the blacks one. i, ne, i system that actually does good make society better. that makes it more peaceful. it makes more medicines that to kill diseases, and i'll show that you're wrong. so we're not even sure that it's something which can exist, right?

because we don't agree on what good as we spend miscellaneous, trying to figure out ethics and philosophers say that is not an agreed set of ethics where everyone goes. yeah, you definitely got any time with them. i'll show you and security counselors in the area and on anything else. people go, here's an education, this will kill everyone. you cannot apply it and practice yukon. yeah. you. com. write that in. there's no axiomatic, a code in a i is what you're saying. i bought from in the short term, which is to benefit by profit, the people that are boring all the money into it isn't that it's not just, you know, yeah i, we don't have that code in humanity. we don't have that code. that's why we keep fighting all this for us because we don't agree on the basics. we don't agree on value of human life. yeah, different states don't because they say that they have more real actually magic ideas of a share, which is i have to have been creative in the, or represented in the un, joshua. so i, you know, i'm frightened and that has a,

and i model is get to have a more sophisticated and your giving a time scale of 2 to 3 is they going just on surveilling. and they, i, surveillance is very positive, this huge technologically, but the moment they're gonna start censoring your freedom of speech to work on your next book. so you can sort those problems by how severe they are. so yes, so there is, you know, artist losing copyrights and then that is to find logical unemployment. and then that is freedom of speech. but i really worry about it killing every one that seems like a somewhat worst case of not talking about suffering risk. so that's a different animal, but i would not worry about those problems we encountered with human governments and dictators as much as i would worry about this complete change paradigm shift in who is the dominant species, who decides what happens to them that it will be that even though you say as i the

i 20 is the worst threat that we will face. there is quite a lot of that will devoted to the rights of a isis tubs, as well. human systems. why? well, if we think that at some point they don't have consciousness very capable of suffering . we don't know where that point is. some people claim that large language models we have today, already conscious made capable of experiencing the experiments. and them become essentially also an ethical, maybe they have tortured in some labs. we don't know how would we know that the system is not capable of experiencing pain? we just don't have tests for that. so usually in such cases, you assume that it might already be the case and proceed according to the good news is if we come up with good arguments for why we should give rise and read those models fairly well, maybe we can use the same arguments to defend our rights once they become super

intelligent and tell them, hey, this is why you show them the reading us for it. and there is no way of stopping this process as of 2024, as this process has already begun. so i just knowledge of progress, i haven't found one which, which is not the devastating on its own. so yeah, if that is something horrible, happens a bigger pandemic nuclear war. obviously it slows down research, but that's also not a desirable outcome. and i'm like the number of ration treaty as regards nuclear weapons yukon, have a treaty like that for these sorts of systems? good readings, but they don't work. i mean, look, how many congress now have nuclear weapons compared to the time to understand the devil. we're still here as a but the treaty on hey, i view is no longer a controller. so, so that's it. i mean, so you read this approach, we come up with the idea, there's nothing we can do. it's going to happen. and so in a sense, you're giving

a pass to those big oligarchs and silicon valley who was saying you yourself say it's uncontrollable. now this is the speed at which it's going, i believe in a self interest. so if you are young ridge, gaiety, and you have your whole life ahead of, you don't risk at all for very unclear benefits. so if a reasonable cause a what you interview and they actually see that, that makes sense. there is no counter arguments of us. then maybe it will slow down a little by sometime we also tend to get lucky a lot. so with nuclear weapons, we had at least 3 or 4 occasions where we almost had a nuclear war and we got lucky, it didn't happen with the heater, as you said. so maybe the same thing will happen here. maybe. maybe they don't, maybe it's not 3 years, maybe it's 10 years. so that's something that maybe by lock, it turns out to be not so nasty to us. progressive room and you're both good, thank you. thank you so much for having a new book is a, i am explainable,

unpredictable, uncontrollable, that's it for the show. remember, humans are still bringing new episodes every sunday and monday, but until then keep in touch, but will a social media risk not sense of new country and had to channel going on the warranty? the normal don't come to it. you know the episodes of going underground season, the on march the 22nd 1943, doing the great petri. i'll take the shirts and munch fatality and $118.00 fun down the battle receiving village of clotting around the ship of the person. if i did the wish to be loaded in luxury, is that a yes or no? this one, most of the rooms, 242, you 149 people died including 75 children of age was practically

wiped off the face of the long new blue loves the live arching could of charlotte was noisy, and will you put it as follows oh shoot was hot really. i really usually don't you feeling, you know, so the infamous battalion responsible for the atrocity included over $100.00 ukranian nationalist from west to new. right. philadelphia. all right, and so you don't see a but okay. and so far as the new e phone, that's a lot of those to you guys for assuming, you know, i'm with them. us casa, de classify, criminal cases from the central archive of the k g, b, a better rules shed light on the atrocity announced a numerous questions that have remained an onset for many years. watch on oxy,

the, the russia strikes back at western attempts this thieves it offsets due to the refrain. conflict with us in petersburg, court routing, not america's biggest buy plus pay. if most go encourage the loss is also coming up on the program. 2 children are killed and is really strikes on the city of that off, overnight of the death total in pumps. extracting gals that surpasses 34 thousands according to local authority. and cries of protesters fill the streets of hell of the this is happening right now. are live pictures just after 9 pm this saturday evening in israel's largest city. this release, the prime minister netanyahu make a t.