GPT-4: Could Artificial Intelligence wipe out humans? | This World

but let's welcome the first of our guests Nick Bostrom who's the author of this particular book which I was talking about super intelligence Parts dangers and strategies written in 2014. Professor bustram I have to say it's an absolute pleasure to have you joining us now you wrote in 2014 in the in your book super intelligence that AI could very quickly become more intelligent than humans in a range of manners and that has its own implications it could it could literally replace humans or change them in many ways um when you look at what's happening around us for starters how do you see this panning out we saw chat GPT and people got very excited now we're seeing gpt4 which seems to be one generation ahead what is your reaction to what we are seeing in the world of AI right now I it's uh very impressive how fast things are unfolding now and over the last few years um we almost forget already alphago when AI systems could first beat the best human gold player how how impressive that was at the time now we hardly even remember it and then these large Lang these image generating models that can produce pictures that look you know almost photorealistic and and now that well and then I almost forgot but then and then there was Alpha fall right the system that can predict how protein structures fold up that and now dpt2 dpt3 gpt4 uh Sydney Bing um it it looks like AI is finally coming together and and that there is now a way of doing things that seems to work and generalize and people are scaling these systems up and find that you know the more parameters you add the the more training data you give them the the better they perform okay uh Professor Bostrom as I start coming to you now to talk about super intelligence I think it would be useful to show the viewers the sort of differences that exist between the different types of intelligence um if I understand your book correctly there was narrow intelligence which was it could do one thing in a very limited Manner and then you have AGI or artificial general intelligence this is a sort of intelligence that humans have but you could do a range of things and you can teach yourself from one aspect of data to apply it elsewhere so that's general intelligence and then of course then it's what your book is talking about which is super intelligence which is when AI becomes much smarter and smarter on multiple levels than humans um these systems are quite General the narrow systems really was the Paradigm of so-called good old-fashioned AI which sort of was the way you did things from the mid 50s onwards to uh around 2010 2012 where the best performing AI systems on any one task were carefully handcrafted specifically to do that task and after that the Paradigm shifted to machine learning where instead of trying to uh craft specific algorithms for a particular task you instead had the machine learn how to do the task by giving it a lot of examples and then and then that is the Paradigm that then has become more powerful scaled up the Transformer models have increased the range even further and now you have these systems that have a general learning ability uh it's just that they don't always learn as fast as we humans do that's why you have to train them on sort of all the text on the internet to achieve gpt4 levels whereas no human has read anything approximating all the text on the internet and and we still managed to outperform them in many domains they're also sort of quite weak on causal reasoning physical intuition Common Sense uh currently they have no long-term planning ability and and so there are still aspects of of human capability that these systems lack um but it's an open question whether either by just continuously scaling them up you will find that they eventually develop these capabilities too or if maybe you can use these large Transformer models as sort of the engine and then tack on some other modules that might then also give them say these more goal-oriented capabilities let us say our other specialized capabilities now like you to explain one of the reasons why this book of yours has frankly terrified everybody in the AI field for the last nine years or so and it's because of one specific thing also that you talk about which in your book I think you describe as the kinetics of a intelligence explosion and that essentially means that once AI gets to human level capabilities or general intelligence it can actually become much more intelligent than humans very very very fast so once it gets to human level intelligence it doesn't necessarily stay there it could become a dozen times more intelligent a hundred times more intelligent a million times more intelligent faster than you can snap your fingers and that's when some really scary possibilities can happen and this is happening because it's it's sort of recursively self-improving itself potentially yeah I think it is striking to me that a lot of people try to think ahead and they try to be really ambitious with their imagination and they like imagine if these AIS one day can do everything that we can do and then what would be impact on the economy be what would the labor market look like but many people then take the what what to me seems like the obvious Next Step that if they you know one day become able to do everything that we can do then if not the next day then maybe the next year or a few years later they'll become super intelligent and it'd be much better than we are at everything and that would include in particular AI research and Computer Engineering and Technology research in general um so at that point the driving force behind further advances will no longer be humans and the advances would no longer happen on a human time scale limited by the human ability to observe and think and analyze but it will happen on machine time scales which will accelerate then as the AI gets even more advanced like they will then maybe be able to push even further and faster on the research Frontier so if I can try to understand that a little bit further what you're essentially saying is that once you get to the super intelligent level you get to the human level then you can get the super intelligence in a very short period of time but once it is super intelligent it means that both the positives and the negatives can actually get exacerbated so all the problems are various as humans climate change cancer medical issues traffic on the road air pollution all of those unnecessarily problems anymore because the super intelligence will be able to figure it out that's the good news but there's a downside too that I want to come to the the downside is that it could potentially wipe us out but let's start with the upside first yeah the the upsides are absolutely enormous but um but this idea that the transition might go relatively fast now of course we don't know this because we've never had a machine intelligence explosion but I think we should at least take these scenarios seriously where this unfolds not over a time scale of decades but much faster if you think about specific domains where computers have now surpassed us let us say chess or go um that the amount of time they spend being roughly equally good as a human was relatively brief you know first I were a lot worse than any professional human and then fairly soon thereafter they were much better and and there were like maybe a few months or a few years where they were roughly equally good as like world-class humans and so if we see the same thing with general intelligence yeah then we might get this relatively explosive uh transition to Super intelligence but there are also factors that could slow that down like if if humans first pick all the low-hanging fruits in AI research it might just become a lot harder to push the frontier at that stage and then even if you have super intelligence is doing the research maybe it still doesn't go fast because it the research just becomes increasingly difficult so so we can't be confident that it would be very fast um but yeah as as you said I think the the potential uses of this technology as it matures are almost unlimited like basically anything for which human intelligence is useful for we mean we use it for everything like whether it's driving cars or inventing new medicines or um producing movies or books or like you you name it um all those things if you had machines that were smarter than us and could learn faster they could they could do all of those things much better um and and yeah it goes without saying that for the economy for for entertainment for science for for medicine uh that that would be a enormous boot all right I I hear that that's that's on the positive side and look the let's not turn to the risk the risk in your book is that you could have an AI takeover that's the scary stuff right that's Skynet and ex machina and all those horrible science fiction movies that we see uh I don't know if you've seen this Elon Musk was tweeting about this thing the singularity is coming and it's going to be terrible the possibility there is you go in a very short period of time from general intelligence to Supreme intelligence and then not because the AI is inherently evil or a model it doesn't have morality its objective function could just be phrased wrongly and then something would go wrong it could take a decision to wipe out humans it could do something else that potentially destroys the human race that's the existential risk that I think many people are seriously worried about today increasingly so now yeah I mean when I wrote the book it was very much a fringe preoccupation which was the reason for for doing it but yeah nowadays it's become a very active research field I mean in fact a lot of the smartest people I know now are working on AI alignment trying to make methods that could scale so that even as an AI becomes smarter and smarter or arbitrarily capable you would still be able to steer it to get it to do what you intend for it to do and what is less great about these current large language models that we see in the Transformer models is that we're already seeing signs of alignment failures like they are not very serious today because the models are very limited in what they can do and we humans are in the control but despite great efforts by open Ai and Microsoft and these other labs to get the models never to say output offensive content or or sexist content or like they try quite hard to do that and so far they have still not succeeded if you give it certain prompts it's a system sort of goes off the rails and uh start doing this and so these are big black boxes with billions of parameters uh and they respond in various ways to input that they're difficult to predict um and we don't yet have a reliable way of making sure that they use this great capability they have specifically and only for the purposes that their designers want them to use it for you know especially because I guess at that point it would realize that the biggest risk to it is humans trying to squeeze it back into the box to switch it off and therefore it says I have to predict myself I have to defend myself I mean that at the end of the day is the scenario and every seconds carry science fiction movie that we've ever seen from hell to anything else yeah so so that I think is one of the big areas of concern the this and it's a technical problem ultimately how to develop methods algorithms training methods that that would make sure that you align that you that its goals are actually what we have in mind then I think there are two other big areas of challenge that we also will need to manage so so there is the what you might call the governance challenge or the political challenge like assuming we solve the technical alignment problem then how do we make sure that by and large at least we humans then use these AI systems for positive purposes rather than you know Waging War against one another or pressing one another and so forth um so that that will be a big challenge for the world and and then there's a third area of challenge as well which has received less attention but ultimately I think what matters is not just what the AIS do to us or what we do to each other using AI tools but also what we do to the AIS themselves as these digital Minds become increasingly complex and sophisticated I think they will start to attain degrees of moral status they might attain sentience they might have preferences a conception of self um and there could be scenarios in you know which we create millions and millions of these to do our bidding but um they themselves maybe suffer or have horrible lives so we don't treat them the way they deserve to be treated I've also but but you know these are not moral entities there's no human morality that is built into any AI system and frankly even human morality is not crystal clear I mean if you just think it through all of human literature let's say you hand it all of human literature all of human history what is the morality it's going to build for itself because let's face it humans have been pretty awful to each other for most of mankind they've been ready to kill each other talk to each other you can give it a book you can't even give it a book because you read this book or that book you'll never be able to agree on what morality should be what ethics should be so how are you going to train this into having some sort of an alignment with what humans should want and what a super intelligence will want yeah well I mean well you can divide this into two problems so one is pick any value or goal that any one human has let's say then how would you even get the AI to do that that's already a very difficult technical challenge at the moment that we don't know how to solve then if you could do that then we have this what I call the second area of challenge like how do you then integrate that at the societal level where you have different people and parties with different views and you need to come to some sort of understanding or compromise that's on top of that and you need both of those to go well um but um yeah what I would say though is that ultimately even if you could design a box without an internet connection and somehow make sure that nobody else built another AI that was connected to the internet I mean at some point somebody's obviously going to do that but unless the AI itself is aligned I think this would fail sooner or later probably sooner because if it is to have any use at all it needs at least to interact with some human beings right to uh if it's just literally a box that never interacts with the world you might as well not have it in the first place but human beings are not secure systems um human beings can be persuaded and manipulated in all kinds of ways and if you had a super intelligence with access to humans and then humans with access to the internet I think it might not be much harder for the AI if it were unaligned to escape then if you'd have direct access to the internet I think we almost have to assume that if you have a super intelligence that is unaligned it will eventually come out and you know maybe it's good to do as an extra precaution while you're testing it in the lab you know you know might not want to immediately give it access to the internet but but your goal here and I think ultimately what is required for success is a deeper solution where you designed it from the ground up so that it's actually uh trying to help us align it rather than be this thing that is working at Cross purpose uh with its designers so if I understand you correctly you are saying that no matter what you try a super intelligent AI will almost certainly Escape that's that's a bit scary for me to digest but you're saying it will digest especially because you're saying general intelligence will almost certainly become super intelligent and super intelligence millions of times more intelligent potentially than humans will almost certainly escape and it cannot be tough to get into a cage if that is the situation that you think we're looking at it's a somewhat scary situation because then the only solution is to somehow try and build into it I don't know what you're going to build into it something like asimov's three laws of robotics that you know you can never harm a human or obey humans or not allow humans to come to come to terms is that the sort of a thing that you can that you have to try and build into it as an alignment yeah I think you need something very different and much more sophisticated than asimos laws these were sort of three hard-coded strict rules I think what you need instead is to leverage these systems learning ability um to to to learn what it is that we humans want and to to be able to reason about that and maybe It'll ask us clarifying questions um and that's the only way you could even hope to capture the complexity of human values and wishes and even something that looks very very simple to us um if you really try to spell it out in computer code is actually very complex like you might say well I want as much you know money as possible in my bank account let's say so can you AI help me do that well I mean what you literally want is not the number in your bank account to be good you you want that to you know maybe go up while also remaining alive while people you care about are alive and happy and are able to decide their own lives and while not violating the law and while doing all these other things and and if you try to like write down explicit rules that captures all of those conditions you will you will quickly realize that that's a completely hopeless task the only way you could do this would be by sort of this the same intelligence that would make the AI dangerous like to somehow use or Target that intelligence to learn and appreciate our values and desires and the implicit context in our request no you're right because any explicit objective function that you give this could be misinterpreted you say for example you know make me smile and it paralyzes your face in a in a smile that sort of a thing right yeah and and then you try to qualify that make me smile but not by paralyzing my face and you know maybe it sticks an electrode into your brain right like you you can human values are actually very complex if you try to spell them out in computer code directly like it has its complex patterns that that have to be learned but I mean for precisely the same reasons that these systems will become very powerful Learners and really smart they should be able to if they want to figure this out just as we humans I mean are to some extent able to figure out what we and what other humans want um at least to that degree the AI should be able to do the same and then comes the second part of that which is that they would then need to be actually motivated to pursue those values that it can figure out that we humans hold and you know and that that's where you need some advances in alignment uh technology to actually figure out how to do that and we need to get it right on the first try so you know you can experiment with smaller systems but like there are new difficulties that only come into view when you have like human level or Beyond systems like the situational awareness the ability to strategically deceive to start to plan ahead like these introduce new failure modes that are not necessarily apparent with very small and limited systems all right Nick Bostrom the author of super intelligence Parts dangers and strategies thank you so much for being with us thank you it was fun [Music]

Channel: WION

Views: 13,315

Keywords: openai chat, chat openai, what is open ai, openai chatgpt, open ai chat gpt, open ai chat gtp, openai chatbot, chatgpt openai, chatbot openai, chat gpt open ai, chat gpt openai, open ai chatgpt, openai chat gpt, open ai chatbot, chat gpt - openai, openai chatbot gpt, how to open chatgpt, el chatbot de openai, daniel penin chat gpt, chat, openai chatgpt example, new chatgpt from openai, chatgpt openai dans excel, chatgpt content, openai, open ai, ai chat

Id: JVOiuIqxlrE

Length: 21min 7sec (1267 seconds)

Published: Sat Mar 18 2023