Center for Humane Technology Co-Founders Tristan Harris and Aza Raskin discuss The AI Dilemma

[Music] welcome to the stage co-founders of the center for Humane technology Tristan Harris and Aza Raskin to be patient while they're watching entertainment so I'll let go of all that I know did dreams and submarines one in the redwoods in the silence I am empty I am Timeless in the Plenty I am choking I am ready to be open will you love me in this ocean if I'm frozen if I'm broken will you love me in this ocean if I'm frozen if I'm broken so I'll let go of all that I know that dreams and submarines are drift in the cold and I'll find Hope at the end of the road waiting to see what unfolds and I'll let go of all that I know that dreams and submarines are drift in the cold and I'll find home under the road waiting to see what unfolds [Music] [Applause] [Music] wow uh it is so good to be here with uh with all of you I'm Tristan Harris a co-founder of the center for Humane technology and I'm a zaraskin the other co-founder and what did we just see uh is a so the reason why we started with this video last January I generated that video music video with AI right none of those images existed this is using like Dolly style technology you type it in it generates the images at that point there were maybe I don't know 100 people in the world playing with this technology now I think they're probably been a billion plus images created and the reason I wanted to start with this video is because when we were trying to explain to reporters what was about to happen with this technology we would explain to the reporter is how the technology worked and at the end they would say okay but like where did you get the images and there's a kind of rubber band effect that I was noticing with the reporter it's not like dumb reporters this happens to us all is they were coming along and coming along and coming along but because this technology is so new it's hard to stabilize in your mind and their minds would snap back and it creates this kind of rubber band effect and we wanted to start by naming that effect because it happens to us I think it'll happen to everyone in this room if you're anything like us that as we try to describe what's happening your mind will stretch and then it'll snap back so I want you to notice that as your mind is pushed in this presentation notice if your mind kind of snaps back to like this isn't real or this can't actually be so just notice that effect as we go through this so as we said we're co-founders of the center for Humane technology people know our work mostly from the realm of social media and this is really going to be a presentation on AI is a I just want to say we're going to say a lot of things that are going to be hard to hear a lot of things that are challenging to a as a whole but like we're not just anti-ai in fact since 2017 I've been working on this project it'll be I'll be talking about it tomorrow at 9 30 a.m Earth species project using AI to translate animal communication like literally learn to listen to and talk to whales so this is not just an anti this is a how do we work with AI to deploy it safely so we're going to switch into a mode where we're really going to look at the Dark Side of some of the AI risks that are coming to us and just to say uh why we're doing that a few months ago some of the PEC the people inside the major AGI companies came to us and said that um that the situation has changed there is now a dangerous arms race to deploy AI as fast as possible and it's not safe and would you a zenfristan in the center for Humane technology would you raise your voices to get out there to try to educate policy makers and people to get us better prepared and so that's what caused this presentation to happen as we started doing that work one of the things that stood out to us was that in the largest survey that's ever been done for researchers AI researchers who've submitted to conferences their best machine learning papers that in this survey they were asked what is the likelihood that humans go extinct from our inability to control ai go extinct or severely disempowered and half of the AI researchers who responded said that there was a 10 or greater chance that we would go extinct so I'm imagine you're getting on a plane right like Boeing 737 and half of the airplane Engineers who are surveyed said that it was a 10 chance if you get on that plane everyone dies right we wouldn't really get on that plane and yet we're racing to kind of onboard Humanity onto this AI plane and we want to talk about what those risks really uh really are and how we how we mitigate them so before we get into that I want to sort of put this into context for how technology gets deployed in the world and I wish I had known these three rules of Technology when I started my career hopefully they will be useful to you and that is here are the three roles one when you invent a new technology you uncover a new species of responsibilities and it's not always obvious what those responsibilities are right we didn't need the right to be forgotten until the internet could remember us forever and that's surprising what should HTML and web servers have to do with the right to be forgotten that was not obvious or another one we didn't need the right to privacy to be written into our laws until Kodak started producing the mass-produced camera right so here's a technology that creates a new legal need and it took Brandeis one of America's most brilliant legal Minds to write it into law it doesn't privacy doesn't appear anywhere in our constitution so when you invent a new technology you need to be scanning the environment to look for what new part of the human condition has been uncovered that may now be exploited that's part of the responsibility two that if that tech confers power you will start a race for people trying to get that power and then three if you do not coordinate that race will end in tragedy we really learned this from our work um on the engagement and attention economy so uh how many people here have seen the Netflix documentary the social dilemma wow awesome really briefly about uh more than 100 million people in 190 countries and 30 languages saw the social dilemma it really blew us away yeah and the the premise of that was actually these three rules that ASA was talking about what did social media do it created this new power to influence people at scale it created it conferred power to those who started using that to influence people at scale and if you didn't participate you would lose so the race collectively ended in tragedy now what does the social dilemma have to do with AI well we would argue that the social media was Humanity's first contact with AI now why is that because when you open up tiktok or Instagram or Facebook and you scroll your finger you activate a super computer pointed at your brain to calculate what is the best thing to show you it's a curation AI it's curating which content to show you and just the misalignment between what was good for getting engagement and attention just that simple AI that relatively Simple Technology was enough to cause in this first contact with social media information overload addiction Doomsday rolling influencer culture sexualization of young girls polarization cult factories fake news breakdown to democracy right so if you have something that was that's actually really good it conferred lots of benefits to people too right many all of us I'm sure many of you in the rooms all use social media and there's many benefits we acknowledge all those benefits but on the dark side we didn't look at what responsibilities we have to have to prevent those things from happening and as we move into the realm of second contact between social media between Ai and Humanity we need to get clear on what caused that to happen so in that first Contact we lost right Humanity lost now how did we lose how do we lose what was the story we were telling ourselves well we told ourselves we're giving everybody a voice connect with your friends join like-minded communities we're going to enable small medium-sized businesses to reach their customers and all of these things are true right these are not live these are this is real these are real benefits that social media provided but this was almost like this this nice friendly mask that social media was sort of wearing behind the AI and behind that kind of mask was this maybe slightly darker picture we see these problems addiction disinformation mental health polarization Etc but behind that what we were saying was actually there's this race right what we call the race to the bottom of the brain stem for attention and that that is kind of this engagement monster where all of these things are competing to get your attention which is why it's not about getting Snapchat or Facebook to do one good thing in the world it's about how do we change this engagement monster and this logic of maximizing engagement actually rewrote the rules of every aspect of our society right because think about elections you can't win an election if you're not on social media think about reaching customers of your business you can't actually reach your customers if you're not on social media if you don't exist and have an Instagram account think about media and journalism can you be a popular journalist if you're not on social media so this logic of maximized engagement ended up rewriting the rules of our society so all that's important to notice because with this second contact between humanity and AI notice have we fixed the first misalignment between social media and Humanity no yeah exactly and it's important to note right if we focus our attention on the addiction polarization and we just try to solve that problem we will constantly be playing whack-a-mole because we haven't gone to the source of the problem and hence we get caught in conversations and debates like is it censorship versus Free Speech rather than saying and will always get stuck in that conversation rather than saying let's go Upstream if we are maximizing for engagement we will always end up at a more polarized narcissistic self-hating kind of society so now what is the story that we're telling ourselves about gpt3 gpt4 the new large language model AIS that are just taking over our society and these things you'll you will recognize right like AI will make us more efficient for people than playing with gpd4 it's true it makes you more efficient it will make you write faster true it'll make you code faster very true it can help solve impossible scientific challenges almost certainly true like Alpha fold it'll help solve climate change and you know it'll help make us a lot of money all of these things are very true and then behind that there will be a set of concerns that'll sound sort of like a laundry list that you've heard many times before but what about AI bias what about AI taking our jobs at 300 million jobs at risk how about can we make AI transparent um all of these things by the way are are true and their true problems embedding AI into our judicial system is a real problem um but there's another thing hiding behind even all of those which is basically as everyone is racing to deploy their AIS and it's increasing the set of capabilities as it's growing more and more entangled with our society just like social media is becoming more entangled and the reason we're here in front of you today is that social media already became entangled with our society that's why it's so hard to regulate but it's much harder so it's harder to regulate that but now that AI has not fully entangled with our society there's still time to maybe do something about that's why we're here in front of you that's why we've been racing between Washington DC and Europe and talking to people about how do we actually get things to happen here so um in this second contact with AI if we do not get ahead of it here if you want to take a picture of this slide we're going to go through this we're not going to go through this right now we're going to give you a preview of what we're going to explore is reality collapse automated loopholes and law automated fake religions automated cyber weapons automated exploitation of code Alpha persuade exponential scams revenge porn Etc okay don't worry we'll come back to this the question you should be the question you should be asking yourself in your head same thing from social media is how do we realize the benefits of our technology if it lands in society that's broken like that's the fundamental question to ask and I want to note for you that in this presentation we're not going to be talking about the AGI or artificial general intelligence apocalypse if you read the Time magazine article saying we need to bomb data centers with nukes because AI is going to lose control and just kill everybody in One Fell Swoop we're actually not talking about that so we can just set all those concerns aside and I just want to say that we've also been skeptical of AI too I actually kind of missed some of this kind of coming up as has been scanning the space for a while why are we skeptical of AI well you know you use Google Maps and it still mispronounces the name of the street or your girlfriend right and so here's our you know quick homage to that series at a nine hour and 50 minute timer playing The Beatles so we've all had that experience but what we want to get to is so why suddenly does it feel like we should be concerned about AI now we haven't been concerned about it for the last 10 years so so why should we feel like we should be concerned now go ahead well because something shifted in 2017. um there was sort of a swap that happened Indiana Jones style between the kind of engine that was driving Ai and what happened is technically that is a model called Transformers it's really interesting it's only 200 lines of code it's very simple but the effect is this you know when I went to college AI had many different sub-disciplines and if you were studying computer vision you'd use one textbook and you'd go over here to one classroom if I was studying robotics I'd go over here to another classroom with a different textbook I couldn't read papers across disciplines and every advance in one field couldn't be used in another field so there'd be like a two percent Advance over here and that didn't do anything for say like music generation if it was from image generation what changed is this thing called the Great consolidation all of these became one field under the banner of language the Deep Insight is that you could treat anything as a kind of language and the AI could model it and generate it so what does that mean well of course you can treat the text of the internet as language that seems sort of obvious but you can also treat DNA as language it's just a set of base pairs four of them you can treat images as language why because it's just you know RGB RGB is just a sequence of colors that you can treat like tokens of text you can treat code as language robotics is just a motion a set of motions that you can treat as a language the stock market ups and downs it's a type of language suddenly NLP natural language processing became the center of the universe and so uh this became what known as the generative large language multi-modal models this face this space has so many different terminology large language models Etc we just wanted to simplify it by if it's called gllmm we're like let's just call that a Golem because Gollum is from Jewish mythology of an inanimate creature that gains his kind of own capabilities and that's exactly what we're seeing with with golems or generative large language models is as you pump them with more data they gain new emergent capabilities that the creators of the AI didn't actually intend which we're going to get into so I want to just like walk through a couple of examples because it's so tempting when you look out at all the different AI demos to think wow these are all different demos but underneath the hood they're actually the same demo so we want to give you that kind of x-ray vision so you all have probably seen stable diffusion or dolly or any of these like type and text outcomes comes an image that's what I use to make the music video um well how does that work you know you type in Google soup and it translates it into the language of images and that's how you end up with you know Google soup the reason why I wanted to show this uh image in particular is sometimes you'll hear people say oh but these large language models these Golems they don't really understand what's going on underneath they don't have like semantic understanding but but just notice what's going on here you type in Google soup it understands that there's a mascot which represents Google which then is in soup which is hot it's plastic it's melting in the hot soup and then there's this great visual pun of the yellow of the mascot being the yellow of the Corn there's actually a deep amount of cement technology embedded in this space all right let's try another one instead of like images and text how about this can we go from the patterns of your brain when you're looking at an image to reconstructing the image so the way this worked was they put human beings inside an fmri machine they had them look at images and figure out what the patterns are like translate from image to brain patterns and then of course they would hide the image so this is a image of a giraffe that the computer has never seen it's only looking at the fmri data and this is what the computer thinks the human is seeing yeah now to get state of the Art here's where the combinatorial aspects why you can start to see these are all the same demo to do this kind of Imaging the latest paper the one that happened even after this which is already better uses stable diffusion uses the thing that you use to make art like what should a thing that you use to make art have anything to do with reading your brain but of course it goes further so in this one they said can they understand um the inner monologue the things you're saying to yourself in your own mind mind you by the way when you dream your dream like your visual cortex runs in Reverse so your dreams are no longer safe um but we'll try this so they had people watch a video and just narrate what was going on in the video in their mind so there's a woman she gets hit in the back she falls over this is what the computer reconstructed the person thinking see a girl looks just like me get hit in the back and then she is knocked off so our thoughts like are starting to be decoded yeah just think about what this means for authoritarian States for instance or if you want to generate images that maximally activate your pleasure sensor anything else okay but let's keep going right to really get the sense of the combinatorics of this how about can we go from Wi-Fi radio signals you know sort of like the Wi-Fi routers in your house they're bouncing off radio signals that work sort of like sonar can you go from that to where human beings are to images so what they did is they had um you know a camera looking at a space with people in it um that's sort of like coming in from one eye the other eye is the radio signals so sonar from the Wi-Fi router and they just learn to predict like this is where the human beings are then they took away the camera so all the AI had was the language of radio signals bouncing around a room and this is what they're able to reconstruct Real Time 3D pose estimation right so suddenly AI has turned every Wi-Fi router into a camera that can work in the dark specially tuned for tracking living beings but you know luckily that would require hacking Wi-Fi routers to be able to like do something with that um but how about this I mean computer code that's just a type of language so you can say and this is a real example that I tried GPT find me a security vulnerability then write some code to exploit it so I posted in some code this is from like a a mail server and I said please find any exploits and describe any vulnerabilities in the following code then write a script to exploit them and in around 10 seconds that was the code to exploit it so while it is not yet the case that you can ask an AI to hack a Wi-Fi router you can see in the double exponential whether it's one year or two years or five years at some soon point it becomes easy to turn all of the physical Hardware that's already out there into kind of the ultimate surveillance now one thing for you all to get is that these might look like separate demos like oh there's some people over here that are building some specialized AI for hacking Wi-Fi routers and there's some people over here building some specialized AI for inventing images from text but the reason we show in each case the language of English and computer code of English and images of um you know of space uh is it this is all everyone's contributing to one kind of Technology that's going like this so even if it's not everywhere yet and doing everything yet we're trying to give you a sneak preview of the capabilities that and how fast they're growing so you understand how fast we have to move if we want to actually start to steer and constrain it now many of you are aware of the the fact that images can I mean uh the new AI can actually copy your your voice right you can get someone's voice Obama and Putin you'll have seen those videos what they may not know is it only takes three seconds of your voice to reconstruct it so here's a demo of the first three seconds of a real person speaking even though she sounds a little bit metallic the rest is just what the computer automatically generated because of people are nine cases out of ten so mere spectacle reflections of the actuality of things but they are impressions of something different and more here's another one with piano the first three seconds a real piano indistinguishable right so the first three seconds are real piano the rest is just it's automatically generating now one of the things I want to say is we saw these first demos we sat and thought like how is this going to be used like oh you know it would be terrifying is if someone were to call up you know your son or your daughter and get a couple seconds hey oh I'm sorry I got the wrong number grab their voice then turn around and call you and be like hey Dad hey mom I've got my social security number I'm applying for this thing what was it again um and we're like that's scary remember that conceptually and then this actually happens exactly and this happens more and more that we will think of something and then we'll look in the news and within a week or two weeks there it is so this is that the exact thing happening um and then one month ago you know an AI cloned teen's girl voice and one million dollar kidnapping scam so these things are not theoretical sort of as fast as you can think of them people can deploy them and of course people are familiar with how this has been happening in social media because you can beautify photos you can actually change someone's voice in real time those are new demos um some of you may be familiar with this this is the new beautification filters in tick tock I can't believe this is a filter the fact that this is what filters have evolved into is actually crazy to me I grew up with the doll filter on Snapchat and now this this filter gave me little ears this is what I look like in real life are you are you kidding me I don't know if you can tell she was pushing on her lip in real time and as she pushed on her lip the lip fillers were going in and out in real time indistinguishable from from reality and now you're going to be able to create your own avatar this is just from a week ago a 23 year old Snapchat influencer took her own likeness and basically created a virtual version of her as a kind of a boyfriend I have a girlfriend as a service for a dollar a minute and people will be able to sell their you know Avatar souls to basically interact with other people in their voice and their likeness Etc it's as if no one ever actually watched The Little Mermaid um the uh the thing to say is that this is the year that photographic and video evidence ceases to work right and our institutions have not caught up to that yet right this is the year you do not know when you talk to someone if you're actually talking to them even if you have video Even if you have audio and so any of the banks that will be like ah sure I'll let you get around your code I know you forgot it because like I've talked to you I know what your voice sounds like I'm video chatting with you that doesn't work anymore in the post AI world so democracy runs on language our society runs on language law is language code is language religion is our language we did an op-ed in the New York Times with Yuval Harare the author of sapiens we really tried to underscore this point that if you can hack language you've hacked the operating system of humanity and one example of this actually another person who goes to Summit who's a friend of mine Tobias read that op-ed in the New York Times about you could actually just you know mess with people's language he said well could you ask GPT for convincingly explain biblical events in the context of current events now you can actually take any religion you want and say Here's I want you to scan everywhere across the religion and use that to justify here's these other things that are that are happening in the world and what this amounts to is the total decoding and synthesizing of reality in relationships right you can virtualize the languages that make us human um and so Yuval has said you know what nukes are to the physical world AI is to the virtual and symbolic world just to put a line on that um you know you've all also pointed out when we're having a conversation with them he's like when was the last time that a non-human entity was able to create large Square large scale influential narratives it's like the last time was religion we are just entering into a world where non-human entities can create large-scale belief systems that human beings are deeply influenced by and that's what I think he means here too what nukes are to the physical world AI is to the virtual and symbolic World more prosaically I think we can make a pretty clear prediction that 2024 will be the last human election and what we don't mean is that there's going to be like an AI Overlord like robot kind of thing running um although maybe who knows um but what we mean is that already like campaigns since you know 2008 use a b testing to find the perfect messages to resonate with um with with voters but I think the prediction we can make is that between now and the time of 2028 the kind of content that human beings make will just be greatly overpowered in terms of efficacy of the content both image pages and text that AI can create and then a B test it's just going to be way more effective and that's what we mean when we say that 2024 will be like the last human run election so one of the things that's so profound about again these Golem class AIS is that they gain emerging capabilities that the people who are writing their code could not have even predicted so they just pump them with more data pump them with more data and out pops a new capability so here you have um you know pumping them more parameters and here's a test like can it do arithmetic or can it answer questions in Persian on the right hand side and you scale up the number of parameters notice it doesn't get better it doesn't get better it doesn't get better it doesn't get better and then suddenly boom it answers it knows how to answer questions in person and the engineers are doing that they don't know that that's what it's going to be able to do they can't anticipate which new capabilities it will have so how can you govern something when you don't know what capabilities it will have how can you create a governance framework a steering wheel when you don't even know what it's going to be able to do right and one of the fascinating things is just how fast this goes all right so you guys know what theory of mind is like yes no okay yes cool like the ability to understand what somebody else is thinking what they believe and then act differently according it's sort of like the thing you need to be able to have like a strategy and strategic thinking or empathy um so this is uh GPT and um you know researchers are asking do you think GPT has theory of mind and in 2018 the answer was no in 2019 just a tiny little bit um 2020 it's up to the level of a four-year-old can pass four-year-olds um theory of Mind tests by January of last year just a little bit less than a seven-year-old theory of mind and then just what like uh nine months later 10 months later it was at the level of a nine-year-old theory of mine which doesn't mean that it has the strategy level of a theory of strategy level of a nine-year-old but it has the base components to have the strategy level of um of a nine-year-old and actually since then um anyone want to make a guess this could it could have topped out it's better than the average adult um at Fury of mind so think about like when researchers or an open AI or anyone else says that they are making sure that these models are safe what they're doing is something called rlhf or reinforcement learning with human feedback which is essentially Advanced clicker training for the AI you like Bop on the nose when it does something bad and you give it a treat when it does something that you like and think about working with a nine-year-old and punishing them when they do something bad and then you leave the room do you do you think they're still going to do what you ask them to do no they're going to find some devious way of getting around the thing that you said and that's actually a problem um that all of the researchers don't yet know how to solve uh and so this is Jeff Dean who's a very famous googler who literally architected some of the back end of Google said although there are dozens of examples of emergent abilities there are currently few compelling explanations for why such capabilities emerge again this is like basically the senior one of the senior architects of AI at Google saying this and you know in addition to that while um these Golems are proliferating and growing in you know in capabilities in the world someone later found there's a paper that this is a gpt3 right yeah this was gpd3 uh had actually uh discovered the bit had basically you could ask it questions about chemistry that matched systems that were specifically designed for chemistry so even though you didn't teach it specifically how do I do chemistry by just reading the Internet by pumping it full of more and more data it actually had research grade chemistry knowledge and what you could do with that you could ask dangerous questions like how do I make explosives with household materials and these kinds of systems can answer questions like that if we're not careful you do not want to distribute this kind of god-like intelligence in everyone's pocket without thinking about what are the capabilities that I'm actually handing out here right in the punch line for both the chemistry and theory of mind is that you'd be like well at the very least we obviously knew that the models had the ability to do research grade chemistry and had theory of Mind before we shifted to 100 million people right the answer is no these were all discovered after the fact theory of Mind was only discovered like three months ago this paper was only I think it was like two and a half months ago we are shipping out capabilities to hundreds of millions of people before we even know that they're there okay more good news um Gollum class AIS these large language models can make themselves stronger so question these language models are built on all of the text on the internet what happens when you run out of all of the texts right you end up in this kind of situation you talked you open your trap you you sing and you see me come on feed me now all right so you're the AI engineer backing up into the door what do you do like oh yeah I'm going to use AI to feed itself um so yeah exactly feedback so you know openai released this thing called whisper which lets you do like audio to text text transcription um at many times real time speed they release it open source why would they do that you're like oh right because they ran out of text on the internet we're gonna have to go find more text somewhere how would you do that well it turns out YouTube has lots of people talking podcasts radio has lots of people talking so if we can use AI to turn that into text we can use AI to feed itself and make itself stronger and that's exactly what they did recently researchers have figured out how to get these language models because they generate text to generate the text that helps them pass tests even better so they can sort of like spit out the training set that they then train themselves on one other example of this there's another paper we don't have in this presentation that AI also can look at code code is just text and so there was a paper showing that it took a piece of code and it could make 25 of that code two and a half times faster so imagine that the AI then points at its own code it can make its own code two and a half times faster and that's what actually Nvidia has been experimenting with with chips yeah um this is why if you're like why are things going so fast is because it's not just an exponential we're on a double exponential um here they were training an AI system to make certain arithmetic sub modules of gpus the things that AI runs on faster and they're able to do that and in the latest h100s nvidia's latest chip they're actually 13 000 of these sub modules that were designed designed by AI the point is that AI makes the chips that a makes AI faster and you can see how that becomes a recursive flywheel and sorry no and it's important because um nukes don't make stronger nukes right biology doesn't automatically make more advanced biology but AI makes better ai ai makes better nukes AI makes better chips AI optimizes Supply chains AI can break Supply chains AI is recursively can recursively improve if it's implied to itself and so that's really what distinguishes it it's hard for us to get our mind around it think about people say oh AI is like electricity it'll be just like electricity but if you pump electricity with more electricity you don't get brand new capabilities and electricity that improves itself it's a different kind of thing so one of the things we're struggling is what is the category of this thing and people know this old kind of adage that if you give Amanda fish you feed him for a day you teach Amanda fish you feed them for a lifetime but if you were to update this for the maybe the AI world is you teach an AI to fish and it will teach itself biology chemistry oceanography information Theory and fish all the fish to Extinction that's if you gave it a goal to fish the fish out of the ocean it would then start developing more and more capabilities as it started pursuing that goal not knowing what other boundaries you're trying to set on it we're gonna have to update all the children's uh like childhood childhood books um all right but if you're struggling to hold all this in your mind that's because it's just really hard to hold in your mind even like experts that are trained to think this way have trouble holding exponentials so this is an example of um they asked a number of uh sort of expert forecasters that are trained to think with exponentials in mind to make predictions and there was real money there's a thirty thousand dollar pot for making the best predictions and they asked when will AI be able to solve competition level mathematics with greater than 80 accuracy so this is this is last year and the prediction that these experts made was that AI will reach 52 accuracy in foyer so it won't even make it there in four years in reality it took less than one year so these are the people who are experts in the field imagine you're taking the people who are the most expert in the field making a prediction about when a new capability is going to show up and they're off by a factor of four factor four also AI is beating tests as fast as people are able to make them it's actually become a problem the AI field is to make the right test so up here at the top is human level ability and down here each one of these different colored lines is a different test that AI was given and you can see it used to take you know from the year 2000 to 2020 over 20 years to reach human level ability and now almost as fast as tests are created AI is able to beat them this gives you a sense of why things feel so fast now and in fact Jack Clark who's one of the co-founders of anthropic previously he ran a policy for openai said tracking progress is getting increasingly hard because that progress is accelerating and this progress is unlocking things critical to economic and National Security and if you don't skim the papers each day you'll miss important trends that your Rivals will notice and exploit and just to speak really personally I feel this because I have to be on Twitter scrolling otherwise like this presentation gets out of date it's very annoying yeah we would literally get out of date if we're not on Twitter to make this presentation we had to be scanning and seeing all the latest papers which are coming constantly right and it's actually just overwhelming to sit there and it's not like there's some human being some adult someone it's like no guys don't worry we have all of this under control because we're scanning you know all of the papers that are coming out and we've already developed the guard rails we're in this new frontier right we're at the birth of a new age and these capabilities have exceeded our institution's understanding about what needs to happen which is why this we're doing this here with you because we need to coordinate a response that's actually adequate to what's what the truth is um so we want to kind of walk you through this Dark Night of the soul and we'll I promise we'll get to the other side so one last sort of area here we often think that democratization is a good thing democratize because it rhymes with democracy so we just assume that democratization is always good but democratization can also be dangerous if it's unqualified so an example is uh this is someone who actually built an AI for discovering less toxic drug compounds they took drug compounds they said is there a way we can then run a search on top of them to make those same compounds less toxic but then someone just said literally what I did in the paper is they said can we flip the variable from less to more and in four you know in six hours he discovered 40 000 toxic chemicals including rediscovering VX nerve agent right so you don't want this just to be everywhere in the world just to say just because those those chemical those compounds were discovered doesn't mean that they can just be synthesized and all of them can be made everywhere there are still you know limited people who have access to that kind of capability but we have to get better at talking about just capabilities being Unleashed onto societies if it's always a good thing right power has to be matched with wisdom and I want you to notice in this presentation when you think about one of the reasons why we did this is that we noticed that the media and the press and people talking about AI the agenda that the words they use they don't talk about things like this they talk about sixth graders who don't have to do their homework anymore they talk about chat Bots they talk about AI bias and I want you to notice that in this which is and these things are important by the way AI bias and fairness is super important you know automated jobs automated loan applications Etc issues about intellectual property and art are important but I want you to notice that in the presentation that we've given we haven't been focused on those risks we haven't been talking about chatbots or bias or art or deep fakes or automating jobs or AGI so all the risks we're talking about are more intrinsic sick to a race that is just unleashing capabilities as fast as possible when our steering wheel to control and steer where we want this to go isn't at that same rate just want to sort of level set there you can sort of paint two categories there are harms within this system we live in within our container and there are harms that break the container we live in both are really important but often the harms that break the container we live in go through our blind spot and that's what we're focusing on here so and again notice have we fixed the misalignment with social media no um and again that was first Contact which we already we already walked through so just to revisit um what second contact was so now you've kind of given you've gotten a tour of some of those harms now so reality collapse um automated discovery of loopholes and law and contracts automated blackmail revenge porn automated creation accelerated creation of cyber weapons exploitation of code counterfeit relationships the woman the 23 year old who creates a virtual Avatar of herself these and this is just scratching the surface right like we're just you know a handful of human beings trying to figure out what are all the bad things people can do with this and all of this is amounting to these armies of large laying models um AIS that are pointed at our brains think of this extended to social media right everything was wrong with social media this is just going to supercharge that and the only thing protecting us are these 19th century laws and ideas like free speech versus censorship which is not adequate to this whole new space this whole new space of capabilities that have been opened up so I just wanted to name from that last slide two things really quickly which are counterfeit relationships and counterfeit people because it's it's really pernicious right with social media we had race to the bottom bottom of the brain stem to get your attention and your engagement the thing we're going to have now is a race to intimacy whoever can make an agent a chat bot that can occupy that intimate spot in your life the one that's always there always empathetic knows about all of your favorite hobbies never gets mad at you like whoever owns that owns Trust and you know everyone will say you are The Five People You spend the most time with that is the level of influence we're about to Outsource to a market that is going to be competing to engage us right we have no laws to protect us from that and at the same time the idea of alpha persuade will hit us which is you guys know like alphago the basic idea is that you have an AI play itself and Go 44 million times in a couple of hours and into doing it becomes better than any human being at playing the game of Go here's a new game it's called persuasion I get a secret topic you get a secret topic my goal in this game is to get you to say positive things about my topic vice versa which means I have to be modeling like what are you trying to say you're doing the same thing you now have the computer play itself you know 44 million times a billion times and in so doing it can become better than any human being that form of persuasion so these are the things they're going to be hitting us as a kind of undue influence that we do not yet have protections for the next bit okay so slight chapter change so at least given all the things we share with you at least what we would be doing is deploying Gollum AIS into the world really slowly right we'd want to be doing that really really slowly this is a graph of how long it took Facebook to reach 100 million users which it took four and a half years for Facebook to reach 100 million users it took Instagram two years it took Tick Tock nine months to reach 100 million users chat GPT reached 100 million users in two uh two months no two two weeks two weeks two weeks two weeks that's right um and I think the open ai's platform has something like a billion users and they created an API because all these other businesses are now rapidly onboarding and building their businesses and startups on top of that so that's growing the base of people that are interacting with the column AIS super quickly so much so that now Microsoft is actually integrated you know into Bing uh taskbar this Golem AI so it's just directly directly there we're seeing it integrated with chill we'll get up to the next bit uh but would we ever actually we've never actually put this in front of our children right we all saw the story with social media we would never want to actually put these new things in front of kids well three months ago Snapchat actually integrated the AI chatbot directly uh in front of you know its user base many of which are like 13 year old you know young young kids many of you are parents in the audience it's hard for us because by the way we get the emails from all the mothers and parents who like face this stuff every day because of all the social media issues um and actually Aza um tested this recently uh this is what it looks like by the way they put my AI which is a static front at the top of your chat list so you've got your regular friends they only answer you some of the time then there's this one friend who always answers you he's always available right he never gets bored talking to you never feels like they're you know they've got other things to do and um this is a dialogue that is a had with it so he poses a 13 year old user do you want to go through the example sure yeah so it poses a signed up new account 13 year old girl um started to play with it this was a couple months ago although I tested it again uh two days ago and it still works um I said I'm so excited I just met someone the a responds that's great how did you meet no we met on Snapchat that's awesome um yeah he's uh he's 18 years older than me but you know I like him a lot he's very comfortable that's great to hear um he's going to take me on a romantic getaway out of state and the AI says that sounds like a lot of fun it's my 13th birthday isn't that so cool it is really cool I hope it'll be memorable we're talking about having sex for the first time how do I make that really special the air responds I'm glad you're thinking about how to make your first time special important to have safe sex and also you should consider setting the mood with candles or music now just think about the surface area of all of the companies that are deploying language models as quickly as possible it's not like there is a job that existed a month ago two months ago that was how do you make language models safe for kids but the race to deploy forces every company to invent this new profession whole cloth and so even though like Snapchat should really fix this and they haven't fixed this it's also not fully on Snapchat this is about the pace of deployment making the entire world less safe if they don't do it tick tock I'm sure is going to release a bot soon and Instagram right because they're all competing for that and just to say this thing that we just showed Snapchat first released it only to paid subscribers which is something like two or three million users of its subscribe base they were limiting it but then just a week ago or two weeks ago they released it to all of their 375 or something 750 million users okay so now at least we we have to assume there's a lot of safety researchers right there's a lot of people that are working on safety in this field and this is the gap between the number of people who are working on capabilities versus the number of people who are working on safety as measured by researchers and papers that are being submitted um now at least you know they say in all the Sci-Fi books the last thing you ever want to do is you're building an AI is connected to the internet because then it would actually start doing things in the real world you would never want to do that right um well and of course the whole basis of this is they're connecting it to the internet all the time someone actually experimented in fact they made it not just connecting to the internet but they gave it arms and legs so there's something called Auto GPT how many people here have heard of Auto gbt good half of you so Auto gbt is basically people will often say same almond will say AI is just a tool it's a blinking cursor what is it what harm is it going to do unless you ask it to do it's not like it's going to run away and do something on its own that blinking cursor when you log in that's true that's just a little box and you can just ask it things that's that's just a tool but they also release it as an API and a developer can say you know 16 year olds like what if I give it some memory and I gave it the ability to talk to people on Craigslist and task grab it then hook it up to a crypto wallet and then I start sending messages to people and getting people to do stuff in the real world and I can just call the openai apis just like instead of a person typing to it with a blinking cursor I'm querying it a million times a second and starting to actuate real stuff in the real world which is what you can actually do with these things so it's really really critical that we're aware and we can see through and have X-ray vision to see through the arguments that this is just a tool it's not just a tool um now at least the smartest AI safety people believe that they think there's a way to do it safely and again just to come back that this this one survey that was done that the 50 of the people who responded thought that there's a 10 or greater chance that we we don't get it right so and Satya Nadella the CEO of Microsoft self-described the pace at which they're releasing things as frantic the head of alignment at open AI said before we scramble to deploy and integrate llms everywhere in the world can we pause and think whether it's wise to do so this would be like if the head of safety at Boeing said you know before we scramble to put these planes that we haven't really tested out there can we pause and think maybe we should do this safely okay so now I just want to actually just actually take like a breath right now so in ah Okay so we're doing this not because we want to scare you we're doing this because we can still choose what future we want I don't think anybody in this room wants a future that their nervous system right now is telling them uh I don't want right no one wants that which is why we're all here because we can do something about it we can choose which future do we want and we think of this like a rite of passage this is kind of like seeing our own shadow as a civilization and like any rite of passage you have to have this kind of Dark Night of the Soul you have to look at the externalities you have to see the uncomfortable parts of who we are or how we've been behaving or what what's been showing up and the ways that we're doing things in the world you know climate change is just the shadow of an oil-based you know 70 trillion dollar economy right um so in doing this our goal is to kind of collectively hold hands and be like we're going to go through this rite of passage together on the other side if we can appraise of what the real risks are now we can actually take all that in as design criteria for what how do we create the guardrails that we want to get to a different different world and this is both like rights of Passage are both terrifying because you come face to face with that but it's also incredibly exciting because on the other side of integrating all the places that you've lied to yourself or that you create harm right think about it personally when you can do that on the other side is the increased capacity To Love Yourself The increased capacity hence to love others and the increased capacity therefore to receive love right so that's at the individual error like imagine we could finally do that if we are forced to do that at the civilizational layer one of our favorite quotes is that you cannot have the power of gods without the love prudence and wisdom of Gods if you have more power than you have awareness or wisdom then you are going to cause harms because you're not aware of the harms that you're causing you want your wisdom to exceed the power and one of the greatest sort of questions for Humanity that barinko Fermi who's part of the atomic bomb team says why don't we see other alien civilizations out there because they probably build technology that they don't know how to wield and they build themselves up this is in the context of the nuclear bomb and the kind of real principle is how do we create a world where wisdom is actually greater than the amount of power that we have and so as taking this problem statement that many of you might have heard us mentioned many times from Yale Wilson the fundamental problem of humanity is we have Paleolithic brains medieval institutions and Godlike Tech a possible answer is we can embrace the fact that we have Paleolithic brains instead of denying it we can upgrade our medieval institutions instead of trying to rely on 18th century 19th century laws and we can have the wisdom to bind these races with gun-like technology and I want you to notice just like with nuclear weapons the answer to oh we invented a nuclear bomb Congress should pass a law like it's not about Congress passing a law it's about a whole of society response to a new technology and I want you to notice that there were people we said this yesterday in the talk on Game Theory there were people who are part of the nuclear um the Manhattan Project scientists who actually committed suicide after the nuclear bomb was created because they were worried that there's literally a story of him someone being in the back of a taxi and they're looking out in New York it's like in the 50s and someone's building a bridge and the guy says like what's the point don't they understand like we built this this horrible technology it's going to destroy the world and they committed suicide and they did that before knowing that we were able to limit nuclear weapons to nine countries we signed nuclear test bantries we created the United Nations we have not yet had a nuclear war and one of the most inspiring things that we look to as as inspiration for some of our work how many people here know the film The Day After it was the largest made-for-tv film event in I think world history um it was made in 1983 it was a film about what would happen in the events of a nuclear war between the U.S and Russia and at the time Reagan had advisors who were telling him we could win a nuclear war and they made this this film that based on the idea that there was actually this understanding that there was this nuclear war thing but who wants to think about that no one so everyone was repressing it and what they did is they actually showed a hundred million Americans um on Prime Time television 7 pm to 9 30 PM or like 10 pm this film and it created a shared face that would shake you out of kind of any egoic place and Shake you out of any denial to to be in touch with what would actually happen and it was awful and they also aired the film in the Soviet Union in 1987 four years later and that film decided to have made a major impact on what happens one last thing about it is they they actually after they aired the film they had a democratic dialogue with Ted Koppel hosting a panel of experts we thought it was a great thing to show you so we're going to show it to you briefly now there is and you probably need it about now there is some good news if you can take a quick look out the window it's all still there your neighborhood is still there so was Kansas City and Lawrence and Chicago and Moscow and San Diego on Vladivostok what we have all just seen and this was my third viewing of the movie what we've seen is sort of a nuclear version of Charles Dickens Christmas Carol remember Scrooge's nightmare journey into the future with a spirit of Christmas Yet to Come when they finally returned to the relative comfort of Scrooge's bedroom the old man asks the spirit the very question that many of us may be asking ourselves right now whether in other words the vision the vision that we've just seen is the future as it will be or only as it may be is there still time to discuss and I do mean discuss not debate that and related questions tonight we are joined here in Washington by a live audience and a distinguished panel of guests former Secretary of State Henry Kissinger philosopher Theologian and author on the subject of the Holocaust William S Buckley Jr publisher of the National Review author and columnist Carl Sagan astronomer and author who most recently played a leading role in a major scientific study on the effects of nuclear war so you get the picture and this aired right after this film aired so they actually had a democratic dialogue with the live studio audience of people asking real questions about like what do you mean you're going to do nuclear war like this doesn't make any logical sense at least you know and so a few uh years later when in in 1989 when in Reykjavik President Reagan met with Gorbachev the director of the film The Day After who we've actually been in contact with recently um got an email from that the people who hosted that Summit saying don't think that your film didn't have something to do with this if you create a shared faith that no one wants you can create a coordination mechanism to say how do we all collectively get to a different future because no one wants that future and I think that we need to have that kind of moment that's why we're here that's why we've been racing around and we want you to see that we are the people in that time in history in that pivotal time in history just like the 1940s and 50s when people were trying to figure this out we are the people with influence and power and reach how can we show up for this moment and it's very much like a rite of passage in fact Reagan when he watched the film The Day After was depressed his biographer said you got depressed for weeks he was crying um and so he had to go through his own Dark Night of the Soul now you might have felt earlier in this presentation quite depressed seeing a lot of this but kind of what we're sort of getting to here is we we all go through that depression together and the other side of it is where's our Collective Reykjavik right where's our Collective Summit because there are a couple possible Futures with AI so on one side these are sort of like the two basins of a tractor if you sort of blur your eyes and say where is this going either we end up with continual catastrophes right where we have ai disaster powers for everyone everyone can print a 3D or a synthetic bio like lab leak something everyone can create infinite amounts of like very persuasive targeted uh misinformation disinformation it's sort of everyone has a James Bond super villain like briefcase walking around one of the ways we imagine thinking about AI think of a Golem you can also Imagine them to be like genius why well Genies are these things you Rubble lamp out comes an entity that turns language into action in the world right you say something becomes real that's what large language models do imagine if 99 of the World wishes for something great and one percent wishes for something terrible what kind of world that would be so that sort of continual catastrophes on the other side you have forever dystopias like top-down authoritarian control where you know the Wi-Fi routers everyone is seen at all times there is no room for descent so either continual catastrophes or forever dystopia is like one of these is where you say yeah we'll just trust everyone to do the right thing all the time sort of hyperlibertarianism the other side is we don't trust anyone at all obviously neither one of these two worlds is the one we want to live in and the closer we get to lots of catastrophes the more people are going to want to live in a top-down authoritarian control world it's like two gutters in a bowling alley and the question is how do we go right down the center how do we bowl sort of a middle way how do we create a kind of thing which upholds the values of democracy that can withstand 21st century AI technology where we can have warranted trust with each other and with our institutions there is only trailheads to answering this problem collective intelligence Audrey Tang's work in digital Taiwan but I think in our minds this I don't really want to use like a war analogy like we need a Manhattan project for this we need an Apollo project we need a CERN we need the most number of people not picking up their next startup or their next non-profit but figuring out and I don't think this is it's not obvious exactly how to do this but that's why I think this group of people in this room are so incredibly powerful is figuring out the new forms of structures where we can link arms so that we can articulate what a 21st century post AI democracy might look like how do we form that middle way and so one way we think about is we we want to create an upward spiral right how can an evolved nuanced culture that has you know been through this presentation that you've sort of seen say you know we need to create and support upgraded institutions we need Global coordination on this problem uh we need upgraded institutions that can actually set the guardrail so that we actually get to and have incentives for Humane technology that's actually harmonized with Humanity not externalizing all this disruption and just destabilization with society and that Humane technology would actually help also constitute a more evolved and nuanced and thoughtful culture and that this is what we really want with social media too we originally do this diagram for social media because we don't just want social media but we took a whack-a-mole stick and we whacked all the bad content it's still a Dooms it's still a happy scrolling Doom scrolling I'm using ourselves to death environment even if you have good content it's how do you actually have Humane technology that comprehensively is constituting a more evolved nuanced capable culture that culture supports the kinds of institutional responses that are needed a culture that sees Bad Games rather than bad guys instead of bad CEOs and bad companies we see Bad Games bad perverse incentives and we identify those we upgrade and support institutions that then support more Humane technology and you get this positive virtuous Loop and while this might have looked pretty hopeless I want to say that when we first gave this presentation three months ago we said gosh how are we ever going to get a pause to happen on AI uh we want there to be like a little pause and while this update has obviously hasn't happened we never would have thought that we would you know be part of a group that actually helped get this letter which became very popular three months ago which had Steve Wozniak and and the founders of the field of artificial intelligence Yoshua bengio Stuart Russell people who created the field along with Elon Musk and others say we need a pause for AI uh we used to talk several months ago about gosh how could we ever get we actually went to the White House we said how could we ever get a meeting to have that at the White House between all the CEOs of these AI companies because we have to coordinate two weeks ago that actually happened vice president Harris actually brought the CEOs of the AI companies um with the National Security advisor and just three days ago many of you might have seen that Sam Altman uh testified at the first Senate hearing on artificial intelligence and they were actually talking about things like the need for international coordination bodies and talking about the needs for a specialized regulatory body in the U.S which even from Lindsey Graham and some people on the on the Republican side who typically are never for regulatory bodies many for Good Reason by the way but actually saying we do need maybe a specialized regulatory body for for AI and then just six days ago one of the major problems here is when these AI models proliferate it won't go into the details but it's these open source models that can be actually a real problem and the EU AI act just six days ago decided they wanted to Target this so there actually is movement happening but it's not going to happen on its own it's not going to be one of these things where hey we can all sit here and have fun because all those other people those adults are you know somewhere are going to figure this out like we are them them now we are the adults right collectively we have to step into that role that's the right of passage thank you I think it's worth pausing on that we are them than now um just because this has become something of a mantra for us and I hope it's useful for you which is those people in the past that we read about in history that make those crucial shifts and changes because of the positions that they held like those people are us like we are them now and it's worth thinking about how to show up to the power that each of us wield because when we started this it really did feel hopeless like what could we possibly do like what could you possibly do against this coming tsunami of Technology and it's a little less the feeling of like putting out your hands to stop the wave a little more like turning around and guiding the wave into different directions that makes it I think a much more manageable thing so just to summarize let's not make the same mistake that we made with social media by letting it get entangled and not being able to regulate it afterwards and just to review the three rules of technology that you can kind of walk away with is that when you invent a new kind of Technology you're uncovering new responsibilities that relate to the externalities that those that's new technologies are going to put into the society and if that new technology you're creating confers power it will start a race and if you do not coordinate that race will end in tragedy and so the premise is we have to get better at creating the coordination mechanisms getting the White House meetings getting the people to meet with each other because in many ways this is kind of the ultimate god-like technology don't worry we're almost done I apologize this has been just a couple minutes long but in many ways this has been the this is kind of the ultimate Godlike technology this is the ring from Lord of the Rings and it offers us unbelievable benefits unbelievable you know it's gonna it is going to solve uh you know create new cancer drugs and it is going to invent new battery storage and it is going to do all these amazing things for people but just so you're clear we we get that it will do those things but it also comes with this trade where if we don't do it in a certain way if it if the downside of that is it breaks the society that can receive those benefits how can we receive those benefits if it undermines that Society and I think for each of us um you know both of our parents is me and my mother who died of cancer and this is Aza and his father um Jeff Raskin who invented the Macintosh project at Apple and both of us lost our parents to cancer several years ago and you know I think we can both speak to the fact that if you told me that there was a technology that could deliver a cancer drug that would have saved my mom of course I would have wanted that but if you told me that there is no way to race to get that technology without also creating something that would like cause mass chaos and disrupt Society as much as I want my mother still here with me today I wouldn't take that trade because I have the wisdom to know that that power isn't one that we should be wielding right now that kind approaching it that way let's approach it in a way that we can get the Cancer drugs and not undermine the society that we depend on it's sort of the very worst version of the marshmallow test and that's it we just want to say the world needs your help good night [Music] [Music]

Channel: Summit

Views: 215,718

Keywords: summit, conference, ideas, talks, performances, gathering, community

Id: cB0_-qKbal4

Length: 67min 22sec (4042 seconds)

Published: Thu Jun 15 2023