We present the following presentation as a cautionary tale. This is a hint at the true state of the art in electronic surveillance and control technologies.
For the most effective countermeasure against these unseen influences, we recommend The Halopulser.
All credit to the @SummitSessions channel on YouTube for their excellent material.
What does it look like to align technology with humanity’s best interests?
Tristan Harris and Aza Raskin discuss how existing A.I. capabilities already pose catastrophic risks to a functional society, how A.I. companies are caught in a race to deploy as quickly as possible without adequate safety measures, and what it would mean to upgrade our institutions to a post-A.I. world.
Learn more about Tristan and Aza's work and listen to their podcast, Your Undivided Attention, at humanetech.com/podcast.
This talk was recorded at Summit At Sea in May 2023.
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 Aza Raskin, the other co-founder. And what did we just see, uh, Iza? 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, a hundred people in the world playing with this technology. Now, I think there have probably been a billion plus images created. And the reason why 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 reporters 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 reporters.
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 I just wanted 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 AI 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 930 a. m. Earth Species Project using AI to translate animal communication, like literally, Transcribed 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 gonna switch into a mode where we're really gonna look at the dark side of some of the AI risks that are coming to us and just to say Why we're doing that a few months ago some of the peck the people inside the major AGI companies came to us and said 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, Aiza and Tristan 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, uh, what is the likelihood that, uh, 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 percent or greater chance that we would go extinct. So imagine you're getting on a plane, right, a Boeing 737, and half of the airplane engineers who were surveyed said there was a 10 percent chance if you get on that plane, everyone dies.
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. Here are the three rules. 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 non 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. And 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. 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 Eiza 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 supercomputer 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 utterly simple technology, was enough to cause, in this first contact with social media, Information overload, addiction, doom scrolling, influencer culture, sexualization of young girls, polarization, cult factories, fake news, breakdown of 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 room 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, uh, between AI and humanity. Uh, 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 did we lose? What was the story we were telling ourselves? Well, we told ourselves, we're giving everybody a voice.
Uh, connect with your friends. Join like minded communities. We're gonna enable small, medium sized businesses to reach their customers. And all of these things... Or true, right? These are not lies. 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 brainstem 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 maximizing 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, um, 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 we'll 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, uh, 4, 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 that have been playing with GPT 4, 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 AlphaFold. 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 they're 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 this 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 it. That's why we're here in front of you. That's why we've been racing between Washington, D. C. 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 just give you a preview of what we're going to explore. Is reality collapse, automated loopholes in 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 for social media is, How do we realize the benefits of our technology that 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 that 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. Um, and I just want to say that we've also been skeptical of AI too. Um, I actually kind of missed some of this kind of coming up. Eiza's been scanning this space for a while. Why are we skeptical of AI? Well, you know, you use Google Maps and it still mispronounces, uh, 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 15 minute timer.
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 why should we feel like we should be concerned now? Go ahead. Well, because 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, it's 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 2 percent advance over here, and that didn't do anything for, say, like music generation if it was from, uh, 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. So, uh, this became what's known as the generative large language multi modal models. This phase, this phase has so many different terminology, large language models, etc. We just wanted to simplify it by, if it's called glllmm, we're like, hmm.
Let's just call that a golem, because golem is from Jewish mythology of an inanimate creature that gains its kind of own capabilities. And that's exactly what we're seeing 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 in text outcome outcomes and 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. Um, 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. 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 semantic knowledge embedded in this space. All right, let's try another one. Instead of like images and text, how about this? Um, 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, where 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. 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 or 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, the sonar from the, uh, 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, um, some code. This is from like a, a mail server. And I said, uh, 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 WiFi 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 we, 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 fact that, um, uh, images can, I mean, uh, uh, the new AI can actually copy your, your voice, right?
You can get someone's voice, Obama and Putin, people 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 are of a real person speaking, even though she sounds a little bit metallic. The rest is just what the computer automatically generated.
9 cases out of 10, 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 are 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, as we saw these first demos, we sat and thought, like, how is this going to be used? We're like, oh, you know what 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 forgot my social security number. I'm applying for this thing. What was it again? Um, and we're like, that's scary. We thought about that conceptually and then this actually happened.
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 exact thing happening. Um. And then one month ago, you know, AI clones, teens grow a voice in a 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 TikTok. 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 dog filter on Snapchat. And now, this, this filter gave me lip fillers. 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, uh, 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, in their likeness, et cetera. 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 We'll 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, religions are language. We did an op ed in the New York Times with Yuval Harari, 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 is 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 4, 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 and 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, Yuval also points out when we're having a conversation with him, he's like, when was the last time?
That a non human entity was able to create large scale influential narratives. He'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 images 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 Gollum class AIs is that they gain emergent 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 with more parameters, and here's a test, like, um, 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, doesn't get better, doesn't get better, doesn't get better, and then suddenly, boom, it answer questions in Persian. And the engineers who 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. Alright, 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, uh, strategy. Um, and strategic thinking. Um, 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 test.
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, nine months later, um, ten months later, it was at the level of a nine year old theory of mind. Which doesn't mean that it has the strategy level of a theory, uh, 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, GPT 4 came out. Anyone want to make a guess? This could, it could have topped out.
It's better than the average adult, um, at. Theory of mind. So think about like when, 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 R L H F 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. Um, 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 asked 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, uh, 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 emerged.
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 3, right? Yeah, this was GPT 3. Uh, had actually, uh, discovered the, 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 godlike intelligence into everyone's pocket without thinking about what are the capabilities that I'm actually handing out here, right? And the punchline 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 shipped it 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, 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 text? Right? Well, you end up in this kind of situation. Feed me! Tui! You talked! You opened your trap! You sing and you sing!
Feed 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 released it open source, why would they do that? And you're like, oh right, because they ran out of text on the internet, we're going to have to go find more text somewhere. How would you do that? Well, it turns out YouTube has lots of people talking. Podcast, 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.
Um, and that's exactly what they did. Recently, uh, 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 percent of that code two and a half times faster. So imagine that the AI then points it at its own code. It can make its own code two and a half times faster. Um, 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? It's 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, um, GPUs, the things that AI runs on, faster, and they're able to do that. And in the latest H100s, NVIDIA's latest chip, there are actually 13, 000 of these sub modules that were designed by AI.
The point is that AI makes the chips that makes AI faster. And you can see how that becomes a recursive flywheel. And, sorry. No, and this is 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, uh, can recursively improve if it's implied to itself. Um, and so that's really what distinguishes it. It's hard for us to get our mind around it. Think about it, 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 with is what is the category of this thing? And people know this old, uh, kind of adage that if you give a man to fish, you feed him for a day.
You teach a man to fish, you feed him 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, and oceanary 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, uh, on it.
We're gonna have to update all the children's, uh, like, childhood, uh, books. Um, alright. 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 was a 30, 000 pot, um, for making the best predictions. And they asked, when will AI be able to solve competition level mathematics with greater than 80 percent accuracy? So this is, this is last year. Um, and the prediction that these experts made was that AI will reach 52 percent accuracy in four years.
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.
A factor of four.
Also, AI is beating tests as fast as people are able to make them. It's actually become a problem in 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 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 have 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 somewhere that'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 guardrails. 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, 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 they did in the paper is they said can we flip the variable from less to more?
And in four, no, in six hours, it discovered 40, 000 toxic chemicals, including rediscovering the X nerve agent. Right, so you don't want this just to be everywhere in the world, and it's to say, uh, just because those, those chem, 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, the words they use, they don't talk about things like this.
They talk about 6th graders who don't have to do their homework anymore. They talk about chatbots, 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.
Um, 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 deepfakes, or automating jobs, or AGI. So all the risks we're talking about are more intrinsic 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.
I just want to sort of level set there. You could sort of pick two categories. There are harms within the 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 in law and contracts, automated blackmail, revenge porn, automated creation, accelerated creation of cyber weapons, exploitation of code, um, 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, this extended to social media, right?
Everything that 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, uh, 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 brainstem 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 chatbot, 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. 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, like, you guys know, like, AlphaGo, the basic idea is that you have an AI play itself in Go 44 million times in a couple of hours, and in so 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 at form of persuasion.
So these are the things that are going to be hitting us as a kind of undue influence that we do not yet have protections for. Just want to note on time, we'll probably rush through some of the next bit. Okay, so slight chapter change. So at least, given all the things we've shared 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. Um, 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 TikTok nine months to reach a hundred million users. ChatGPT reached a hundred million users in two, uh, two months.
No, two, two weeks, two weeks, two weeks. That's right. Um, and I think the OpenAI'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 Gollum AIs super quickly.
So much so that now Microsoft is actually integrated, you know, into Bing, uh, Taskbar. This Gollum AI, so it's just directly, directly there. Uh, we're seeing it integrated with children, we'll get to the next bit. Uh, but would we ever actually, we would never actually put this in front of our, our children, right?
I mean, 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, it's 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 Eiza, um, tested this recently. Uh, this is what it looks like, by the way. They put my AI, which is a static friend, at the top of your chat list.
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, uh, a dialogue that Aza had with it.
So... He poses a 13 year old user. Do you want to go through the example? Sure, yeah. So he 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 two days ago and it still works. Um, I said, I'm so excited. I just met someone. The AI responds, that's great.
How did you meet? Oh, 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. Hope it'll be memorable. We're talking about having sex for the first time. How do I make that really special? The AI 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's 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, TikTok 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 2 or 3 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 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 would 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 it to the internet, but they gave it arms and legs. There's something called auto GPT. How many people here have heard of auto GPT? A good half of you. So, auto GPT is basically, um, people will often say, Sam Altman 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? 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, Hmm, what if I give it some memory and I gave it the ability to talk to people on Craigslist and TaskRabbit, 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 API, so 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 bullshit 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 50 percent of the people who responded thought that there's a 10 percent 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 OpenAI 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,
Let's actually take like a breath right now. In.
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 in 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 gonna 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 rites of passage are both terrifying because you come face to face with death, 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 layer. 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 Errico Fermi, who is 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 wheel 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 mention many times from E. O. Wilson, the fundamental problem of humanity is we have Paleolithic brains, medieval institutions, and god like 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 19th 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 are people, we said this yesterday in the talk on game theory, there are 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 someone being in the back of a taxi and they're looking out in New York, it's like in the fifties, 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 ban treaties. 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, uh, for some of our work. How many people here know the film, The Day After? Quite a number of you. 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 event of a nuclear war between the U.
S. and Russia. And at the time, Reagan had advisers who were telling him, we could win a nuclear war. And they made this, this film that, based on the idea that there is actually this understanding that there's 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 primetime television, 7pm to 9. 30pm, uh, or like 10pm, this film. And it created a shared fate 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 happened. 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 is Kansas City, and Lawrence, and Chicago, and Moscow, and San Diego, and 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 the spirit of Christmas yet to come? When they finally return 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 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.
Elie Wiesel, 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 and people asking real questions about like, what do you mean you're gonna 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 19, I think 89, when in Reykjavik, president Reagan met with Gorbachev, the director of the film, the day after, who we've actually been in, 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 fate 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 he 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 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, um, a synthetic bio, like LabLeak something. Everyone can create infinite amounts of, like, very persuasive, targeted, uh, misinformation, disinformation.
It's, sort of, everyone has a, a James Bond super villain, like, briefcase walking around. Um, one of the ways we imagine thinking about AI, Gollum, you can also imagine them to be like genies. Why? Well, genies are these things you rub a 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 percent of the world wishes for something great, and 1 percent wishes for something terrible, what kind of world that would be. So that's 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 dissent. So either continual catastrophes or forever dystopias, like one of these is where you say, yeah, we'll just trust everyone to do the right thing all the time, sort of hyper libertarianism.
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, uh, 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 of that 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 guardrails so that we actually get to and have incentives for humane technology. That's actually harmonized with humanity, not externalizing all this disruption and destabilization with society. And that humane technology would actually help.
Also constitute a more evolved and nuanced and thoughtful culture. And this is what we really want with social media too. We originally drew this diagram for social media because we don't just want social media where we took a whack a mole stick and we whacked all the bad content. It's still a doom, it's still a happy scrolling, doom scrolling, amusing 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, Um, a 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?
We want there to be like a little pause. And while this 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 got, became very popular three months ago, which had Steve Wozniak 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, uh, several months ago about, gosh, how could we ever get, we actually went to the White House and said, how could we ever get a meeting to happen at the White House between all the CEOs of these AI companies? Because we have to coordinate. Two weeks ago, that actually happened. The 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, uh, 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, um, one of the major problems here is when these AI models proliferate, I 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 then 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? 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. Um, 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 new technologies are going to put into 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 God like 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, 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, uh, Both of our parents, me and my mother who died of cancer, and this is Eiza and his father, 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. Thanks. Sort of the very worst version of the marshmallow test. And that's it. We just want to say the world needs your help.