That's really good to be here. It's nice to see all of you.
I prepared my remarks. The only thing I had AI do, well, I had AI do two things. I had AI organize the remarks into a flow and then also on your website you have a trained an avatar of yourself to answer questions. So I asked your avatar all these questions ahead of time and I'm just curious to see to what extent your avatar is going to answer the way that you're going to answer. So, but it's not a gotcha moment. Anyway, I just thought that was interesting. I was like, your avatar was there and it said, "Talk to me." So, I said, "Okay, well, this is a great opportunity."
It's interesting. You know, we worked on that avatar for a long time, like a year and a half now. And it was really bad initially. It kept on saying that I was I would never say that. And then over time, it got better and better. It still says some weird stuff that I'm not happy with. But one interesting thing I learned is when people talk to my avatar, my avatar makes mistakes and hallucinates. One of my internal team members said, "I talked to Avatar and it told me about this great mountain climbing thing you did." I said, "I've never climbed those mountains, but people still seem satisfied talking to my avatar."
I can't remember. I think it was ChatGPT when I was It's the equivalent of Googling yourself, but I was asking ChatGPT what just to give me a quick bio of myself and it said that I was also a survivor of pancreatic cancer and I've never had cancer much less pancreatic cancer and so I was like where the hell did that come from? Yeah, they do they do still hallucinate. What what since we're talking about this on average do you know what percentage of the time the current models are hallucinating?
Yeah, I I don't know. They definitely do hallucinate, but probably less than most people think. Two, three years ago, AI models were much less advanced than they are today. And there were some widely publicized examples of AI models hallucinating like, you know, the lawyer that got into trouble for made up court filing details and so on. The technology has gotten much better and it still does make mistakes, but probably less than most people think.
Great. Okay. So, I have a couple of prepared questions. We'll get into a conversation. You you may have some questions for me and then hopefully we can take some questions from the audience. I just wanted to start with a really broad framing question that we had a we zoomed a week or so ago. So, we had a kind of pregame and I now now know which side of the fulcrum you stand on, but there are there are AI doomsayers and sort of AI utopianists. And so, where on that spectrum would you say you fall?
The boring answer, I'm actually neither of the above. I think I'm much more in the middle. The way that AI narratives have evolved over the last few years has been kind of weird and there's been a lot more doomsaying or a lot more fear narratives about AI being very destructive than most people would have expected. The funny thing that happened was a handful of companies decided that stoking fear about AI was wonderful PR because it turns out that if you go around and say my technology is so powerful I could cause the extinction of the human race. That turns out to be one way to convince people your technology is really powerful and therefore must be worth a lot. After that narrative fell apart, the next one became, now my technology could take all of your jobs, so how much is my technology worth? Got to be a lot, right? And that narrative is slowly falling apart. Oh, and then a more recent one, my technology is so smart it could blackmail people.
We've seen lots of stories about that.
Yeah. And it's been interesting how isolated lab testing in a highly artificial environment has been hyped up as if it was a real thing. So maybe sometimes think about the way we test airplanes. Imagine if we were to test an airplane in a wind tunnel and simulate a hurricane and if in a wind tunnel I managed to create a hurricane conditions to make an airplane break apart. But the newspaper then says, "Oh, airplanes can break apart midair." Well, it never actually happened in real life. That, by the way, is what happened with the blackmailing episode. It turns out that you put AI in a lab and poke and probe it and have some of the most skilled scientists in the world bend over backwards to try really hard to push the AI into the corner. So, has no choice but given all the conscious instructions and no choice but to kind of look like a blackmail someone. That's what happened. What happened in the media was, "Wow, AI is so dangerous. It could blackmail people." And the damaging thing about this is that a lot of these narratives has a tiny gem of truth in it. You know that yes, airplanes in hurricanes will break apart in midair, but I don't know of many airplanes that broke apart in midair, you know, recently because we just don't fly them through hurricanes. And so it's been very difficult to push back because you can't say airplanes will never break apart in midair. And you can't say there's no circumstance under which even a lab test you can't make AI do something bad. But the amplification of these narratives has been very damaging for America. Because it's holding back what should be great opportunities for us to pursue, but instead, you know, we see kids turned off AI, people wanting to not use these wonderful technologies that could make them so much better, wrong.
Yeah, I can speak to this. I'll tell two quick anecdotes. I don't know how many of you read the article in the New York Times about the AI companies doom trolling last weekend. And it seems on the outside of it kind of counterintuitive to a lot of people. Why would the AI companies be saying they could cause the extinction of humanity? I think it was about eight months ago. I was part of a group of filmmakers that were invited up to Silicon Valley and for a two-day symposium and it was a lot of the bigwigs from DeepMind and Anthropic and OpenAI and all these various engineers gave talks on the promise and peril of AI but a lot of it was very scary and they were saying oh within four to six years four billion people are going to be out of a job. Well, they they they took a hand count and I can't name some of the people that were there, but there were people that, you know, in the AI pantheon, you know, how many people think there's a chance that AI will cause the extinction of the human race and you know, a third of these people raised their hands. You know, that there's that's a possibility. Anyway, I'm not pushing back on what you're saying. What I'm what I'm saying is I experienced firsthand this. I was trying to figure out why the hell are they trying to scare us? It seems counterintuitive. And so anyway, you had you have a lot to say. By the way, you have a very informative newsletter. I think it's called the batch.
Which I encourage everyone to join. It's super informative. Also,
And so I'm curious you've you've come away you know with I think a what sensible view of the odds of AI human extinction and all that. Do you want to share your views as well?
Well I'm I'm I'm kind of in the middle in terms of I think there are massive benefits. I think personally now I'm just a guy who does vampire movies and stuff like so what do I know? But I I I'm in the middle. I I think that the the that possibilities of of AI super intelligence is probably farther out than people think. I think the possibility of AI this is what I now think causing the extinction of the human race anytime within our lifetime is highly unlikely. I personally think, and we can talk about this in a few minutes, that the AI displacing workers, the threat might be bigger than you're saying it is, but that's okay. We can talk, we can talk about that. And I've certainly seen that, experienced that in Hollywood to a certain extent. So, I'm in the middle, but I've I've also been using AI myself for the last two or three years. And I've seen lots of ways that it has sped up my work, whether it be through concept art iteration on something I'm doing or whether it be through visual effects or there are different ways that I use it in the writing process and the research process as well. So, you know, it's it's a tool and like any tool, I think, you know, a hammer can be used to build and a hammer can be used to hurt someone. So, it it just depends.
You know, on on on the job issue that that you mentioned, one of the hype narratives has been AI will lead to millions of people unemployed and it does help people convince be convinced maybe that the technology is very valuable, but I'm not seeing it happen for this reason. So, chatting with CEO of a large workforce management company that actually ran surveys of people that were recently laid off and we see all this news about AI leading to layoffs, right? Company comes saying AI is coming, laying off people, blah blah blah. He ran a survey and asked of the people that were actually laid off, were you replaced by AI or not? And the number, what do you think it was? Was it 50%, 20%, 10%? It was 1.4% of recently laid off people actually had their jobs replaced by AI. To be clear, I'm very sympathetic to this 1.4%. I'm very worried. A tiny sliver, a small fraction of people have had their jobs replaced by AI and it sucks for them and we need to do something about that. Give them a safe work with them safety net provide them a way to reskill, upskill, learn, you know, other things to to find new jobs. So that's really important to help this 1.4%. And again, it's one of these things where there's a gem of truth. There's a tiny number of people whose jobs are going away because of AI, but the hype is vastly greater than the reality. And the broader reality is that if you look at the task-based analysis of jobs, these are very rigorous studies that economists will do where we look at the jobs done by say the American economy or the global economy, break down the job into discrete tasks and then figure out what can AI actually automate. It turns out that for many jobs AI could plausibly automate you know 30 40% of the task in someone's work. And it does mean that someone that doesn't use AI at all, they will be much less productive than someone that does. And so that pathy saying that you people won't be replaced by AI, but people will be replaced by someone else that uses AI. That is absolutely true. But at the same time, if AI automates, you know, 30 40% of someone's job, you really need a human to do that remaining 60 70%. Historical trends have shown that when this happens, when technology automates part of someone's job, the complement to it, the other thing that needs to go with it becomes even more valuable. And so historically, the trends have been automation have led to higher wages for people. Which is why when we went from horses to cars, they became tons more jobs for taxi drivers because taxi drivers are a compliment for, you know, the taxi for the car. But slowing down people's willingness to adopt this technology is really hampering people's trajectory to get to these often better, higher paying jobs.
Okay, switching gears for a second. I'm curious, you know, you've briefed lots of policy makers on AI. So I'm curious to what degree do you feel our our government to the extent that you can talk about it is sufficiently up to speed on AI and what would you say is the the biggest mischaracterization or blind spot that you've run into when talking about when talking to our government?
You know, I feel like in the US, the federal government has mostly got its head together in a more reasonable way than one would have hoped a couple years ago.
I was surprised to hear this.
So, I mean, clearly we have a lot to drown to cover, but two years ago, the AI and extinction narratives were making us rounds in Washington DC. And these were part of the regulatory capture attempts of a handful of AI companies that want to say AI will lead to human extinction regulated. And in particular, it turns out that one of the major threats to some leading American AI companies is when you spend tens of billions of dollars training large AI models called foundation models. It turns out that some companies, including overseas ones in China and elsewhere, are training models and releasing them on the internet free for anyone to use. We call these open weight models. But it's really annoying if you spend billions of dollars to train an AI model and then someone else posts an equivalent one on the internet free for anyone to use. It really degrades the value of your investment. So about two years ago, I was actually kind of really surprised by the intensity of the lobbying to drum up fears on AI to try to pass stifling regulations to make sure that, you know, to make it difficult for anyone to publish their AI models on the internet free for anyone to use. Fortunately, a lot of the AI community has largely beaten these efforts back at the federal level, but the lobbying because there are, you know, billions or more likely trillions of dollars at stake, you can imagine the resources and the intensity of the lobbying is intense. There are also very credible reports that some of Americans adversaries are busy funding nonprofits to stoke up unnecessary fears over AI safety to try to slow down American innovation in AI. The credible reports about our adversaries funding American and other European organizations to slow us down is also very concerning. Having said that, I think both White House and Congress have had enough talks with AI. You know, Schumer ran a bipartisan AI insight forum. Bunch of people including me spoke with Senate Congress. And I think is you know could have been much worse. I worry much more about state level regulation of AI because a lot of the lobbying efforts have been now at the state level because you can get a state level stifling regulation passed. It could have deleterious effects across the entire nation and there's so many state houses. So constantly squashing back referee proposals to the state level has been much harder.
So one of the one of the narratives that we hear in terms of you know not not not stifling innovation through overregulation is that at least in America that we're in a race with China. That's one of the things that we often hear and and that if we don't get there first, China will become ascendant and then whatever the the West will be over. I'm curious because you're someone who's worked in AI initiatives for both the US and for China. So, one of the questions that I asked you in kind of our pregame was how would you define the different approaches between the two countries in terms of AI, you know?
So, one thing that I would not have guessed few years ago was that China would end up being much more open in the way it innovates in AI compared to the United States. So again, I think I know the US ecosystem really well since I you know, both Google Brain and Sam and I I might be the only person on the planet that both Sam and Dario has worked for, but so I feel like I know all the teams of people well. But what has happened at the moment that ChatGPT was released in the end of 2022, America was decisively ahead of China. But China since then has played its hand really well leaning into the open-source open-weight ecosystem in which teams in China will do research and publish their research papers free on the internet for anyone to read. And by the way, I'll tell you a lot of the all the American teams read tons of the Chinese research papers and learn from them and benefit from them. Tons of the techniques used by leading American teams now frankly were from research innovations learned from China. So I think there's very much two-way sharing of knowledge. But one thing that China did really well was a lot of its companies will release this technology as open weight models meaning they're published on the internet free for anyone to use including American companies but also
Can you just define what open weights are for the audience? How many of you know what open weights are? Some. Okay, cool. Not bad. Oh, thank thanks for good call. Thank you.
So, when you have a AI model, read tons of text on the internet, look at tons of images, watch lots of videos. I know copyright is the sole point, but after AI has learned from lots of data, it results in a long list of numbers. We call those numbers weights. And an open weight model is when someone has trained an AI model on lots of data and it publish and you publish all of those numbers called weights on the internet free for anyone to download and use. And so when there's an open weight model any of us I could download it you could download it and have AI to run on your own laptop as opposed to the proprietary closed weight models such as the leading models from OpenAI Anthropic Google Gemini. Those are closed weight models because none of us can see the numbers that AI had learned. We can only send our prompt to one of these companies and have them use these weights or these numbers compute the output, compute the response and show us only the response. One brilliant move that Chinese companies did was publishing open weight models. It turns out to help the whole world but you help yourself the most because it rapidly increases the rate of diffusion of knowledge within your companies or within your ecosystem. And so because of factors like that over the last few years, China has raced ahead in AI capabilities. And right now the US closed weight models is ahead of the Chinese models. But China leads the world in open-weight models. And what this means is for a lot of nations that want an open alternative they're adopting Chinese models. So adoption of Chinese models in the US is you know there's a lot but in nations such as Africa the Chinese DeepSeek and a handful of other Chinese models are adopted very widely far more than say the American open weight models which have really fallen behind and this is important for two reasons. But first, just as storytelling we've seen in Hollywood is an important source of soft power, do we want stories told about the importance of liberty and democratic values or do you want stories told that reflect other nations values? It turns out that when someone asks a question of AI such as what happened in Tiananmen Square in 1989, well, whose model they are using will cause the model to give an answer that reflects that nation's values. So getting your stuff out there is a tremendous source of soft power influence. And then the second thing is AI is a key part of the supply chain of how we build a lot of software products. And if more and more companies are building on really a Chinese dominant supply chain that has implications for America that right I think will also weaken our ability to change the way the technology goes and just these past two weeks something a couple very strange things happened in America. First, when Anthropic released its latest model, the Claude 5 model, it put in place restrictions to stop Claude from helping AI researchers in ways that it thought could support Anthropic's competitors. And then like a week or so later, a few days later, the US government imposed export control restrictions on the Claude model causing Anthropic to shut down access all around the world to all customers. But what this showed to the whole world is that private American companies such as Anthropic can shut down can restrict how others can use frontier AI models such as no helping competitors. It also showed the whole world that America can impose export controls to yank access to AI technology. And so what I've seen just over the last week is this has accelerated in many capitals around the world the urgency with which they feel like they need to secure their own supply of AI that the US government and private American companies do not have the ability to turn off at a given moment's notice frankly. And this is also causing many nations to look at open weight models because once they have all the numbers, no one can take it away from them. And I think this I think will probably have a unfortunate effect on America's soft power.
See, I just assumed that happened because Trump was annoyed at Anthropic and wanted to get back at them. But but to be clear, you're a proponent of open weights.
It turns out that if AI ends up concentrated in the hands of a very small number of gatekeepers, probably everyone here will be worse off. I mean, but the the very small number of gatekeepers would be better off because they charge a very high toll, but almost everyone would be worse off. And maybe I sometimes make an analogy to the mobile phone ecosystem. Today, there are two gatekeepers, Android and iOS. And if there are certain innovations they don't want anyone to really try out, then no one's allowed to try them out. But it turns out to be really profitable to be one or two of the gatekeepers. So, I think it would be unfortunate if AI as valuable as is ends up with a handful of gatekeepers. And to be clear, a lot of friends that work at these AI companies, I hope they do really well. I hope they have successful IPOs, you know, do very well financially. I think all that is great. At the same time, I think there is a path for these businesses to do really well without them becoming gatekeepers in a way that stifles the innovation of everyone else that would like to do things on top of this technology.
So, is there anyone aside from you in America that is is leading the charge on the open weight approach?
So, it's a tough one. Nvidia has turned out to be a major producer of open-weight models. Their models aren't quite state-of-the-art, a little bit behind, but Nvidia selling GPUs has a financial incentive to do this. Meta used to be publishing open weight models but recently switched directions to closed weight models and then there's a number of kind of you know small nonprofits academic university labs working on open weight models but because of the resources needed to build AI models has been challenging to get that going and the current White House administration has been reluctant for various reasons some of which I think are sound but we can debate them to put been you know enough resources to fund an open alternative to proprietary models. Some of the Middle Eastern countries I think you know UAE has had some signed open weight initiatives. India number of nations are trying. So I think there's actually a lot of interest in this but how well it plays out I think I don't know remains to be seen.
So China might save us in this regard. Okay switching gears for a second. I have a couple more questions. It's just having China control the open weight supply chain would would be I don't know it created a very different world than the one
No I No I know I'm being being facetious.
I know I know I know. One of the things that I've been thinking about is sort of what I call the Plato's cave analogy. So I don't know, you know, in terms of the way that the models are currently trained, they're effectively trained on data that's been digitized, right? Whether that be text or pictures or, you know, videos or whatnot, sound files, and you know, there's a massive amount of humanities works that have yet to be digitized. Obviously a disproportionate amount of western or possibly you know Chinese works have been digitized but there's a ton that hasn't been digitized and we all know also that the internet is kind of a distorted model of our world right? So one of the things that I've thought about and been worrying about and you said a couple of reassuring things about this is that is that the models are being trained on an inaccurate depiction of our reality, a distorted depiction of our reality. And so, you know, if you if you just go by the internet, everything's brain rot and every woman is like a 36D and every man is, you know, is a silver fox and, you know, all of the all of the AI ads that we're fed, you know, when we're on CNN or any other site. So to what extent is that a genuine concern that the models are are basically trained on a distorted view of our reality that it doesn't accurately depict our reality?
