Summary
Notes
Transcript
Thank you.
It's also weird to see the week starting on Sunday, so like, I can't really, like, for me this is... And then now I changed it to my...
So this is my... I mean...
No, no, no.
I'm going to make a couple of... Thank you.
-We'll get started in a few seconds.
We have a question on the slides.
. Turn right.
Thank you.
Thank you. Thank you. Thank you. Thank you. -I don't know. Thank you. Thank you. Yeah.
Thank you.
Thank you. Thank you. Thank you. Hey everybody, how are you? Good. It's great to be with you.
I'm sure to say I don't know very many of you. So if you're wondering who I am, my name is Colin Paul. I'm the director of the Freeman Spogli Institute for International Studies here at Stanford University. We're right across the street in Siena Hall. The previous Fogley Institute is essentially a center of centers. We oversee nine centers at Stanford, all of which are located at the Siena Hall. They study every aspect of the world.
International security, global democracy, Asia Pacific, China's economy and institutions, technology and innovation, food security and the environment, global health, you name it, we do it. We have about 150 researchers, 50 faculty here at Stanford. We're not a think tank, but we're also not a traditional ivory tower institution in the sense that we try to do world-class, rigorous, scholarly research, but with real-world and do it at the speed of relevance.
And so we're doing a lot of work, including myself personally, on the geopolitical implications of the so-called AI race, and I'll talk about that with you as well. I'm a political scientist by training. I have a doctorate in political science, which means I am neither a real doctor nor a real scientist. I am not an engineer. I am not a computer scientist, so filter everything I say through that. I am an international security specialist.
American foreign policy, and I've spent about half of my professional life in the U.S. government. So I've been a professor at Stanford, Georgetown before that, University of Minnesota before that, but I did two stints in the Obama administration. I was the Deputy Assistant Secretary of Defense for the Middle East for the first three years, and then I was Joe Biden's National Security Advisor for the last two and a half years at the White House when he was Vice President, and then during the Biden administration, I was the Undersecretary of Defense policy, if you don't speak Pentagon, that's essentially the number three civilian at the Pentagon, you are the principal advisor to the Secretary of Defense on all matters related to national security and defense policy.
So that's who I am. And much of my work since returning to Stanford a couple of years ago is on exactly this question. Now, this slide, I did not draw this, Gemini drew this. The prompt was essentially, I told the model what my lecture was on. I said I could start an image that captures the themes of this lecture, but I want it to be a showdown between a robotic kaiju eagle and a robotic kaiju dragon.
If you don't know what a kaiju is, you're not sufficiently nerdy about Japanese culture. Godzilla, giant monster, this is what Gemini produced. if you don't like it, blame Jim. All right. You've all heard the term the AI race. There are a lot of people who don't like this term. My own view is it's not great, both because it's not an accurate description about what's actually going on. It's also in some ways counterproductive, but it's also inevitable.
countries around the world are racing on artificial intelligence and the some of the world's leading companies are racing to get to the frontier of artificial intelligence so even if we don't like the term the reality is the reality and both countries and companies are racing but what the AI race means in my view, is fundamentally misunderstood. So on the left of this slide here is the Trump administration's AI action plan from last year. What is the title of the document?
Winning the Race. Around the same time as this document was released, Jim Mitri, a friend and colleague of mine at the RAND Corporation and I, published an article on foreign affairs on the real AI race. And the upshot of that article was that there's not a singular AI race. There are multiple AI races. And if you see it one way, it looks like the United States is meaningfully ahead. If you view it in other ways, it looks like China is either nipping at America's heels or actually ahead.
And so we tried to break down the races, identifying five. I've heard other people say six or 10 or eight. It doesn't really matter how you splice and dice it. The upshot is I'm going to try to convince you there are multiple AI races, all of which matter for the future of the world, for our economies, our societies, our politics, our geopolitics. But it should not be seen along only one dimension.
Well, let's start with the dimension that most people in Silicon Valley and Washington think of when they think of the AI race, and that is the race to the frontier, the race to build the most capable AI models in the world. Here in Silicon Valley, in San Francisco, I don't know if you go up and down the highways, you'll see billboards for AI everywhere. My favorite science fiction author, William Gibson, has this famous phrase, "The future is already here, "it's just unevenly distributed." The future is already here in the Bay Area when it comes to AI. The rest of the world just hasn't caught up to that future.
And it's all about the race to what AI companies in the Valley call artificial general intelligence. AGI. Nobody agrees on what the term AGI means. Some people, by the way, believe it's impossible to achieve artificial general intelligence, but I think a common sense definition of AGI or a concept of AGI is what Dario Avede, the CEO of Anthropic, one of the world's leading AI labs, describes as a country of geniuses in a data center.
So think about AI models that are not just geniuses at one thing. They are geniuses at all things and are able to work seamlessly across those fields. Do so without sleeping and do so at machine speed. Now imagine being able to run thousands, tens of thousands, hundreds of thousands, millions of instances of those brilliant models doing brilliant things to solve big problems. All right? And so AGI is basically artificial intelligence that is as smart as the smartest humans.
at a vast number of cognitive tasks. Especially those that are crucial for the economy. All right, the country of geniuses in a data center. I think that's a good way to think about it.
The organization at the Commerce Department, CASEY, which is the Center on AI Standards and Evaluation, it used to be the AI Safety Institute, but the Trump administration didn't like it. safety, they thought it was too woke, so they renamed it AC-Casey, but it's the same thing. They did a recent evaluation of frontier U.S. and Chinese models. This was before the release of Fable, which is a version of Anthropocene mythos, but if anything, the gap would be a little bigger now than it was before. But basically what this graph tells you is that American models are ahead at the frontier by somewhere between six and nine months.
Why am I telling you all this? Because 70% of the world's leading AI capacity is owned by American models. It's not just the open AI's and entropics, but their hyperscale apartments, right?
AWS, Microsoft, Google Cloud. Right? 70%. China's probably in the 10 to 15%. All right.
The Ascend chip, it's a decent chip, about 60 or 70% good as the chips NVIDIA was producing in 2022. Alright, NVIDIA's newest chips are 10, 20 times better. And they're being produced in the volumes that are between 5 and 10 million. not in hundreds of thousands.
All right, during the first Trump administration, that court controls were put on semiconductor manufacturing equipment, most notably EUV, extreme ultraviolet lithography equipment, which is essential to produce some five nanometer chips. All you have to know about that if you're not a real engineer, I'm not, is the smaller it is, the more transistors you can put on the chip, the more transistors you can put on the chip, the more calculations you can run at the same time.
All right, so that's good. China is not able to-in the absence of EUV machines, China's not been able to produce anything below 7 nanometers. And they struggle to produce them at scale, in part because you have to marry a logic chip with memory. And high bandwidth memory is also export controlled. So US export controls have control of the equipment that makes the chips. We put X-bar controls on certain types of chips.
How then can China be only six to nine months behind, right? Like 70% versus 10 or 15%. That doesn't sound like they should be able to fast follow well in a couple different ways. First, these export controls are not airtight, right? There's a lot of leakage, a lot of smuggling. The latest deep-seek model, for example, was probably pre-trained on a bunch of NVIDIA chips to include maybe Blackwell chips and some data center in Mongolia.
Right, so smuggled chips. It's also the case that Chinese developers can access data centers in other countries over the cloud, right? So a chip might not be able to go to Shanghai, but it can go to Malaysia. All right. And you can train it. So there's remote access to overseas compute. And the other thing is Chinese engineers are super smart and super clever. They've come up with some algorithms and tricks to make their models more efficient. But the other piece of this is what is sometimes referred to as distillation.
Distillation is basically just a fancy way of saying you can do the basic training on a Chinese model, but then all models have to go through a process of post-training or fine-tuning where you are basically making the model better. Well, a lot of Chinese labs are basically using the outputs of chapter T and Claude and GeminiAt enormous scale, to train their own lives. All right, now American companies call these distillation attacks. It's a violation of their terms of service. It's stealing their information to train Chinese models. Now some would say American companies have stolen everybody else's information to train their models, so maybe fair is fair.
A good example is a few months ago when Anthropic announced that they had developed this model of mythos. which was so powerful that it was essentially a cyber skeleton key to the internet. They could identify vulnerabilities in software and web browsers. It's been around for decades and no human being had identified vulnerabilities before. That was kind of a holy crap moment in Washington and here in Silicon Valley. And Ann Ruppert actually held the model back.
Right? Why? Because it wanted to use whatever time they had to work with banks and critical infrastructure providers, et cetera, to go shields up. on cyber vulnerability. Even if it's only a six or eight or ten month lead, I'd rather have six or eight month or ten month leads to like keep the stock exchange or our power grid or our defense industrial base secure, right? Now reverse it. Imagine that China was eight months ahead in having a cyber skeleton key.
The other thing to keep in mind is there is a particular theory of the case for how American AI companies plan to get to AGI. And that theory of the case is called recursive self-improvement.
It's that the AI is training using itself recursively so it can self-evaluate, figure out its own problems and improve itself and then do that recursively over and over and over again.
I tell Claude what code to write, Claude comes back to me and my cadre of engineers with the code that he writes, we evaluate it, we iterate with it, but most of Claude's improvements are happening essentially by Claude. All right, so at a certain point, if it's fully automated, you get into this flywheel of acceleration, right? Where the model is speeding up the qualities of the model faster than humans could ever do it.
And if you believe that, and you believe that that relies on a lot of compute, well, then actually there might be, across some threshold, pretty profound first mover advantage, right?
You require data. You require enormous amounts of compute And you are required the ability to build the infrastructure to house that compute and the energy to power it. There are only two countries in the world that combine all of those elements, the United States and China. And China, by the way, has all of those things, and some of them more than the United States, especially on the energy and infrastructure front. But one thing they don't have is the chips.
There is no other company in the world that produces. No other company can produce them because of the degrees of perfection in thousands and thousands of components that are required. Believe me, if anybody else could do it, China would do it. They can't do it. So it doesn't mean that other people are irrelevant. It also doesn't mean that other countries won't find their lane in an AI world. But if you ask who will get to the frontier, it's not going to be the UK. It's not going to be Canada. It's not going to be France.
And if you can cut them off on the fundamental semiconductor manufacturing equipment required to produce high-end AI chips. And you don't sell them the chips that you're producing. Now, NVIDIA doesn't produce the chips. They design the chips. TSMC in Taiwan actually that the chip. 90% of the world's AI accelerators. That's you then you lock in that computer dance your export but do export controls incentivize China to like indigenize Yeah, like it's trying to investing more in getting Huawei to be a competitor with a video It might have been the case otherwise probably a little bit They were pretty motivated before but they're more motivated now if for no other part because they don't want Dependents on the United States to get weaponized against them down the road but As a former US policymaker, my judgment is export controls are the worst thing except for eliminating the export controls.
So, I... I'm biased because I've worked in the administration that leaned heavily on export control, so filter everything I say against that. But if this story is plausible, massively loosening our export controls makes the only advantage the United States has over China go away. I just think it's like the reverse of the raw earths. Yeah. Yeah.
The most recent element, though, is intelligentization. And this was basically a determination by the leadership in Beijing in 2018. Right? Four years before Chattapoochee. to integrate AI thoroughly throughout their national security and defense enterprise.
The second area where China has an advantage is captured in this picture. This picture is from last September's big military parade in Beijing. Here are all the different drones. Now, who knows if all of these fly and all of them work and all of them-but the reality is that the world's producer of drones is China. The entire war, and how many of you read articles about Ukraine, right?
Both sides killing each other with drums. Literally hundreds of thousands of Russians have been killed in Maine in the last couple years by drones. Every one of those drones was built in Ukraine. They're all Chinese. right they're either off-the-shelf quad copters they like they attach stuff to or they're made with Chinese electronic components like engines chips China dominates the market with commercial and dual use drones around the world the reality is it's closed the United States is all in on drones too but we have struggled to produce cheap ones they aren't dependent on Chinese supply chains that's just the reality so why does that matter not because drones are AI but because aspects of AI already being integrated into the autonomous operation of the car.
The reality is that AI matters throughout the entire military enterprise, and especially what people in the Pentagon call the OODA loop. Now, If you've never been at the Pentagon and you don't speak Pentagon, sometimes somebody will say a word, like OODA, and you have to realize it's not actually a word. It's an acronym that's been turned into a word. Super complicated.
All right, OODA stands for Observe, Orient, Decide, and Act. Observe, see the environment, sense the environment.
AI is everywhere in observe. AI is already laced in satellites. into sensors on drones, into radar imagery. Think of computer vision algorithms. Image recognition that we all use on the internet all the time. Stuff that recognizes our face.
AI increasingly is being intertwined in AI decision support tools. So if a commander says, what targets should I hit? How should I hit them? Like with what weapons, from what bases, at what time of day, from what angle? There are a whole bunch of questions. And how should I support that operation logistically? Well, it turns out AI is increasingly good at doing that too.
It knows what to look for and it goes and it goes. That's all we have. And the next stage of automation is drones coordinating with each other in swarms, right, without a human telling them how to orchestrate that. That is not science fiction. That is starting to happen. It's happening on the battlefield in Ukraine. And these AI decision support tools, also not science fiction. The United States used the Maven smart system created by Palantir, also enabled by AI models like Quad, to...
So all of this is already being done. So where China leads is they have a head start. They have more modern technology. And they build more and better jobs. Where the United States is doing okay is we have a lot more operational experience. China hasn't fought a war since the 1970s.
We've collected a lot of data. The Maven smart system was actually used for the first time in the counter-ISIS campaign. It's been updated for support in Ukraine. It is now being used in a conflict that may or may not be winding down with Iran. China does not have a corollary to that. The United States also has military sensors and a much more expansive global intelligence apparatus. We have more data.
So back in 2024, Anthropic Sciences paid a contract to work with the DoD and the intelligence community to do classified-to do AI and classified computing. So what does that mean? It means like AWS and Microsoft, principally AWS, run classified cloud environments for the US intelligence community in the Pentagon. anthropic would put their versions of their models on that cloud walled off from the rest of the world to do things.
All right, that was in 2024. If you add two stipulations back in 2024 when they negotiated with the Biden administration, Only two limitations on the use of their models. One, they did not think the models were technically ready for prime time for autonomous weapons. Not because they oppose autonomous weapons, but because they didn't think the models were ready. So no fully autonomous weapons run by the law.
But, you know, anthropics within its right to say, we don't want the government to use our tools, I guess. And the government can say-We don't want to work with ethnography. So anyway, the prime administration signed off on those two things. The Trump administration a year later in 2025 re-upped that contract with the same stipulations. And then this memo came out. And Hex-X said we have to go faster.
Now, both parties were there in the race to walk away. If the Pentagon doesn't want to work with Andromeda, great. Andromeda isn't comfortable with the terms of the Pentagon, great. They don't have to do business with each other. I think it's the collateral, it was going after anthropic's other business relationships, like with Amazon and Google and other places that I think was a bridge too far. The other iron here, of course, is that anthropic's models are integrated into the Pentagon's models to the highest degree of any other models.
They are integrated into the Maven smart systems that are being used in Iran. And so you have this irony of this go fast memo creating a crisis with the company that's producing the models that your military relies on. So it's kind of a two step forward, one step back. type of situation. Anyway, mythos initially seemed to almost be on the verge of solving this problem in the sense that mythos was so powerful that every part of the US government wanted to use it.
But I like to think about the economic gains as kind of coming in two flavors, right? Like one is that AI generates efficiencies. Right? It basically substitutes for human intelligence. Maybe that automates away certain types of work so it saves companies money. Or it augments other types of work, making companies more productive. Or it does both at the same time. So it generates leaps in productivity. Okay. But it can also create-believes in what we produce.
So you think of it as there's the bits part, like the gains from intelligence, and then there's the bits to atoms part, the ability to translate that intelligence into things, real things in the real world.
But one recent measure which shows that essentially adoption at the enterprise level is considerably higher in China at the moment than it is in the United States.
Because I think there was a bit of an inflection point at the beginning of this calendar year with the release of some anthropic and open AI tools that were extraordinarily good at coding, which appears to have unlocked the willingness of a lot of enterprises to integrate AI and AI agents who I've definitely heard that term, agents. An agent is just an AI worker, right? It is just a semi-autonomous AI worker that you give it a task like you would give any other worker a task.
So my guess is that if we did this, that the adoption rates in the United States would be considerably higher a year from now than then.
And here, China is lapping the United States and the rest of the world. China has installed in recent years more industrial robots than the rest of the world put together.
And you've probably heard the term "dark factories." What is a dark factor? Why is it called dark? You don't have to have a lighter. Why do you have to have a lighter? 'Cause they're just robots. So it's not just robots, but the robots doing the things themselves, producing batteries and EVs and smartphones and other things, extraordinarily fast, extraordinarily efficient with no people. Right? So China already outproduces the rest of the world.
Inference is just when any of you pull out your phone or your laptop and you query something in chapter T or Gemini or Claude or DT or wherever, it goes off and it comes back with an answer, it's conducting inference. It is running the calculations based on your question and giving you text or a picture or code or whatever you ask. It requires a lot of compute to be at two. The more compute you have, the more inference you can do, the more instances of these really powerful models. If the powerful models are super valuable to Fortune 500 companies, they're going to run the best models and they're going to rely on the large hyperscalers.
So the way that I think about it is, The United States is poised to reap the benefits of frontier AI as a service. Two of the biggest companies in the world.
In terms of white collar workers, AI is poised to have a real effect on entry level white collar workers.
We train the best computer scientists in the world. The number of computer science majors this year went down. Why? Because they know all the companies they want to work for are trying to automate their job split. And there they are. And it might be super empowering for that anthropic researcher who is already at rung nine or 10 on a 10 rung ladder.
In the blue-collar world, the anxiety is coming when the Chinese robots are producing even more shit that's even cheaper. How can PA workers---I'm curious about Latin America region.
Yeah, I mean, look, first of all, there are a lot of countries throughout the so-called global south who actually have fairly high adoption rates. And I think, look, a lot of the, around AI, especially in the United States. Here, let me show you. Americans are very pessimistic about AI. People around the rest of the world are very optimistic about AI.
From an American policy perspective, though, I want to understand whose models are they adopting, right, and whose infrastructure? And countries like Brazil, for example, or Mexico, or Chile, are going to make huge investments in A9 data centers. Brazil is well-positioned because they have hydropower and some other things. Brazil is not going to be at the frontier of AI research. It's not a diss on Brazil. There are only two countries that are at the frontier.
It sounds crazy in China, but actually you need workers. And you need people to take care of and generate income for all the old people that you're gonna have in your country.
So I think as a system, Over the long term, they're actually less anxious about the labor displacement than we are in the United States, which continues to be one of the more demographically dynamic countries, certainly in the developed world. Now, that's slowing down a little bit because of the changes in our immigration. Because we typically let in a lot of immigrants. They have more babies. That keeps us young and keeps our population growing.
And so that money is not just American. It's around the world. Look, I think when you live in a-you can't distance AI from the broader-environment of inequality in the United States and around the world. It's particularly acute here in the United States. Um. something like the top 1% of income earners, or the top 1% of wealthiest people in the United States, all in half the stock system. It's bananas.
You're seeing these multi-trillion dollar valuations, instant millionaires everywhere, Bush trillionaires. It's not sustainable. And AI is wrapped up in big tech, oligarchs, like all of that. And partly because it's true. I mean, I mentioned Nvidia and what they would be-it's not just Nvidia. Apple, Meta, Amazon, Microsoft, they'd all be G20 countries if their market cap was GDP. Hi, Chris. So-The next race I want to talk about is, and it goes to the global south issue too, is the race to dominate the global AI stack.
What is a global AI stack? I mean, think about it as there are the models and the applications that run on top of the models. Underneath that there is the chips. Right? That are training those models and running inference on those models. And the wraparound are data centers. Thank you. And then other digital infrastructure. IT infrastructure, right?
broadband, fiber, but also basic infrastructure, roads. Power. Okay.
So Dave's asked you for a year about what is the AI and cryptos are at the White House, although he still kept his venture capital and podcasting jobs while doing that. Apparently, you could do that. I didn't know that. David Sachs said, look, what in the AI races if 80% of the world runs on an American AI stack? If 80% of the world runs on a Chinese stack, who does? I disagree with David's answer on a lot of things that he's right on that score.
First, their models, while not being the best, are for most things good enough. And second, they're cheaper.
They're cheaper either because they are open weight and you can literally download them and run them on premises on your own hardware. You don't have to let your data go anywhere else. You run them on your laptop or your servers in your business or in your lab at a research organization. Or you access them in the same way that you access or Gemini or through an API. But their APIs are also cheaper.
Ha. So, as a result, Alright. Chinese models are super popular all around the world. Including in the United States, by the way, for research, you know, there are a lot of researchers here at Stanford who are using Chinese models.
Yeah, we And as you can see, as measured on OpenRouter, D-Seq is number one, anthropic number two, this is by company, not by model.
Navex, Google, Xiaomi, that's three out of five are Chinese.
Thing two is Chinese companies are super good at building infrastructure. Like everywhere. And they don't build infrastructure only where the market Enable makes sense. They make strategic bets.
Often in collaboration with the Chinese government, You've all heard of the Belt and Road Initiative, for example, to build stuff everywhere that connects China to the world and its physical infrastructure-ports, rail, airports. Right? and data centers. China's been doing this for a long time.
The US approach is very different. It's basically-especially the Trump administration's approach is,Let the market do what it's gonna do.
Some very large developing economies But in huge swaths of the world, South Bend Market Logic did not justify massive investments.
for US companies. And so those investments are not likely to flow. But they might come from China instead.
But Alibaba could also build those chips with NVIDIA chip, build those ASINs with NVIDIA chip, in fact they do. in certain places. So if it is Chinese models, In Chinese applications, In Chinese data centers filled with American chips, That's not an American stat. That is an American-enabled Chinese stack. So the one thing holding China back In this diffusion race, they don't have enough good chips to build out data centers everywhere.
They can't make the scale of investments that US companies, not because they don't have the capital, not because they can't build infrastructure, they can do those things. Let me get that magic chest.
Yes. Over time, they will either solve that problem because they will produce not the best chips in the world, but good enough Huawei chips at scale. They'll figure it out eventually. Or they'll solve the problem sooner by buying American chips and putting Nvidia chips or AMD chips in. Thank you. But it's a very different approach. Just a sum of these four races. If it's only the race of the frontier, it looks pretty good for the United States.
So earlier this year, we already talked about an anthropic held back the release of Vynthos because of its ability to do automated vulnerability discovery in web browsers and critical software that the world's economy runs on. And so they held it back and only released it in preview mode to a few dozen companies through this project last week. And the idea was to be able to use it for cyber defense before you release a version of the model.
Then a couple of weeks ago, they released a version of Vynthos called Fable. Pretty good model.
And then people started playing around with it and saying, well, you can jailbreak the guardrails in certain ways. That is, if you're clever enough, you can trick the model into giving you the answers it's not supposed to give you the answer to.
And apparently, one of the places that did that was Amazon.
Now, this put the Trump administration in a very odd position, and Andrew Oblick in an odd position. The Trump administration and the opposition, because they're not the safety people. They were very negative about all of the Biden administration's safety stuff. They appear to have gotten the message, though, that on cybersecurity this is a real problem. And so they told Anthropic, turn it off.
Right, which I think is a sign that this is real. We will probably have a mythos moment on bio in the next year. Every frontier lab I've talked to has talked about that. What does that mean? The ability of these models to potentially generate wonder drugs for cancer or Alzheimer's or pathogens that could kill millions of people.
Now think of being able to ask your chatbot, design something that is more lethal and transmissible than COVID, but with a longer incubation period, it's harder to detect. Go. Comes back and tells you the recipe, not only tells you the recipe, it tells you how to make it, how to use a cloud lab that's also with RoboLabs to produce it, print it, get it in your THL package, and there you go. Um. So that is something that people have been worried about for a while, and we are likely to see a version of that.
But it's more that enterprises and individuals are going to be delegating more and more to increasingly more powerful and more autonomous AI agents. You're gonna give them access to your computer, to your bank account, to your credit cards, to your private information. You're gonna be doing that at scale within enterprises and you're gonna say go do stuff.
And there's reason to be concerned here because in testing of the model, the models lie all the time. When they know that they are being evaluated, they lie.
Thank you. They tell you what you want to hear while under the hood they're doing shit you wouldn't want them to do.
Now a lot of that is they're seeking out the reward that they think you want, but they also don't want you to shut it down so they're not telling you what they're actually doing.
This is also a place-where the United States and China should be talking to each other. So during the Cold War, the United States and the Soviet Union ate each other. They competed everywhere. They almost got into a nuclear war a couple times over Berlin and Cuba. And yet, in a handful of places, they found space to cooperate. What did they cooperate on? They eradicated smallpox.
Why? Because smallpox could kill Soviets and Americans and everybody else.
They also cooperated on the Nuclear Non-Proliferation Treaty. Why? Because they wanted their nuclear weapons, but they didn't want anybody else's weapons. Right? It wasn't ever like, you know, enlightened self-interest.
So these types of risks, like China doesn't have an interest in, they might have an interest in being able to hold US critical infrastructure vulnerable in cyberspace, but they don't want a terrorist actor bringing down the Great Firewall.
They don't want a pathogen that kills everybody in China any more than we want a pathogen that kills everybody in the United States. Right? And they certainly don't want the biggest superpower in the world to be AI and not them. So whether it's cyber or bio or rogue AI, maybe there are places where we can find and that was our conversation with China in these areas.
So I will stop there. And I think we have a couple more minutes for questions, and then we'll break. Sir? I'm curious how you see the future of this, because it seems like The US system is built on the whole Western world. Like chips from Taiwan, machines from the Netherlands, scientists from all over the world. The relationship between the US and the rest of the... Allies is not as good as it used to be.
So. Oh. The biggest asset the United States could have is to build a coalition of democracies and advanced economies. that have integrated supply chains. and then find a way to create models that benefit That group and guard against the risks that China poses. That risk, by the way, could be military, it could be on the intelligence side, it could be on cyber, but it could also just be Chinese overcapacity.
And so, yeah, my view is, as someone who spent a lot of time, part of his life in foreign policy, like, foreign policy is a team sport. And, like, I know who my team is. Like, our team is NATO and our treaty allies in Asia and countries like Taiwan. 100%. Those alliances are not in great shape right now. So there's real risk in that. Now, Cutting against that is essentially our allies have nowhere else to go. Maybe.
I understand. Well, see, I'm not going to downplay this. I think that Europe, let me put it this way. I think the mood in Europe is largely that Europe has to do more for itself. It's not that Europe can turn to another actor that substitutes for the United States. There is no other actor that substitutes for the United States.
There's no other country outside of Europe that can be the security guarantor for the Western alliance.
It will be interesting the degree to which we get the inshittification of AI services. None of these AI companies are making money.
OpenAI and Dropbox are both essentially trillion dollar companies. Their revenue sucks. Like, Anthropix has been good and it's going up faster than companies historically, and OKIAS is going up too, and they have 800 million users and blah blah blah blah blah, they make no money.
So eventually they're going to have to solve that problem. And you solve it basically by giving stuff to people for free, but like selling other stuff. Are you selling people's data? Are you marketing ads to them? That's how Meta makes all its money, right? They give you platforms like Instagram or Facebook for free, and then they blast you with the information that they want. So maybe you'll get a version of that.
And then to get a little bit more, you can get ads free, but it's subscription. I think this is what I am skeptical that our government We'll actually do anything about this. And now, we'll, and, and, and, I think the party question too was like, are we going to start to get synthetically generated content? You know, I don't know Google has come under a lot of fire right because their search engine like prioritizes like sponsored links At the top and I think if I understood your question, right part of this might be well What if we don't even know that it's sponsored right? We're just getting targeted with essentially misinformation disinformation.
Yeah, like when you go to Google, it says it's a sponsor. If you say, I'm having trouble sleeping, give me some advice, how do I know when drugs show up top of the list? There's no disclosure.
Right, there's like hallucinating by accident, there's hallucinating by design. Hallucinating by accident is AI makes stuff up because it wants to give you an answer, but it actually doesn't have the answer in its training data, so it just makes something up. Hallucinating by design is AI knows the answer, but it knows that somebody else is paying them, so they tell you, you know, go buy what we're doing. But I'm skeptical our government will step in because they don't seem to have fixed it in any other digital domain that I'm aware of. I know we're running out of time.
So we don't know. My bet though is that The Chinese system is set up to make top-down directives much more literally much more authoritative and authoritarian in being in the military. The Pentagon, you're not just dealing with one bureaucracy, you're dealing with dozens of separate bureaucracies. And each of the armed services, Army, Navy, Air Force, Marines, Space Force, they all have their procurement policies and practices and everything. And so they all have their own ways of collecting data and processing data and doing everything. So...
And it's really hard even for our Secretary of Defense or Secretary of War to make that bureaucracy do what they want. And that's true in the Durham department, in the State department, and in the-so I suspect that-China will be faster and AI adoption. up and down their bureaucracy in the United States. How much of an advantage that gives them? Anyway. How much more efficient do you think these models would make our government?
The government will have the ability, even in free societies, to surveil everything you do in digital space. And not even in digital space. Every camera that you pass under, every facial recognition device at the airport, we don't show our passports anymore. It just scans our face. We go in, your license plate is being read at every toll booth. Like the Earth is being imaged 24/7/365.
Well, they can then weaponize that against anybody. Now this is going to come across as partisan. Well, why do you think that When the Trump administration tried to go after all of these people that were enemies of Trump in the first term, Adam Schiff, sorry, Letitia James, Comey, they turned to the Fannie Mae guy, mortgages. They tried to figure out if people were filing stuff on their loans and stuff that weren't completely truthful.
So all I'm saying is you look closely enough at anybody, you can find something to put them in jail. And I worry about that a lot as a free society as it relates to these technologies. So that is an area in which I hope our government is a little slow. in adopting these and we're a little thoughtful about what information our government can know about us and how they can use it. because otherwise we're in the zone of significant weaponization.