Your Undivided Attention - Chips Are the Future of AI. They’re Also Incredibly Vulnerable. With Chris Miller

Episode Date: March 29, 2024

Beneath the race to train and release more powerful AI models lies another race: a race by companies and nation-states to secure the hardware to make sure they win AI supremacy. Correction: The latest available Nvidia chip is the Hopper H100 GPU, which has 80 billion transistors. Since the first commercially available chip had four transistors, the Hopper actually has 20 billion times that number. Nvidia recently announced the Blackwell, which boasts 208 billion transistors - but it won’t ship until later this year.RECOMMENDED MEDIA Chip War: The Fight For the World’s Most Critical Technology by Chris MillerTo make sense of the current state of politics, economics, and technology, we must first understand the vital role played by chipsGordon Moore Biography & FactsGordon Moore, the Intel co-founder behind Moore's Law, passed away in March of 2023AI’s most popular chipmaker Nvidia is trying to use AI to design chips fasterNvidia's GPUs are in high demand - and the company is using AI to accelerate chip productionRECOMMENDED YUA EPISODESFuture-proofing Democracy In the Age of AI with Audrey TangHow Will AI Affect the 2024 Elections? with Renee DiResta and Carl MillerThe AI ‘Race’: China vs. the US with Jeffrey Ding and Karen HaoProtecting Our Freedom of Thought with Nita FarahanyYour Undivided Attention is produced by the Center for Humane Technology. Follow us on Twitter: @HumaneTech_  

Hey everyone, it's Aza. So in 2017, something really miraculous happened, something that has never before happened in the history of humanity. And that is, humanity learned how to make machines think. To turn chips or computers, not into just calculators, but into thinkers, from chips to cognition. And this is a big deal because that means chips and compute is about to become one of the most valuable commodities in the world. Actually, it will likely become the most valuable commodity in the world.

And it's a commodity unlike any other commodity because it is intelligence. And so no matter how much of it you have, well, intelligence can figure out how to use more intelligence. There is no upper bound. And I cannot overstate how much the U.S. company, Nvidia, dominates the chip race, and how heavily the whole world relies on it. It controls 80% of the entire GPU market, but then all of those chips are made by a Taiwanese company called TSM. Now, other companies are coming for it, or at least trying. Intel is racing to launch a new AI chip this year.

Meta is planning on using their own custom chips, as is Google, as is Amazon. And meanwhile, the U.S. is racing to onshore the manufacturing of its own chips. We don't want to be dependent on other nations, especially ones that are so physically close to China. The Biden administration just committed up to $8.5 billion under the Chips and Science Act so that Intel can build its own brand-new chip-making centers across the U.S. And all of this is for this tiny little object known as the GPU. And we're going to talk about all of this today with Chris Miller. He is an economic historian.

His day job is the assistant professor of international history at the Fletcher School of Law and Diplomacy at Tufts University. And he is the author of the best-selling book, Chip War, the fight for the world's most critical technology, which I think made it on to the 2023 must-read. lists of foreign affairs magazine, Bill Gates, and Barack Obama. So, Chris, welcome to the show. Well, thank you for having me. Let's start by doing some table setting. People may have heard chips, GPUs, microprocessors, semiconductors. We use these terms sort of interchangeably.

Walk us through what compute is when people say compute and give us a little bit of its history. Well, a chip, as it's probably most commonly known as a piece of silicon in most cases, that has lots of tiny microscopic circuits carved into it. Here is a modern 1966 version of integrated circuits with many hundreds of components on this one circuit. This particular function provides 16 bits of digital memory in this one package. In the early days of the industry, when the first chips were invented in the late 50s and early 60s, a chip would often have a handful of components on it, little switches called transistors that flip circuits on and off. But today, the most advanced ships can have tens of billions of

these little transistors. And as a result, they're tens of billions of times more powerful than they used to be. And for any type of computing that happens in the world today, it's represented by these chips. All the ones and zeros undergirding all software, data storage, are just little circuits flipping out and off on a piece of silicon. Somebody was recently telling me, I don't know if this is right, that the width of the lines etched for when you make these chips are like three nanometers, which is the distance your fingernail grows while I say the sentence.

Like, give us a scale, like, why is it so hard, right? If it's so valuable, like, are these things hard to make? Like, why aren't more people making them? Well, if you take a chip inside of a new smartphone, for example, open up a phone and look at the processor inside, what you'll find is a chip that has roughly 10 or 20 billion tiny transistors carved into it and sort of fit 10 or 20 billion of these devices

into a chip that's roughly the size of a fingernail. Each one of them has to be tiny, roughly half the size of a coronavirus to give you a sense of this scale. And so manufacturing coronavirus-sized devices by the billions is the hardest manufacturing that humans have ever undertaken. It's more complex than anything else that we make.

But in the chip industry, the manufacturing is so complex, so R&D intensive, so costly, that there's just a couple of companies that can produce chips at the cutting edge. Got it.

So walk me through that ecosystem. Like, who can produce chips? What do they depend on? How fragile is it? I just want to understand a little bit of the landscape for everyone. Well, so to talk about the chip industry, you've got to divide

into different types of chips. And so chips do lots of different things. If you take your phone, for example, there's a chip that connects to the Bluetooth, the chip that connects to the Wi-Fi. a chip that manages the camera, but the most important chips are processor chips, chips that run operating systems on your phone

or that train AI systems and data centers. And if you look at the ecosystem of companies that produce advanced processor chips, there's really just a couple in the world. In terms of manufacturing, almost all of the most cutting-edge processor chips are produced by one company, Taiwan's TSMC,

the Taiwan Semiconductor Manufacturing Company, which manufactures almost all GPUs, and a huge share of the advanced processors that go in computers and in smartphones. But TSM actually doesn't design any of the chips. They manufacture chips for other companies. And so whether it's NVIDIA or Apple or Qualcomm or AMD, the world's largest chip design companies

generally rely on TSM to manufacture their chips. So what you're saying is that many people can come up with the designs for chips, but it is TSM that ends up actually making the chips. So they're like they're the final bottleneck. That bottleneck exists in as you said Taiwan, which is very interesting from the geopolitics. I think everyone is familiar with during COVID suddenly cars became hard to get and cell phones became more expensive because chips suddenly weren't able to get to us and became aware of how in some sense fragile or dependent we are. How did that come to be?

That seems very surprising that the most important commodity is essentially controlled by just one company. Well, in the chip industry, the economics have been defined by economies of scale. The more chips you produce, the lower your unit costs fall, and even more importantly of that, the more rapidly you improve technologically. Because for every chip you produce, you gather data about the production process, you tweak your manufacturing as a result, and you get better chips in the end. And so TSM is the world's biggest chipmaker. And as a result, it's also the world's most capable, most technologically advanced chipmaker

when it comes to producing these processor chips because it's learning more, simply because it produces more. And so that has led it to acquire a market position that is extraordinarily hard to displace. And so other companies that are trying to compete, like Samsung or like Intel, right now have much smaller scale than TSM.

And so they struggle to get the technological advancements that TSM can gather from their vast production, and they struggle to get the low cost. I would assume that every major country, China and the U.S., would be trying as hard as they can to undo that bottleneck. Could you walk through what's going on there? You know, U.S. I think, just invested how many billions of dollars into Intel to try to make this happen? And, like, again, like, if there's that much money flowing in, why is it so hard? Well, you're right that many countries have looked at their reliance on chips made in Taiwan as a potential vulnerability. From the perspective of Beijing, China spends as much money each year importing chips as it spends importing oil. And Taiwan is one of the largest sources of those chips.

And so the Chinese government for the last decade has been in the midst of a process of trying to reduce their reliance on imported chips. But the challenge they face is that they're meaningfully behind the cutting edge. And the U.S. finds itself in a somewhat similar position. Key U.S. companies like Apple and NVIDIA rely exclusively on TSM to produce some of their most important chips. And the U.S. has been via the Chips and Science Act, which was passed two years ago, trying to incentivize more chip manufacturing in the United States. And it's giving out grants to a number of companies, including to TSM, to open some new facilities in the U.S.

but the reality is that all of the impact is going to take a lot of time and it's only going to be a relatively small impact in comparison to the scale that TSM currently has. So I guess I sort of want to get it a little into the history then because it seems like if I was Taiwan it would be no accident to want to have China be dependent on me for something so core to their economy. So how did TSM come to be, how did it work

in Taiwan's strategic plans? Well, Taiwan has been a part of the electronics supply chain for now well over half a century. But for a long time, it was in the bottom rungs of value add. It was simpler assembly, for example, it was happening in Taiwan, and the more high-tech parts were happening in California or in Japan.

But around 30 years ago, the Taiwanese government decided to make a bet on an entrepreneur named Morris Chang, who had said, spent his career at Texas Instruments in the United States, but was passed over for the CEO job and was looking for something else to do. And he was approached by the Taiwanese government and asked if he wanted to start up a company in Taiwan to manufacture chips. And he had an idea that would transform the industry. And it was to split the design and manufacturing of chips

into two different parts. Before that time, almost all chips were both designed and manufactured by the same company. But he wanted to only do the manufacturing part, letting him manage. manufactured larger and larger volumes, driving down the cost and also driving up his ability to learn. And that business model innovation explains why TSM and Taiwan are today at the center of the chip industry. And then how does that fit into Taiwan's strategic plan? Well, you're right to ask because although it was a business model innovation that made TSMC so important, TSM is now critical for Taiwan as a whole. In an economic sense, TSM is by far the largest a source of Taiwan's exports and in a strategic sense, TSM, make sure that Taiwan is at the

center of discussions of technology. And so today, it's not just China. It's also United States. That's hugely dependent on Taiwan, which the Taiwanese hope will give both of those countries an incentive to keep the geopolitical relationship somewhat stable. Chris, let's get into the phenomena that describes the incredible progress that's happening in chip manufacturing. It's called Moore's Law. Tell us about it. Well, it was named after Gordon Moore, who's one of the two co-founders of Intel. And in 1965, he hypothesized that the number of transistors per chip would double every year or two over the subsequent decade, all the way to 1975. And that proved true. And it's proven true with some changes in the rate of growth,

but the exponential growth has persisted all the way up to the present. And so we get chips that are roughly twice as good every two years. And that's meant that the capability of chips has far surpassed the rate of growth in basically any other product in all of human history. Nothing comes close to exponential growth sustained over many decades. And it's made it possible to, in an economically viable way, produce larger and larger volumes of computational power, beyond the imaginations of anyone who was producing chips in the 60s or 70s. And if you look at large AI systems today, they're only viable, only possible to think about because we have access to better and better chips with more and more transistors. And so do you see any sort of like

end in sight for this race? What are the limits to the race? Well, Gordon Moore himself predicted that Moore's law is going to end at some point. Exponential growth can't go on forever. And so the question is going to be when. But we haven't gotten to that point yet. And indeed, we've actually had repeated predictions from leading experts that Moore's law is about to end all the way back to the 1980s, and it's been repeatedly wrong. He was initially focused simply on shrinking transistors and putting more of them on chips. And that's still happening. But in addition to that, we can also design ships differently. So what differentiates Nvidia from AMD?

Well, it's the way their chips are designed, and you can get more performance out of certain designs than others. So that's an additional tactic for getting more compute per chip. In addition, you can package ships together differently to have faster data interconnect, for example, between your logic and your memory. That's another way of getting performance. And so all that's to say there's a lot of different techniques for getting improved performance,

so we're not just relying solely on making transistors smaller and smaller and smaller. The second reason that Moore's Law has persisted is because there's a huge economic incentive to make it work. NVIDIA is the latest example of this. By designing chips better, Nvidia turned a type of chip that was initially used for computer graphics into something that's central to AI

and in so doing made itself a company valued over $2 trillion. Well, that will incentivize the next company to find another way to improve computing power and that will keep the Moore's Law dynamic alive. So I would be someone who would bet on Moore's Law continuing for some time just because the incentive to find a way

to keep it working is so large. Imagine sometime in the future, someone trains a model that just has incontrovertibly dangerous capabilities, whether it can understand how to make a virus or other biological weapon in novel ways, or whether it can copy its code and start self-replicating or choose whatever doomsday scenario you feel like, the red lights start blinking. It has a dangerous capability. When do we need to have put in place the safeties, how long will it have taken, like in the past, for us to then make it safe so that, in a sense, you can control anyone else starting to have access

to these incredibly dangerous capabilities. I think it's almost certainly the case that the answer to that question is not a chip-based question. In other words, the chip will do what you tell it to do. Most chips are fairly general purpose in nature. And so what that means is that if you want to put guardrails up around a system, the chip is not the level of what you're going to do it. I think that the economics are such that there are strong incentives to produce general purpose processors,

which means that we're not going to have, I don't think, in a widespread fashion, chip-based guardrails. It will be system-based guardrails that you train on a chip. The second question, though, is for deploying a system, how much computing power does it deploy? We know training is super computationally intensive, but inference will be as well, depending on what scale is happening at. And so you're going to need a lot of chips you already do, and you'll need more. for inference. And so that's the second part of the equation is depending on what type of system you're talking about, what's the computation need for inference and how easy is that to access? Just jumping in here to Orient. For AI, there are generally two portions of training and using an

AI. When you train an AI, well, that's called training. That uses a lot of computation, a lot of power. And then when you use that AI, when you ask it a question, when you ask it to do something, That's called inference, and it happens much faster, and it generally takes much less power, but you use it a lot more. I'm curious, where do you see your greatest concerns? And you've just said you don't think it's going to be on chip, but in systems. So I'm curious to expand that as well. Well, I think if you look historically, whenever a new technology emerges, that is fairly general purpose in its nature. And so it can be used for lots of different ways.

We always don't know the ways it will be used. it's very difficult to predict. And so there's extraordinary uncertainty about how do you set up the types of guardrails that you want because it's difficult to set up guard rails for something that you can't predict how it will be used. You set up guard rails for systems that you know how they'll be used.

On a road, for example, you know the road's going to turn left and so you put guardrails so a car doesn't miss the turn and fall off the cliff. And so where we are with AI right now is that it's extraordinarily difficult, basically impossible to predict all of the use cases that will be envisioned. And so I think that's why,

why the conversation about guardrails is so conceptually tricky. You just don't know where exactly to be putting them up. I think the second challenge is that, like all other general purpose technologies that we've seen historically, AI will be used for a lot of good things and a lot of bad things. And so it will be shaped not just by the technological attributes, but also by the social and political context in which it's being deployed. And so I think if you're going to think about ways in which the technology becomes

problematic. You've got to both think about the technology itself, the process of putting up guardrails around it, and then the social and political incentives about who's actually designing the guardrails. And so all that's to say, I think, you know, it's not just the question of technology. It's also a question of society. Totally. It's something we often say, and this is a quote from Charlie Munger, is that if you show me the incentives, I'll show you the outcome. And I'm sort of like ground zero for understanding this, because I'm the hapless human being that invented Infinite Scroll. And when I invented it, I invented it, I invented it. I invented to it because I'm like, oh, you reach the end of a blog.

Like, you want to see more blog posts. You reach the end of some search results. You want to see more search results to slow more in. But I was blind to the way that it would be the incentives of the race to the bottom of the brainstem for attention that would turn to use Infinite Scrolls a weapon to keep people versus helping them. And so I think the key thing with AI is not to ask is AI good or bad. It's to ask, is the incentive that will pick up and wield AI good or bad? I think the thing that people often forget about AI

is that there is no way of separating the promise and the peril. It's the exact same tech that lets you make incredible AI art and decode brain scans into what somebody is seeing. That tech is the exact same as it is to make fake child pornography. You can't disentangle those. The ones that make great biology and science tutors are the ones that make tutors that help you create bio-weapons. You cannot separate the problem.

in the peril, which makes the governance question, as you're pointing out, incredibly difficult. Yeah, I think that's right. And that's why we can't, I don't think, count on technological solutions to what are essentially social and political challenges. One of the frames that we'll often use, and we've used it on this podcast, is what oil is to physical labor, AI is to cognitive labor. That is, you know, every barrel of oil is worth roughly 25,000 hours of physical human labor time. And after we figured out how to harness oil, you could, like, take people out of the fields and replace them with, like, mechanized workers, essentially tractors. Same thing is going to happen here now with cognitive labor.

What is cognitive labor? It's like when you sit in front of your computer and have to write an email, like writing your email, choosing what words of this cognitive labor, deciding, as a scientist, what experiments you're going to run and how to write them up as papers, that's cognitive labor. And so it sets up very much another race, like an industrial race, between companies as well as countries. And I think that's why we see the U.S. doing their export controls on chips. And so I'd love for you to like, I mean, feel free to disagree or add more nuance to that analogy and then talk about its geopolitical implications. No, I think that analogy is absolutely right. And it's a race with technological and economic, but also defense and intelligence around

too. And so if you talk to people who work in defense ministries or intelligence agencies, they're also asking themselves the question, how will AI change my work? And the answer is it's going to be pretty important and pretty impactful. And so there's a race between militaries and intelligence agencies around this exact topic. And that explains why the U.S. has been trying over the past couple of years to limit China's access to high-end compute. And so starting in 2022. The U.S. imposed restrictions on sort of the highest range of GPUs. And just a reminder for people who might be lost, GPUs are built on chip. So they're a specific kind of chip surrounded by a certain kind of architecture originally made for gaming.

Then in 2023, brought the restrictions a level further down to sort of the second highest range of GPUs, making it illegal to transfer them into China. With the aim of making the cost to compute in China higher, adding inefficiency to the AI training and deployment process, and the U.S. hopes keeping a U.S. edge in AI over China. I was just going to ask, do you know how effective that's been? We've heard, to some of our friends that work with Chinese researchers, they feel like they're at least a year behind, and they're frustrated by the inability to get chips, but we've heard from other people that you can get them on the black market

and the chips get sold to some other country and then import it into China. So do we know how effective it's being? You know, it's a tricky question to answer because any sort of way around the controls is naturally not going to be recorded and publicized. It's also tricky because I hear about shortages of GPUs in China, but I hear about the same thing in Silicon Valley. And so it's difficult to know for sure. There certainly is smuggling happening. There's been well-documented instances that my sense is that on balance it is creating some challenges for Chinese firms to train at the level they like to train and driving up the cost, adding some inefficiencies.

On its own, it's not going to stop China from training any individual system because you can just train a system with a less efficient data center. It takes a bit longer, but you can do it. I think where it's going to have a bigger impact is on the amount of training, especially not by government actors, but by business actors who are going to be more responsive to the price of training relative to the government. So it's going to be a mixed and complex. picture, but I think it is having some impact right now. You studied the role of technology and the downfall of the Soviet Union, and so a question I have is, is it more important for the U.S. to stop China from developing powerful AI for its

military without necessarily developing that power ourselves, or do we have to do both? That is to say, if our goal is to beat China, and we hear that again and again in the halls of Congress, we must beat China, let alone the fact that we beat China to social media, which means we actually lost when it came to polarization and mental health and a whole bunch of other things we really care about to strengthen our society. Like what's required to, you know, quote unquote, win with China is just slowing them down enough. Well, I think the reason this is a tricky question is because nobody's really sure what the most useful application of AI will be in military contexts. Right now, militaries are experimenting. They're deploying AI and

intelligent systems, for example, to gather pictures and use computer vision to identify what's dangerous and what's not, or to make sense of signals intelligence data. It's being deployed to make systems more semi-autonomous, but no one's really sure exactly how it will be the most impactful. And what that means is that both China and the United States are betting not just on the deployment of individual AI systems. They're also hoping that their AI ecosystem will be more developed and give them a broader level of capabilities to draw on in the future. And so that's why there's both sides of this race, trying to slow China down, but also trying to race ahead, are the strategy that's being deployed right now in Washington. One of the goals for many of these AI companies is to automate science itself. That is, essentially, have AI do discovery to discover new types of materials, which if you can discover new kinds of batteries, you can discover new kinds of bombs. If you can discover new kinds of cancer drugs, you can discover new chemical and biological.

weapons. And so I'd imagine that's another thing deeply driving this race is whoever can get to an increased rate of science dominates. Yeah, I think that's right. And the perception among policymakers is that there will be meaningful productivity

improvements to the economy and also to research processes in particular, which mean that it's not just about the question of what will the impact of AI be on our economy, but also if you're deploying AI more rapidly than your competitor, your economy it grows more rapidly, you're in a better strategic position.

That is, again, another incentive to run in this race that you alluded to. Yeah. And I know I'm always very hesitant to reify talking about any kind of like war, especially war between the U.S. and China because we have nukes now. And so like, the stakes are much, much, much higher. But, you know, it reminds me of that moment when the U.S. before World War II cut off our export of oil to Japan, which, you know, I've heard some historian's site as, like, a galvanizing reason for why Japan got into World War II. And I'm curious, does China view the export

controls as some kind of beginning of an act of war? Like, how do they view that? Like, what are the dangers that you see here? Well, I think you're right that there is a danger of pushing China in the direction of thinking it's got no choice but to turn to war. You know, the context in which China to make that type of decision is if it believed it, it was in a race, but it could not win, and therefore it was better off taking a risk now rather than taking a risk in a couple years' time when it was in a worse position. I think there's some key differences, though, between today and the episode of the 1940, 41 period that you were discussing pre-World War II. One is that in Japan in 1941, had they not gone to war and had the oil export bans remained in force

Japan would have run out of oil, and their economy would have frozen in 1941 or early 1942. For China right now, if it can't access cutting-edge GPUs, there might be substantial long-run costs, but the short-run cost is pretty limited. China's economy is doing badly, but not because they can't get GPUs for its own internal reasons. And so there's not the sense of crisis that the oil export bans created to Japan in 1940-41. The second dynamic is that Japan knew it couldn't synthetically.

produce oil domestically. Chinese leaders think they're going to catch up in terms of producing GPUs domestically. And I'm skeptical, but I think they've got a shot. And that, I think, is to them a better bet to take than rolling the dice on a very risky war that would, as you say, be disastrous for everyone. How do you think this is going to change

as the capabilities of the new models get to be known? So the estimates that we're getting from across the different labs is that we're around 9 to 18 months away from the AI models having the capability of programming at roughly the human level. That is, you don't have to hire a Google engineer anymore. You can just spend the money, use the compute, to generate essentially a synthetic programmer.

And of course, what happens there is this is the first time really that human beings and money become fungible, because before you would have to, hire people, find them, train them in your culture. Now you just pour in money and you get out digital programmers. They increase your capacity. They make you more money, which means you make more programmers. And it seems like as soon as that happens, which is very soon, the

countries that have access to asymmetric power, asymmetric amount of compute start to asymmetrically take off. And that seems at that moment China might feel themselves falling further and further behind. you know it's it's certainly possible i think if you talk to policymakers in western countries and i think the same is true in in china although it's a bit more opaque i think in in basically all policy contexts political leaders really struggle to see around technological corners they're not technologists and anyway predicting the future is very hard especially around points of technological discontinuity and so the assumption that they naturally make is that things will look pretty similar in two years

and in four years with marginal change. And in most facets of life, that's a good assumption to make and at points of technological discontinuity, that's a bad assumption to make. But I think that is the reigning assumption among almost all policymakers, not because they're thinking carefully about the problem,

just because that's the natural assumption to fall into. Yeah, and this is such an important point because even the people who work on these AI systems every day are making systematic errors in their predictions for how fast things come. And the systematic error that they make is that they're constantly getting it wrong in the direction of things happening much faster than they think. And we're seeing that across the board. Every benchmark that we sort of, the AI industry sets, we are hitting those benchmarks much, much faster.

And that's because even the latest Nvidia chips, the H100s, we are starting to get to the place where AI is used to make AI better. So I believe that AI was actually used by NVIDIA to make their H-100 chips more efficient, which makes their AI more powerful, which lets them make even more efficient chips. So we're sort of starting to enter this double exponential. What do you see as the ramifications of policy makers sort of rubber banding back into more of a linear mindset, rather than this exponential or even double exponential mindset? It's tricky to think through the ramifications of an AI acceleration on the chip. industry. Because as you say, on the one hand, it is the case that both chip design companies,

the companies that make the software tools that are used to design chips, and also chip manufacturers, they're all using AI to try to improve the processes. On the other hand, the better AI gets, especially if it happens rapidly relative expectations, the more chance there is that you have disruptions to existing business models and potentially even a challenge to the market positions of leading firms. So, for example, you know, if you, if, you, if, you, if, if you believe that AI will be at human levels of programming in X number of years, it will probably also be at human levels of chip design. And so what's the ramification of that?

Well, that's very, very difficult to say with any confidence. But I think it makes projections like this quite difficult. So refocusing for a second on Taiwan, because I think if you had asked people a couple years ago, they would have said, you know, if China were to attack Taiwan, in the end, the U.S. would win. But now China's navy has gotten bigger. China has the home court advantages, especially with supply chains. Given Taiwan's central role in producing chips combined with China's desire to get AI supremacy,

does that make war more or less likely? So today, China relies on chips made in Taiwan. just as much as anyone else. And so if there were to be a war, Taiwan's chip making facilities to be knocked offline on basically day one, and so China would lose access to chips, not only for AI,

but also for all types of manufactured goods. And when you say knocked offline, do you mean because somebody would blow it up intentionally? What do you mean by that? Well, you don't really even need to blow it up intentionally. That could happen, but if you just constricted energy flows into Taiwan, Taiwan imports much of its energy via liquefied natural gas,

that alone would be enough to shut down Taiwan's chip-making facilities. They also import lots of chemicals from Japan and elsewhere, so any type of blockade scenario would result in the facility shutting down. And so they're as a result quite vulnerable to any sort of geopolitical escalation. Today, everyone would suffer in perhaps varying degrees, but everyone would suffer if we lost access to ships made in Taiwan. But China is trying to become much more self-sophobic. and it's going to make progress towards that goal over the coming years.

And so I think by 2030 or so, China will have a lot of the chip-making capabilities that it needs domestically. The question is, will it have the high-end capabilities that are needed for producing AI chips, for example, at scale? That's uncertain. It depends on the rate of technological progress in China. But if it does, then it would face less cost from knocking TSM offline. It sort of feels like we're essentially heading into another kind of Cold War here,

like sort of a compute Cold War. And I'm curious, what are the lessons, if any, that we can draw from our last Cold War with the Soviet Union for how that might play out this time? Well, the Soviets realized to some extent they were in a compute Cold War as well as a real Cold War, but they didn't really have a good strategy for competing in it. They were fixated on copying U.S. technology rather than developing their own. They never scaled up production domestically.

And as a result, their compute resources were hopelessly inefficient and behind the technological curve. China is trying to avoid that mistake by investing in their own chip industry, by trying to scale up using their vast domestic market. And so they're still behind, but they're much closer to the cutting edge than the Soviets ever were. So that's why the U.S. is fixated on this issue of compute in a way that it really hadn't been for several decades because it's got to compete now because the competition is really so close. As you're just talking about as AI gains ability, say, to code or to replace human cognitive labor, those countries that have access to compute obviously out-compete the ones that don't and starts to increase wealth inequality at the country level and the person level. And it's curious to hear you, like, sort of expand on that. Well, I think the dynamics are actually complex and perhaps more contradictory.

If you ask yourself, why are wealthy countries today wealthy, it's because their cognitive labor is highly valued. And so insofar as machines are capable of substituting for some of that, actually, you might have some pretty contradictory facts. You mean that people that currently have, like, high-paying jobs suddenly will be out of those jobs. And so the internal dynamics of that country, might get very challenging. Is that what you're saying? Well, that's right. And if a country has a high level of income due to cognitive labor,

and suddenly cognitive labor is mechanizable in a large-scale way, you know, this undermines their business model. And so I don't think it's necessarily obvious that at a country-level AI, if it's successful at replacing large numbers of cognitive tasks, that it necessarily benefits the countries that are on top right now, It could benefit certain groups in those countries and others, but it seems like it's a much more complex story than just a winner-takes-all at the country level. The other aspect is that if you think of the two main inputs

being compute and power, we've talked a lot about compute, who produces the compute. Also, who produces the power becomes pretty important where power efficiency dynamics trend. And it's been interesting countries in the Persian Gulf, for example, like Saudi Arabia or the UAE, try to play a bigger and bigger role in data center construction, for example, arguing essentially we've got the power. You can put your compute anywhere in the world, but our power resources

are what's actually limited. And that's something you, I don't think, would have expected if you were to ask about what's the economic implications of more and more capable AI. But if, in fact, power is one of the limiting factors, that could be something that ends up being pretty important. And there's a lot of focus right now, for example, in cooling data centers more effectively, the more you cool them, the less power they draw. And also in designing chips in different ways so that they use less energy per unit to compute. And so that I don't have a high confidence view as to what the slope of that line will be, but it's a pretty important question. What keeps you up at night? When you send your mind to all the places that things

could go poorly, like what is at the top of the list for you? Well, I do think that the chip shortage that we experienced during the pandemic was just a tiny fraction of how bad things could be, how disruptive things could be if we were to lose access to Taiwan. It's not just the high-end data centers that are used

for AI, it's medical devices, its cars, and its household goods. And so the entire world manufacturing sector today is tied directly to chips that are largely, not exclusively, but largely made in Taiwan. And so the world economy sits on top of this

I think pretty fragile foundation of Silicon which for a very long time we've taken for granted. I don't know exactly how to ask this question but maybe I'll just start by asking you where are your worries about the AI capabilities

themselves? Is that something that worries you of like the what they might do, what proliferation of intelligence into society like without bounds sort of like amoral intelligence might do, or is that not where you placed sort of like the biggest risk? Well, I guess I think if you look at the history of computing, you've got to conclude that

more and more capable systems will inevitably follow. The economic incentive to create them is so large. The competitive dynamics, the race that you alluded to, is so fierce that they're going to be created. And so the question is, how do you produce the right context? in which there would be used for better and not worse outcomes. And that gets back to the question of what are the social and political incentives, what are the economic incentives around them, and what are the types of guardrails that you want. But I'm not someone who believes that technological improvements are going to be stopped,

and that's just not going to happen given the incentive to keep building. And so the question is, well, what's the context in which you want this building to happen? Right. It's a little different. But the feeling is, instead of saying we must prevent climate change, we must prevent climate change. It's like realizing that we're in mitigation or adaptation. Like climate change is going to be here. We're going to do something about it. So what I'm hearing you say is that the history of compute says the physics and the incentives of it are that we're going to get faster chips. We're going to get them at cheaper cost. We're going to discover new physics that accelerates this whole thing, not just like 10x, but. like a thousand X or more, and then it's about figuring out how to mitigate the consequences rather just preventing it. Yeah, and I think promising 1,000 X improvements in most industries sounds like a pretty big promise, but in the chip industry, you know, 1,000 X we've done several times before. Right.

Well, in fact, since the beginning of chips to now, like what is the total multiple of increase in speed that we've seen? You know, there are different ways you can measure, but if you take the first commercially available chip had four transistors. And so there's about 10 billion times the number of transistors on a new Nvidia GPU.

Is there any other industry that has had similar rates of improvement? No, nothing else comes remotely close. And so I think that speaks to the challenges of trying to do on-chip governance. There are efforts underway

to explore ways to make this robust. But I think like any sort of restriction you put on a piece of hardware, it's only as good as the incentive to break it is small. And so that I think is one. I think the second thing is that it's easier to envision straightforward limitations, but harder to envision more sophisticated ones.

And so you can't go above a certain speed or you can't go in a certain location. That's relatively straightforward. But saying you can't train something that's dangerous is much, much more difficult. to envision how you'd even begin to go about that type of process. Chris, I have to ask you now a final question, and this one comes from Sasha, our executive producer. And her question, I think, is one that many parents have, which is, given the direction this is all going,

what is the future prospect for our kids? What should they study now to have them have any kind of job in the future? Oh, I don't know. That's a hard question. Well, I guess I think the history of most big technological shifts suggest that what you want above all is the ability to harness those for economic purposes. And you don't know what the right way to harness those are going to be. But in terms of skills, I think the skills that are going to be the most valuable are the skills that let you manipulate technology to do useful things with it.

And so that probably doesn't mean traditional STEM, just like I think my generation practiced a lot less long division than my parents did because we knew we could rely on calculators. Maybe the trajectory of change is going to be more rapid, but I think the basic dynamic of trying to find ways to harness technology, the skill sets you need to do that, will be the source of economic value in the future. Awesome.

Chris, thank you so much for coming on your undivided attention. Understanding chips and GPUs and geopolitics, it's a lot to hold in your head, but it's also critical to understand where we're going. So thank you very much for elucidating it all for us. Well, thank you for having me. All right, just one more thing. One more thought.

Chris doesn't believe that on-chip governance, the ability to control how much compute is being used for what is going to be practical. Now, I don't know if that's true or not. I've heard other opinions, but if that is true, it means the cavalry isn't coming. We're not going to get a technical solution to choosing how in what ways compute, that is cognition, that is intelligence, will be used by humanity. What that means is

that overall governance, trustworthy governance, is where the solution is going to have to come from. And I think something we could probably all agree on is that if hundreds of billions of trillions of dollars is going into creating AI capabilities, then at least, you know, one to 10% of that should be going into figuring out what is the form of trustworthy governance that steers the entire thing. Your indivited attention is produced by the Center for Humane Technology,

a non-profit working to catalyze a humane future. Our senior producer is Julia Scott. Kirsten McMurray is our associate producer. Sasha Fegan is our executive producer. Mixing on this episode by Jeff Sudaken, original music and sound design by Ryan and Hayes Holiday. And a special thanks to the whole Center for Humane Technology team for making this podcast possible.

You can find show notes, transcripts, and much more at HumaneTech.com. If you liked the podcast, we'd be grateful if you could rate it on Apple Podcast, because it helps other people find the show. And if you made it all the way here, Let me give one more thank you to you for giving us your undivided attention.