美国主导完整 AI 技术栈的计划：Sriram Krishnan 解读

Hi, listeners. Welcome back to No Priors. Today, Elad and I are here with Sriram Krishnan, a top White House official currently serving as the senior White House policy advisor on artificial intelligence. A former tech executive and venture capitalist, he's one of the lead authors on the American AI action plan released this past week. We talk about the national implications of the AI race, what position we hold today, the workforce and energy needs of the future, and how to win.

Sriram, thank you so much for joining us today for No Priors.

Thank you for having me. I'm a longtime fan. I've never been invited before. I was always a bit sad, but thank you for having me for the very first time. And first, I have to point out, for folks who are listening on audio, that Elad has never looked as good, as dashing, or as handsome as he does now. Elad, you dressed up for me. I'm honored.

This is how you can tell that Sriram is in politics now. He has the liquid tongue of gold with which he coaxes everybody into doing his bidding, so it's very good.

For our audience, Sriram has been a well-known Silicon Valley figure. He worked at Andreessen Horowitz. He worked at a number of the marquee companies and names in Silicon Valley over the last decade-plus, and now he's in government, where he's working on a variety of exciting initiatives around AI and other areas. Could you tell us a little bit more about your role? Should we be calling you Your Excellency, or is there some special title we should be using now that you're in government?

You don't have to, but I will take it. Thank you. It's fascinating for me to be here talking to you in this capacity because I've known both of you forever and ever. We've had hundreds of interactions, and I've also been such a fan of the pod. Congratulations.

Just to give a little bit of backstory, I've been in Silicon Valley for a long time. I feel very old. I did a tour of all the large consumer social media companies, and then I was at Andreessen Horowitz for the last 4 years, competing actively for CDZA term sheets with both of you, I'm sure. All this while, I had no real intention of joining government. I wasn't particularly interested in policy.

1. Sriram Enters Government

But what wound up happening is that, a couple of years ago, I moved to England to head up all of Andreessen's international efforts. At the time, the UK was a hotbed of all the AI policy debates. They had the AI Safety Summit at Bletchley Park, and this was the peak of, I would say, the effective altruist versus EAC drama that was going on. I got pulled into a lot of those discussions, and I remember thinking to myself, "Wow, a lot of people who were in very senior roles in governments in the United States back then and in other parts of the world didn't know what they were talking about when it came to AI."

I was convinced that they were doing the absolute wrong thing on many topics—for example, open source or helping startups—and it was just really, really bad in a way that I think the industry didn't really appreciate until much later. That got me interested in policy, which, by the way, was a word I didn't even understand. We can even get into what that means.

When President Trump was inaugurated, in the first week, he did 2 things. One, he rescinded the Biden executive order on AI, which was bad and awful in many, many ways, which we can get into. Then he signed a new executive order that basically said America should dominate and win on AI. He then called upon a few of us and said, "You guys need to come up with a plan within 6 months to figure out how America is going to dominate and win." That set us off to the races, and I think everything that has happened since then culminated in the event we had yesterday, where we put out this long, 28-page document on America's AI Action Plan. We had a bunch of executive orders. That's a little bit of the history.

That's great. Could you tell us a little bit more about the main things that you considered as you put together this plan? What are the things that you worry about geopolitically? How do you think about AI and competition—big tech versus small tech? It feels like there are a lot of threads in that, and it'd be great to get a view of the main issues that created this plan. Then it'd be great to talk through the plan itself.

2. DeepSeek Starts The Race

One of the catalytic moments that happened was the day before I started this job. I got a call, and this was the weekend DeepSeek came out. There's actually some chatter I've heard online that China timed it so it could come out right after the president was sworn in. We were like, "Hey, we just want you to come in and brief a lot of people at the White House on DeepSeek because we're like, 'Hey, what is this? Is it cheaper? Is it faster? Do they have some magic way of training these models that only costs a few million bucks and not hundreds of millions of dollars? What's going on?'"

You might remember the narrative that existed that weekend. David and I helped brief all of the White House leadership. It was really a starting gun because that moment was profound. It immediately told us a few things: America doesn't have a huge lead on AI; it actually has a very, very small lead.

At the time, DeepSeek was the only reasoning model that was not OpenAI's. I don't think Claude had come out with a reasoning model yet. I don't think Google had one yet. It was the only non-OpenAI reasoning model, and it was very high up in the leaderboards. It was a bit unclear what their cost claims were and how they had gotten there. I think we know a lot better now.

Yeah, and all of that ended up turning out to be very overstated, right? Basically, there was a claim that it was a few million dollars to train the model, and they didn't really talk about the hundreds of millions of dollars that were probably spent to get to that point. The last training run was what they paid for.

Yes, absolutely. I would say there were claims that were inflated and claims we take seriously. The inflated claims were, to your point, the kind where they put out the final training run and not all the ablations and training costs. If you look at the paper, by the way, I don't think they make the claim that the total cost was a few million bucks. I think that's what the press imputed.

But I do think they deserve a lot of credit. I always make a point of saying, "Look, DeepSeek did some very, very good technical work." If you think about it, they didn't have as good hardware as the American model companies do. What they did with KB caching, MLA, and the various theories about how they actually got chain-of-thought was really novel. Maybe they did it by themselves. Maybe they had some help from American companies. But there was some really novel work there.

I think it told us that we don't have a lead that we can take for granted. They were definitely the best open-source model. Even today, I would probably say the Chinese models—DeepSeek and Qwen—are the best open-source models. Symbolically, it told us that we are now in a race—in a very close race, by the way.

What does it mean to win the AI race? Why do we need to win it? What would losing mean? How would we know we've won?

3. Why America Must Win

I suspect you folks agree that AI might be the most transformational economic and cultural force of our lifetime. I believe that if the country or ecosystem that winds up getting ahead is going to have these cyclic effects, you're going to power productivity, have drug discovery, discover new materials science and new technologies, which then feed back into your infrastructure and your economy. You're going to get this flywheel effect where whoever winds up getting ahead could wind up really accelerating ahead in a classic network-effect ecosystem way that all of us in Silicon Valley will understand.

Now, that is purely in the civilian economic context. You can also imagine a military context. Think about everything from drones to autonomous weapons. I'm pretty sure it's not in our best interest to have another country develop that same economy of scale and flywheel and race ahead of us. So that's the race.

One interesting question that we've been pondering, which we can get into, is how do you actually measure what it means? How are we doing in the race? One measure I've been playing around with—and maybe I'd love to get your take on—is that Google just announced this morning that they run inference on 1 quadrillion tokens per month or per quarter; I forget which one.

Let's say the world runs inference on, I don't know, maybe 10 quadrillion tokens a month. We don't know what the number is. What share of those tokens are being inferred using American hardware and American models? How do we maximize that market share? That's one of the mental models I've been playing with.

In a way, you can think about it as America Inc. We have a product stack starting with GPUs from NVIDIA and AMD and a bunch of others. We have a model layer with, obviously, OpenAI, Grok, and Gemini, and many, many others.

We have an application layer. You had many, many, many of them on your podcast, from agents to all kinds of software. How do we make sure this American stack is dominating that market share of token inference? Very good metric.

4. Models Become Cultural Exports

Yeah, it's really interesting because one other thing that you didn't mention, I feel, is a cultural exportation through the models. If you look at prior waves of culture spread, it was the movie industry, it was social media, and then now it's these models, because a lot of people go to these models as a source of truth for history, information, and other things.

There have been some famous examples in some of the Chinese models where there's omission of Tiananmen Square or omission of other facts. Relatedly, there are some things in some of the U.S. models that seem very politically slanted or otherwise not quite great. But it's interesting to also think about it from the perspective of broader cultural exports, so I just wanted to add that to your points on defense and scientific progress in other areas. I think that's another key thing.

Exactly. It's something we are actually addressing, and you're absolutely right. I grew up in India, and a lot of my exposure to Western culture was on the internet and Google. Obviously, a large part of the internet was American, and that kind of introduced me to Americana. Imagine if, in 1995, the internet was not run by America, but run by one of our adversaries.

In a similar context, you're absolutely right. When DeepSeek came out, I think there were all these great examples of lots of stuff in there that probably doesn't align with American values. Now, we are actually addressing this. The president signed an executive order yesterday called “No Woke AI in the Federal Government.”

What it does—and this is probably going to be one of the spicy bits for your audience—is basically say that, from day one of the Trump administration, we have tried to fight back against DEI, wokeness, critical race theory, whatever you want to call it, in all parts of the federal government, along with all kinds of propaganda. What this EO does is actually very simple: It says that all models that the federal government will procure—aka, what your taxpayer dollars will be spent on—have to do 2 things. They have to be truth-seeking, and they can't have artificial ideological bias added. If bias is added, you just have to be transparent about where you're getting that bias from.

It should be very simple for most people, but to your point, if you're saying nothing happened in Tiananmen Square in the early '90s, that cuts to the heart of that. It also cuts to the heart of many, many other things from the culture wars that we have now been trying to fight against.

Hey, Sriram, you used to work in social media for a long time, right? This sounds a little bit familiar in terms of: Is it a platform? Is it a publisher? What is the information consumption that most consumers have? Where does that analogy apply or break down?

It's a good question. I think in some ways that's for the industry and the ecosystem to answer a bit. You're right, I spent a lot of time at Facebook, now Meta, and Twitter. One of the things I saw when I was at Twitter was how easily you could inject cultural bias into your algorithms.

I have so many stories about how, if you pick the right kind of Twitter accounts, which then feed into the trending algorithm, which then feed into Twitter Moments, and then every journalist or editor wakes up, next thing you know, it's one of the news stories of the land, and BuzzFeed will write a piece saying, “People on the internet are talking about this.” I saw this over and over again, and it left me with a profound appreciation of how algorithms can shape culture.

One of the things I always say is that Twitter, or X, is the memetic battleground upon which we fight a lot of these ideological battles. So when it comes to AI, I think it's probably going to be very similar. My kids use ChatGPT to answer everything, from history to geography to silly kids' questions. You can easily imagine a world where people inject their own cultural biases into this.

In the EU, we have a few good examples. We have examples of the Pope being seen as a Black person and misgendering someone being seen as worse than a thermonuclear explosion. A lot of it is meant to say that you can easily imagine a world where these systems, which are at the heart of so many things that the government is going to use and we are all going to use, are artificially injected with an ideology, at least without being transparent about it.

What are some of the other main points of the announcement from yesterday?

One of the ways David and I try to think about this with some of the people we work with is that it should make sense as a strategy for almost any technology company. I hope that—please go read the document. It's actually pretty readable, and hopefully, for those of you who work in the tech industry, it should make sense.

5. The Action Plan Has Three Pillars

We think that if America is going to win the race with China, it needs to do 3 things, and they ladder up to this strategy. The first is that we need to build infrastructure. At the heart of this, if you go back to the scaling laws, what do we need? We need computation, and we need data.

In the United States, it's been really challenging with the grid we have and the crazy permitting around constructing new data centers to get some of these projects off the ground. So the first part of the action plan really dives into what the president calls “Build, Baby, Build,” playing on “Drill, Baby, Drill.” It's all about how we make sure we are building infrastructure, because obviously some of the other countries are doing that, too.

As an example, one of the things it talks about is making permitting on federal land a lot easier for data centers when it comes to old environmental laws or other regulations that get in the way. Think of that as: Let's make sure we are building the infrastructure to power these models as we scale up. So that's number 1.

The second pillar is innovation, which I would describe as making sure all these amazing companies—everyone that you know of, or maybe some companies that don't exist yet—can build applications and models, or anything they want, as fast as they can. What we are trying to do is a couple of things I really want to highlight.

The first is that we want to cut through red tape. Until a year and a half ago, I was in California along with all of you. California almost passed SB 1047, which, if that had happened, would have been the end of open source in the United States, by the way. We would not have had a Llama, and we would not have had an Owen mini coming out.

A lot of states want to do versions of this, and we think that AI is a national priority. If we're going to compete with China, we need to make sure that these are things we deal with at the national level rather than having every single state—especially states that have ideologies that you and I may not agree on—try to set their own rules.

By the way, some people may not understand this, but I understand it: If you have a small state set rules, it can often become the de facto law for the country. If you're a company, you're like, “Well, I have to operate in this state, or I have an office there, so let me just do that for everybody.” It's kind of just like what the EU does.

So we want to make sure we cut through red tape. Let's make sure regulation happens at the federal level. That's very, very key, because I think that's going to enable not just the big companies, but every Series A, Series B, acqui-hire company—whatever the kids are doing these days—to get off to the races. That's number 1.

The second part is open source. I think we probably talked about this a bit offline. Open source is one of the big reasons I actually got into the policy world. The Biden administration really, really tried to scare people about open source and talked about how unsafe it was. SB 1047 obviously tried to basically ban it in many ways.

What the EO does is say that open source is a space where the United States needs to win. It actually points to some resources that are going to be made available for research, because I think you and I know that open source is what everyone uses—from a kid in their bedroom or dorm room, all the way to a startup, to somebody who wants lower-cost inference in their IoT device, or a robotics startup.

For context, too, much of the internet runs on open-source software, right? The server software and other things—much of that is open source. The protocols are all open for the internet. That's also true for crypto.

It's interesting because removing open source from things like AI actually just centralizes power, right? It centralizes power into a small number of companies that could then be controlled by the government. To some extent, the fact that you all are supportive of open source means you actually are supportive of a thousand flowers blooming, but also a lack of direct government control in literally everything AI. So it's a very interesting counterstance to take.

By the way, Elad has our talking points better than I have, because that is absolutely right.

One very fundamental difference I think we have with the Biden administration is that the Biden team really looked at AI as something to be centralized and controlled. Everything was about how do we make sure that we regulate these 3 or 4 companies, and only 3 or 4 companies can build AI. They had to submit their models for testing. It was all about control in a centralized fashion.

Now, when I moved to D.C., one of the things I realized is that's kind of the way D.C. thinks: control and centralize in one place. You and I know that's not how Silicon Valley thinks, and one of the reasons Silicon Valley is the envy of the world is because anybody, any day, can go to a Y Combinator seed round or raise a round, or just go off to the races, and they could build something amazing that catches everyone's imagination. I think what we want to do is enable just that, rather than say, “Okay, we want to centralize power within a 10-mile radius of where I am right now.”

Yeah. In general, central planning tends to lead to very bad economic outcomes, and so that's the collapse of the Soviet Union, et cetera. It's something that's been tried many times before in many industries, and it tends to lead to a very bad place in terms of innovation and economics.

I think one of the things that people underprice about open-source models is they're going to happen, and it's a strategic weapon. They're happening, and Western companies are using Chinese open-source models very broadly already. If you believe that not every model is going to be ideologically neutral or aligned with American and democratic values, then you probably have a problem, right?

The ability to support whatever your point of view is—pluralism and openness and innovation—and have some control as an ecosystem versus in a centralized way is a very different point of view than, “We'll let China develop it.”

Yes, and I think you're making a profound point. You're already seeing that where, when somebody's using DeepSeek or Qwen, that's an expression of soft power. I would much rather have them using a model built by somebody who kind of agrees with us and has our values. That's number 1.

The other issue I would point to is that these models—we don't know what's inside them. Interpretability is still a nascent field, and you could very easily see ways where you plug a model into Cursor or Windsurf, and you generate a piece of code, and then 2 years down the road, it turns out that code had a little if statement saying, “If I'm running in some piece of critical infrastructure, go do something else.”

We don't have ways to validate all that. There are a lot of reasons why we want to make sure that our American models or Western models wind up winning, and this is something I think we are going to put a lot of focus on.

6. Powering The AI Buildout

Just because you have such a good view into this, can we talk a little bit about infrastructure and energy, since you kind of made that point number 1 in terms of what sort of stack we need? People hear these claims from the leaders of the large labs that we're building a data center the size of Manhattan, or that it's the energy that a city uses at any point.

Can you contextualize how much capacity we really need to build and what the biggest bottleneck is? Is it the grid? Is it energy sources? Is it workforce? When you want to solve this problem, as a systems person, what is the first problem?

Okay, so the first thing I would say is it is a system, and this system wasn't really battle-tested for decades. Somebody showed me this number: I think the United States basically had 1% to 2% of power-usage growth for a very, very long period of time.

You can imagine this whole system of everything from gas turbines, coal, and renewable energy. There was regulation, which really stopped nuclear, and then you had these poor state utility companies, which often didn't have the incentive to innovate. You basically ran the state. You weren't really getting new demand or competition. You had a grid that wasn't really pushed because, again, you didn't need to.

Then you have essentially a patchwork of environmental laws and regulations, everything from water to emissions to a whole other set of things, which I'm sure I'm forgetting. Somebody explained it to me as this tangled spaghetti mess of things which, again, until 2 years ago was just fine because you and I were not dramatically using more energy than we were using 10 years ago.

Now, that obviously changed. The scaling laws arrived, and everybody is trying to build new things. I think the way we are trying to attack it is at every single step of the way. One is: how do we make generation better? Second, how do we make sure we make constructing these data centers better—making it easier to get these regulations and red tape out of the way, making sure we put focus on the right energy sources, and making sure we have those lined up?

We are trying to take an approach to all of this, but it is a complicated problem because there are so many different players, so many different states, and such a patchwork of laws and regulations involved. But I think what I would encourage folks to do is look at the executive order on infrastructure that the president signed yesterday, which I think is going directly at this.

We also have something called the National Energy Dominance Council, which works very closely with Secretary Burgum and Secretary Wright at the Departments of the Interior and Energy, and I think you're going to see a lot more from us on that front. The short answer, Sarah, is that it's complicated. I think we are taking a very, very strong approach to this, but there's going to be more to come.

How do you think energy infrastructure is going to feed into these big data center build-outs? One theory I heard is that fiber is cheap and easy to lay, while the grid is hard to build out, and so therefore you're going to centralize data centers near sources of cheap power. Then you just run fiber into them versus moving things around based on other types of capacity from a telecommunications or other perspective.

Are there specific sources of energy that you think are going to power this AI revolution, or things we need to be invested in? Obviously, the president has issued some executive orders around nuclear. I'm curious how you think about what that future really will be and what the major sources of energy are that we really need to be dependent on, and how does that all shape up from an infrastructure perspective?

What I think we see our role as is getting rid of the red tape. Let's make sure the permitting on these things is super easy. Nuclear is another case where I think for decades and decades, the climate lobby and the doomers have kind of stopped any real efforts over there. So I think we're seeing a lot of effort to get the red tape out of the way, get construction going, and see where we get.

The other thing that I think is interesting from an infrastructure perspective is manufacturing capability and supply chain. A subset of the AI supply chain is dependent on China or other countries. Are there certain areas of the supply chain that we should be repatriating back? Or how should we be thinking about American manufacturing more generally?

I say that America needs not just engineers, but people up and down the stack. It needs electricians and technicians. We need to get construction going, and we need to get these jobs and this whole ecosystem back in the U.S.

If you look at the action plan, there's a bunch of stuff in there about this. I think I mentioned 2 parts of the action plan: building, and then innovation on cutting out red tape and open source. The president also talked a little bit about copyright yesterday.

7. The American Stack Goes Global

The third piece of the action plan, which I also think is a pretty dramatic switch away from how the Biden folks thought about it, is making sure the world uses our standards and our technology. Just for context—and again, unless you are a policy wonk, you may not be super familiar with this—during the Biden era, there was something called the Biden Diffusion Rule, which was a 200-page document that basically made it illegal for America to export GPUs.

It was really hard for Jensen or Lisa Su to get their GPUs out to other countries, even some of our allies who want to help us out and who are really enthusiastic about AI, but we were not actually giving them GPUs. So we listened to that order, and one of the things we talk about is how do we make sure that we get all of our allies around the world using the American stack?

That means how do we make sure—and we just did this in the Gulf with the American AI Acceleration Partnership—that we are getting our GPUs over? One of the advantages of doing that is we get our GPUs over, we probably get them to run our models as opposed to models from another country, and we go from there.

Having an American stack that we can export and that the world standardizes on, I would say, is the third part of the action plan.

One other topic where I think people believe China has a lead right now is in certain areas of robotics, whether they are humanoid or otherwise.

It's drones. It's potentially catching up on self-driving and autonomy. If you think about that from a societal perspective, it's obviously automotive. If you look at the European market share of cars, BYD and others are really taking enormous amounts of share. These are the same technologies that would also be used from a defense perspective.

To some extent, one could argue that there's 2 parts of AI. There's the digital-form side of it, and then there's the real-world robotics, drones, and interactive side. How do you think about that in the context of American policy, and what in the action plan addresses the capabilities to build these physical-world products?

The action plan actually has a section in it making sure we're set up for robotics. I think that's obviously going to become super key within the next 18 to 24 months. I would say it ladders from everything else that we talked about, both in the US and internationally.

The first is making sure that our model companies can actually build as fast as they can, and our startups can innovate as fast as they can. The second piece is that we want to make sure that the world is using our robotics companies and our models and not, say, DeepSeek or Qwen. That's actually one of the things, because when I was talking to a bunch of robotic startups, you're seeing a lot of distilled DeepSeek and a lot of distilled Qwen out there.

What we want to do is make sure that we have an open-source response, an American response, which pushes our products as a standard out there. It is a focus. I think it's going to increasingly come into focus in the next 6 to 12 months, and we are spending a lot of time on it.

Related to that, there's always a question of how things actually get done in politics and how it translates into the real world. I think you've gotten something like 90 different agency actions listed in this action plan. How do you think about these things actually translating into industry, the economy, action by companies, and other players? What are the mechanisms that you all have to ensure that these things come together or happen? And if they don't come together, what's Plan B?

Well, there is no Plan B. We want to get this done. One of the things with the Trump administration you will see is that the administration moves really, really fast, which is why in the first week we had a bunch of executive orders.

Look, we already work on all of it. We had 3 executive orders signed yesterday: 1 for infrastructure, 1 for export, which kind of ties to a lot of things we talk about, and 1 to stop ideology and wokeness and DEI. I think you're going to see a lot more. We are already at work on pretty much all of it. There is no Plan B. We're going to get this done.

The other part I would say from yesterday is I've been inundated with just a great response from the industry. A lot of folks that you and I know are really excited to see the government actually maybe understand AI and are actually happy to make sure that American companies can go build American AI. I think they're also very excited to go partner with us. So it's go, go, go. No time to waste. We're getting it done. There is no Plan B.

It's actually exciting, because I think, to your point on understanding AI and government, when I've looked at prior administrations, whether Republican or Democrat, a lot of the people who went into them from tech weren't the core driving forces of tech in the technology world. In other words, they'd get great people, very nice people, but they weren't at the top of the industry. They weren't necessarily the deepest technical experts in some cases. Obviously, there are counterexamples to that.

One thing that's striking about this administration is that the caliber of tech people they actually got this time around is very high relative to prior administrations. I think that impacts the understanding. It impacts how you all are thinking about the world. I found that very exciting and inspiring, in terms of just having a really strong technical basis for what you're doing. So I think that's really good.

Thank you. There's a lot of great people in the administration from the tech industry, not just in AI. For example, you have Emil Michael as the undersecretary for R&E, who's running DARPA in the Pentagon. You have many, many others.

One of the things I think about is that we bring an understanding of how the tech industry works, what is possible, and what isn't. We bring a sense of urgency. We also just really deeply understand the technology. You'd be shocked at how often I've seen David Sacks in a meeting explain how inferencing works, what high-bandwidth memory is, and how the world has shifted from a pre-training context to a post-training context.

We can really mix it up on the technical details, and we obviously also have a lot of strong social ties to the industry, so we can call upon them to help us out. I think it just adds a very different flavor of understanding of AI where, again, to go to my earlier point, I think D.C. kind of just suffered from a lack of real technical understanding of both the industry and the products involved.

I'm exposing my cards a little bit here, but it sounds from both your policies and what you're saying, Sriram, that you're on the same page. Do you think that the US should be a technocracy, just taking that simple statement?

What does a technocracy mean?

Leading with technology and then having a bunch of people in technology leading the country.

I'm not sure I would think of it that way. The way I see it is America has been blessed to have the leading technology ecosystem in the world, and that is an ecosystem which is in an intense competition right now. I think we could have easily lost that competition, and it's still a very, very close race. We need to do everything we can to protect, preserve, and extend our lead.

But at the end of the day, if you look at this administration, we are still trying to make sure that we serve the American worker, the American workforce. If you look at the action plan, that is at the heart of everything we do. I don't think I see it exactly the way you describe it. I see it more as: we have something in a technology ecosystem that is the envy of the world.

The president, by the way, when he was on stage yesterday, talked about a lot of the inventions that the United States had made. We did the integrated circuit. Shockley invented the transistor. We had the Fairchildren. The internet came from us. We did PageRank and Google. We did the iPhone in Cupertino. So many of these things are used by the world.

What do we do to make sure we preserve that lead, especially when it comes to AI? If you look at AI, there are so many potential timelines that AI could take. I have read AI 2027 from Daniel. I have read much more optimistic takes on AI. I think there's going to be an event horizon beyond which you and I can have reasonable discussions on how AI could play out.

But in any one of those scenarios, I want to make sure that the United States is well-positioned, where we can take advantage of the productivity, the science, and the technology breakthroughs that are going to happen, and then be set up for whatever happens next. I'm not sure I really answered the question you phrased.

No, no, no, you did. I was trying to ask the question in a bit of a triggering way, because I think a lot of people would say, "It shouldn't just be driven by the technologists," and it's like, what good does that do us in winning the AI race?

I think that's actually a really profound claim that you made, which I hear as: the country that builds the most capable AI systems gains a lot of upstream control and influence that has been traditionally very American, right? And we should all care about that.

You use the examples of accelerating life sciences, new materials, optimizing industry, being more efficient in healthcare and education, and things that matter to every American and to compounding national wealth. Sometimes a lot of this discussion becomes an argument about what parties have influence versus what position we want to have as a country, right? And whether or not we want that edge.

That's right. I think, very simply, we want to win.

I have 2 questions for you before we run out of time. One is just going back to this idea of you being the strong proponent of open source and open weights. What is the strongest counterargument to the people who would raise concern that the p(doom), the probability that there's some sort of cycle of key man risk or some factor of abuse of these powerful models, increases with open-source models?

If you look at the action plan, it's kind of a manifestation of how we think about things, right? We do talk a lot about risk. We talk a lot about having systems in place to identify cyber risk, bio risk, et cetera.

I think the difference from the Biden administration or the folks who talk about p(doom) a lot on Less Wrong is that we are just inherently more optimistic. If folks haven't seen it, I encourage them to watch the vice president's speech in Paris, where he talked about how we want to embrace AI with optimism rather than fear.

I think one of the things that happened is that there was a lot of fear mistakenly placed on open source. I think there were 2 kinds of fears people talked about. One was what you talked about: What are the risks if these models could do really bad things?

The second was, “Are we actually giving away our secrets to China?” What DeepSeek showed us is that China is actually building these models just fine by itself, and it’s actually American models that were far behind. Immediately, I think that argument got refuted.

On the p(doom) question, I think that’s a perfectly fair question, and I think we need to be vigilant about it. The action plan talks about it. But we have to remember, we are in a race with China, and there are going to be catastrophic consequences if Chinese models are running on every robot, every camera, every car, and every device around the world. We just have to face that reality.

I think also the people who are driving the p(doom) arguments, to some extent, are coming from 1 or 2 large companies that have closed-source models. I think we also forget the incentives of who’s actually pushing for this. If you have a big pharma company, they work with the government to prevent other entrants into the industry, and this is exactly what it felt like was at least partially happening in the AI world.

Now, that may be for perceived altruistic causes or because they’re worried about humanity. But I do think the reality is that a small number of companies have been pushing this narrative pretty strongly: that open source is bad. These are companies that control the closed-source models.

Absolutely. I think there are a few things going on. One is people pushing for regulatory capture. Second is, obviously, the schools of thought from effective altruism, and a lot of people worried about this—all of it kind of mixed together.

Here’s my rebuttal to that: I think one of the things that open-source software has shown us on the internet is that, by default, open source is just safer and more secure. What does Linus’s Law say? More eyes make every bug shallow. Over the last 20 years, what has the security industry learned? The more scrutiny you put your libraries and browser-rendering engines through, the safer they become. We have seen that time and time and time again.

I think the same holds true for open source and open weights. If you have a model up on Hugging Face and somebody downloads a 500-gigabyte file, and there are thousands of students and researchers just pounding away on that, I think there’s a good chance they’re going to find issues a lot better than a very small safety team inside a large lab. I’m a big fan of open source sometimes being a lot more secure than closed source as well.

Awesome. Thanks so much, Sriram.

Thank you, Your Excellency, Your Governorship, Your Grace. I’m not sure, again, what the right title is. Your Policy Advisorship.

Feel free—Elon, right? The more inflated it is, the more it helps my ego. So thank you.

It was Your Excellency. That’s what we started with.

We really appreciate the time today, Your Sriramship. Thank you for joining.

Thank you so much. Such an honor. I love the work you folks do, and thank you for having me.

Find us on Twitter at nopriorspod. Subscribe to our YouTube channel if you want to see our faces. Follow the show on Apple Podcasts, Spotify, or wherever you listen. That way you get a new episode every week. And sign up for emails or find transcripts for every episode at no-buyers.com.