Sacks, Andreessen & Horowitz: How America Wins the AI Race Against China

The Europeans have a really different mindset for all this stuff. When they talk about AI leadership, what they mean is that they're taking the lead in defining the regulations. They get together in Brussels and figure out what all the rules should be, and that's what they call leadership.

1. Europe’s AI “leadership”

It's almost like a game show or something. They do everything they can to strangle them in their crib, and then if they make it through a decade of abuse of small companies, they're going to get the money to grow. Ronald Reagan had a line about this: “If it moves, tax it. If it keeps moving, regulate it. If it stops moving, subsidize it.” The Europeans are definitely at the subsidizing stage.

2. Why AI and crypto fit together

David, welcome. Thanks for joining. So, David, you're the AI and crypto czar. Why does it make sense to have those as a portfolio? What do they have to do with each other? And then I'll have you lay out the Trump plan in those 2 categories, and how we're doing.

They're 2 technologies that are relatively new, and so there's a lot of fear around them. People don't necessarily know that much about them. They don't really know what to make of them.

From a policy standpoint, we can talk about the similarities and differences. The approaches are a little different. With crypto, the main thing that's needed is regulatory certainty. All the entrepreneurs I've talked to over the years say the same thing: “Just tell us what the rules are. We're happy to comply, but Washington won't tell us what they are.”

In fact, during the Biden years, you had an SEC chairman who took an approach that has been called “regulation through enforcement.” That basically means you just get prosecuted. They don't tell you what the rules are; you get indicted, and then everyone else is supposed to divine what the rules are as you get prosecuted, fined, and imprisoned.

That was the approach for several years. As a result, the whole crypto industry was in the process of moving offshore. America was being deprived of this industry of the future.

3. Making America the crypto capital

President Trump, during his campaign last year, gave a now-famous speech in Nashville in which he declared that he would make the United States the crypto capital of the planet and that he would fire Gensler. That was the big applause line. He talked about how surprised he was at what a big ovation he got. He said it again, and the crowd erupted again.

In any event, he promised to provide this clarity so that the industry would understand what the rules are and be able to comply. In turn, that should provide greater protection for consumers and businesses—everyone who's part of the ecosystem—and make America more competitive.

I think that's the mandate on crypto. It's pro-regulation in a way. We want to put regulations in place. In a way, AI is kind of the opposite. I think the Biden administration was too heavy-handed. They were starting to regulate this area without even understanding what it was.

No one had really taken the time to understand how AI was being used or what the real dangers were. There was intense fearmongering, and as a result, the Biden administration was in the process of implementing very heavy-handed regulations on both the software and hardware sides. We can drill into that.

With the Trump administration, the approach has been that we want the United States to win the AI race. It's a global competition. Sometimes we mention that China is probably our main competitor in this area. They're the only other country that has the technological capability, talent, know-how, and expertise to beat us in this area. We want to make sure the United States wins.

In the United States, it's not really the government that's responsible for innovation; it's the private sector. That means our companies have to win. If you're imposing all sorts of crazy, burdensome regulations on them, that's going to hurt, not help.

The president gave a very important AI policy speech a couple of months ago, on July 23, where he declared in no uncertain terms that we had to win the AI race. He laid out several pillars for how we do that: pro-innovation, pro-infrastructure—which also means pro-energy—and pro-export. We can drill into all those things if you want, but that was the high line.

Again, with AI, the idea is: How do we unleash innovation? With crypto, it's been more about how we create regulatory certainty.

4. AI strategy shift: Innovation over fear

In terms of my role, why am I doing both? I think the common denominator is that these are new technologies. They both obviously come from the tech industry, which has a very different culture from Washington. I see my role as helping to be a bridge between what's happening in Silicon Valley and what's happening in Washington, helping Washington understand not just the policy that's needed or the innovation that's happening, but also, culturally, what makes the tech industry different and special, and how that needs to be protected from the government doing something excessively heavy-handed.

We're going to talk a lot about AI today, but just on crypto, I've had this interesting experience this year, after the election, as people adjusted to the change of government. I've had this discussion with a number of people in politics who were previously anti-crypto and have been trying to figure out how to get to a more sensible position.

5. The crypto crackdown and de-banking years

I've also talked to people in the financial services industry who followed it from a distance and maybe participated in the various debanking efforts without really understanding what was happening. The common denominator has been that they're saying, “Marc, I didn't really understand how bad it was. I basically thought you guys in tech were just whining a lot, pleading as a special interest, and doing the normal thing. I figured the horror stories were made up—people getting prosecuted, entrepreneurs getting their houses raided by the FBI, and the whole panoply of things that happened.”

I didn't really understand. Now, in retrospect, now that I go back and look, I'm like, “Oh, my God. This was actually much worse than I thought.” Do you have that experience? As you're in there and now have a complete view of everything that happens, do you think people understand how bad it was?

I think it's a great point. I didn't really know either. You generally heard about it. We knew that debanking was going on. By the way, it wasn't just crypto companies that were being debanked; their founders were being debanked personally.

If you were the founder of a crypto company, you couldn't open a bank account. That's a huge problem. How do you transact? How do you make payments? How do you pay people? It basically deprives you of a livelihood. It's a very extreme form of censorship.

That was definitely happening. Then, of course, you had all the prosecutions that the SEC was behind. So, yeah, it was really bad.

I remember back in March, I think, we had a crypto summit at the White House, and 1 of the attendees said, “A year ago, I would have thought it was more likely that I'd be in jail than that I'd be at the White House.”

It was a really big milestone for the industry. They'd never received any kind of recognition like that. The idea that this was even an industry that you'd hold an event for at the White House—at a minimum, I think crypto was seen as very déclassé.

In any event, it's been a huge shift. We've basically stopped that. It was very unfair because these founders wanted to comply with the rules, but they weren't told what the rules were. That was all part of a deliberate strategy, I think, to drive crypto offshore.

One of the things that's very different between crypto and AI, that we've noticed, is that on the crypto front, everybody just wanted rules, and the industry was relatively unified. Whereas in AI, we've seen very interesting calls coming from inside the house, with certain companies really going for regulatory capture.

6. Anthropic, regulatory capture, and AI gatekeeping

People who have early leads are saying, “Let's cut off all new companies from developing AI,” and so forth. What do you make of that, and where do you think that's going?

I think it's a very big problem. I actually recently criticized 1 of our AI model companies for engaging in a regulatory capture strategy.

Yes. Very fair criticism, by the way.

It is very fair. Of course, they denied it, and then—should I tell the story?

Yeah, sure.

Rarely do you get vindicated on X so thoroughly and completely as I did on this. After this company—it was basically Anthropic—denied it, what basically happened is that Jack Clark, who’s a co-founder and head of policy for Anthropic, gave a speech at a conference where he compared fear of AI to a child seeing monsters in the dark or thinking there were monsters in the dark. But then you turn the lights on, and the monsters are there.

I thought that was such a ridiculous analogy. I mean, it’s basically puerile. It’s so childish as to be almost self-indicting, because you’re basically admitting the fear is made up, not real. In any event, I said, “Well, this is fear-mongering and part of the regulatory capture strategy.”

Of course, they denied it, but then a lawyer who was in the crowd at his speech said, “Well, yeah, but Jack’s not telling you what he said during the Q&A, in which he basically admitted that everything Anthropic was doing with things like SB 53, which is supposedly just implementing transparency—” He said, “No, he admitted that all of that was just a stepping stone to the real goal, which was to get a system of pre-approvals in Washington before you can release new models.”

7. Why "permissionless innovation" built Silicon Valley

He admitted as part of the Q&A that making people very afraid was part of their strategy. So, again, this is as much of a smoking gun as you could ever get in a spat on X. But the reason why I think that approach is so damaging is that the thing that’s really made Silicon Valley special over the past several decades is permissionless innovation, right?

It’s the 2 guys in a garage who can just pursue their idea. Maybe they raise some capital from angels or VC firms—basically, people who are willing to lose all of their money. These are young founders. They could also be the future dropout in the dorm room. They’re able just to pursue their idea.

The only reason that I think has happened in Silicon Valley, whereas you look at industries like pharma, healthcare, defense, or banking—these highly regulated industries where you just don’t see a lot of startups—is because they’re all heavily regulated, which means you have to go to Washington to get permission to do things.

The thing I’ve seen in Washington is that the approvals get set up for reasons, but those reasons very quickly stop mattering. It just matters how good your government affairs team is at navigating through the bureaucracy and figuring out how to get those approvals. It’s not something that your typical startup founders are going to be good at. It’s something that big companies get good at because they’ve got the resources, and that’s exactly what regulatory capture means.

So, the whole basis of Silicon Valley’s success—the reason why it’s really the crown jewel of the American economy and the envy of the rest of the world—is permissionless innovation. We see all these attempts by all these other countries to create their own Silicon Valley. The reason that’s the case is because of permissionless innovation.

What is being contemplated, discussed, and implemented with respect to AI is an approval system for both software and hardware. This is not theoretical. This has already been happening.

On the hardware side, one of the last things that the Biden administration did in its last week was impose the so-called Biden diffusion rule, which requires that every sale of a GPU on Earth be licensed by the government—that is to say, pre-approved—unless it fits into some category of exception. Basically, the overall idea is that compute is now going to be a licensed and pre-approved category. We rescinded that.

On the software side, like I said, the goal very clearly is to start with these reporting requirements to the government, to the states. Then where that ramps up to is that you have to go to Washington to get permission before you release a new model.

This would drastically slow down innovation and make America less competitive. These approvals can take months. They can take years. When a new chip is released every year and we have licenses that have been sitting in the hopper for 2 years, the requests are obsolete by the time they finally get approved. That would be even more true with models, where the cycle time is 3 or 4 months for a new model.

What exactly is the bureaucracy in Washington going to know about this technology that would put them in a good position to approve it? In any event, this is what is being contemplated right now, and I think it would be a disaster for Silicon Valley and for innovation, and therefore for American competitiveness. I think we will lose the AI race to countries like China if this is the set of rules that we have.

8. “Woke” vs. Orwellian AI and narrative control

Yeah. One of the really diabolical things about their argument is, if they really believed there was a monster, then why are they buying GPUs at a rate faster than anybody? The other thing that we know from being in the industry is that their reputation is that they have literally the worst security practices in the entire industry with respect to their own code.

If you were building this monster, the last thing you’d want to do is leave a bunch of holes around for people to hack it. So they don’t believe anything they’re saying. It’s completely made up to try and maintain their lead.

It’s psychotic. I think it’s a heady drug to basically say that we’re creating this new superintelligence that could destroy humanity, but we’re the only ones who are virtuous enough to ensure that this is done correctly, right?

It’s a good recruiting tool.

Yeah. It’s “Join the virtuous team.”

Yes. I think that’s right.

But, yeah, I think that is definitely true. Of all the companies, that particular one has been the most aggressive in terms of regulatory capture and pushing for these regulations.

Let’s just bring it up a level. It doesn’t have to be about them. There are now something like 1,200 bills going through state legislatures right now to regulate AI. 25% of them are in the top 4 blue states, which are California, New York, Colorado, and Illinois. Over 100 measures have already passed. I think 3 of them just got signed in the last month in California alone.

Let me tell you what Colorado is actually doing. Colorado, Illinois, and California have all done some version of a thing called algorithmic discrimination, which I think is really troubling in terms of where it’s headed.

What this concept means is that if the model produces an output that has a disparate impact on a protected group, then that is algorithmic discrimination. The list of protected groups is very long. It’s more than just the usual ones.

For example, in Colorado, they’ve defined people who may not have English-language proficiency as a protected group. So I guess if the model says something bad about illegal aliens, then that would basically violate the law.

I don’t know exactly how model companies are even supposed to comply with this rule. Presumably, discrimination is already illegal. So if you’re a business and you violate the civil rights laws and engage in discrimination, you’re already liable for that.

If you happen to make that mistake and use any kind of tool in the process of doing it, there’s no reason to go after the tool developer, because we can already go after the business that made that decision. But the whole purpose of these laws is to get at the tool. They’re making not just the business that is using AI liable; they’re making the tool developer liable.

I don’t even know how the tool developer is supposed to anticipate this, because how do you know all the ways that your tool is going to be used? How do you know that this output—especially if the output is 100% true and accurate and the model is just doing its job—was used as part of a decision that had a disparate impact? Nevertheless, you’re liable.

The only way that I can see for all the developers to even attempt to comply with this is to build a DEI layer into their models that tries to anticipate, “Could this answer have a disparate impact?” If it does, we either can’t give you the answer, or we have to sanitize or distort the answer.

You just take this to its logical conclusion, and we’re back to woke AI, which, by the way, was a major objective of the Biden administration. That Biden executive order on AI that we rescinded as part of the Trump administration had something like 20 pages of DEI language in it. They were very much trying to promote what they called DEI values in models.

We saw what the results of that were. We saw the whole Black George Washington thing, where history was being rewritten in real time because somebody built a DEI layer into the model. I almost feel like the term “woke AI” is insufficient to explain what’s going on because it somehow trivializes it.

What we’re really talking about is Orwellian AI. We’re talking about AI that lies to you, that distorts an answer, that rewrites history in real time to serve a current political agenda of the people who are in power.

It’s very Orwellian. We were definitely on that path before President Trump’s election. It was part of the Biden executive order. We saw it happen in the release of that first Gemini model. It was not an accident that those distorted outputs came from somewhere.

To me, this is the biggest risk of AI, actually. It was not described by James Cameron; it was described by George Orwell. In my view, it’s not The Terminator; it’s 1984. As AI eats the internet and becomes the main way that we interact and get our information online, it’ll be used by the people in power to control the information we receive.

It’ll contain an ideological bias that essentially will censor us. All that trust and safety apparatus that was created for social media will be ported over to this new world of AI. Marc, I know that you’ve spoken about this quite a bit, and I think you’re absolutely right about that.

On top of that, you’ve got the surveillance issues where AI is going to know everything about you. It’s going to be your personal assistant, so it’s kind of the perfect tool for the government to monitor and control you. To me, that is by far the biggest risk of AI, and that’s the thing we should be working toward preventing.

The problem is that a lot of these regulations that are being whipped up by these fearmongering techniques are actually empowering the government to engage in this type of control, which I think we should all be very afraid of, actually. Erik Torenberg

9. The AGI hype cycle and Goldilocks reality

Sam Altman said earlier this week that by 2028, he expects to have automated researchers. I’m curious, just in terms of the state of AI model development—or just progress in general—what do you think are the implications?

Some people have been saying that AGI is 2 years away, like the “AI 2027” paper and Leopold Aschenbrenner’s “Situational Awareness: The Decade Ahead.” I’m curious what your reading is of the state of play in terms of AI development and what the implications are.

My sense is that people in Silicon Valley are pulling back from the—let’s call it—imminent AGI narrative. I saw Andrej Karpathy give an interview where, all of a sudden, he’s re-underwritten this, and he says AGI is at least a decade away. He’s basically saying that reinforcement learning has its limits.

Reinforcement learning is very useful. It’s the main paradigm right now that they’re making a lot of progress with. But he says that the way humans learn is not really through reinforcement. We do something a little different.

I think that’s a good thing because it means that humans and AI will be synergistic. The AI’s understanding, if it’s based on RL, will be a little different from the way that we intuit and reason.

In any event, I sense more of a pullback from this imminent AGI narrative—the idea that AGI is 2 years away. Of course, it’s unclear what people mean by AGI, but it was used in this scary way, as if it were a superintelligence that would grow beyond our control.

I feel like people are pulling back from that and understanding that, yes, we’re still making a lot of progress, and the progress is amazing. At the same time, what we mean by intelligence is multifaceted. There’s progress being made along some dimensions, but it’s not along every dimension.

I’ve described the situation we’re in right now as a bit of a Goldilocks scenario. The extremes would be the scary Terminator situation—imminent superintelligence that’ll grow beyond our control. The other narrative you hear in the press a lot is that we’re in a big bubble; in other words, the whole thing is fake. The media is basically pushing both narratives at the same time.

In any event, I think the truth is more in the middle. That’s the Goldilocks scenario: we’re seeing a lot of innovation, and I think the progress is impressive. I think we’re going to see big productivity gains in the economy from this.

I like the observations that Balaji made recently. There were a couple of things that really struck me. One was that AI is polytheistic, not monotheistic. What we’re seeing is many instead of just 1 all-knowing, all-powerful god. We’re seeing a bunch of smaller deities—more specialized models.

We’re not on that kind of recursive self-improvement track just yet, but we’re seeing many different kinds of models make progress in different areas. The other observation was that AI was middle-to-middle, whereas humans are end-to-end, and therefore the relationship is pretty synergistic. I think all of those observations resonate with me in terms of where we’re at right now.

10. AI’s rise: Specialized models, agents, and synergy

Yeah, and that’s very consistent with what we’re seeing as well. Ideas that we thought would for sure get subsumed by the big models are becoming amazingly differentiated businesses, just because the fat tail of the universe is very fat, and you need really specific understanding of certain scenarios to build an effective model.

That’s just how it’s going. No model has figured out how to do everything.

Yeah, the models work best when they have context. The more general your prompt, the less likely it is that you’re going to get a great response.

If you tell the AI something very general, like, “What business can I create to make $1 billion?” it’s not going to give you something actionable. You have to get very specific about what you’re trying to do, and it has to have access to relevant data. Then it can give you specific answers to a prompt.

I think this is partly Balaji’s point: the AI does not come up with its own objective. It needs to be prompted. It needs to be told what to do. We’ve seen no evidence that, at this stage, that’s changing. We’re still at step 0 in terms of AI somehow coming up with its own objective.

As a result, the model has to be prompted, then it gives you an output, and that output has to be validated. You have to make sure it’s correct because models can still be wrong. More likely, you have to iterate a few times because it doesn’t give you exactly what you want, so you reprompt.

We’ve all had this experience. This is why the chat interface is so necessary: it takes you a few times to iterate to get to the output that actually has value for you. Again, humans are end-to-end and the AI is middle-to-middle. We haven’t seen any evidence that that fundamental dynamic is changing.

We’re obviously at the outset of agents. With agents, you can give an objective to an agent, and then it’ll be able to take tasks on your behalf. But I suspect that agents will work better as well when they have a much narrower context. They’re much less likely to go off the rails and start going in weird directions.

If you give it a very broad task, it’s just not likely to completely figure it out before it needs human intervention. But if you give it something very narrow to do, then it’s much more likely to be successful.

If you tell the AI, “Sell my product,” it’s very unlikely that it’s just going to figure out what that means and how to do it. But if you’re a sales rep using AI to help you, there are probably very specific tasks that you can tell it to do, and it would be much more successful doing that.

This also speaks to the whole job-loss narrative. I think this is going to be a very synergistic tool for a long time. I don’t think it’s going to wipe out human jobs. I don’t think the need for human cognition is going away. It’s something that we’ll use to get a big productivity boost, at least for the foreseeable future.

I don’t know if any of us can predict what’s going to happen beyond 5 or 10 years, but that’s just what I’m seeing right now. I’m curious what you guys are seeing on this front.

Yeah, generally consistent with that. Things are improving. On agents, the early agents would go completely bananas and off the rails the longer the task ran. People are working on that.

I do think everything’s working better in a context. At least from what we’ve seen, that will continue. Even to your point on super-smart models, there are 12 video models out there, and there’s not 1 that’s the best at everything, or even close to the best at everything. There are literally 12 that are all the best at 1 thing.

That’s a little surprising, at least to me, because you’d think the sheer size of the data would be an advantage, but even that hasn’t quite proven out.

It is, depending on what you want. Do you want a meme? Do you want a movie? Do you want an ad? It’s all very, very different.

I think this gets to your main point. Mark Zuckerberg said something that I really liked: “Intelligence is not life.” These things that we associate with life—we have an objective, we have free will, we’re sentient—just aren’t part of a mathematical model that is searching through a distribution and figuring out an answer, or even a model that, through a reinforcement learning technique, can improve its logic. So the comparison to humans falls short in a lot of ways, is what we’re saying. We’re just different.

The models are very good at things. They’re already better than humans at many things. The other thing I’d bring up related to this, which is somewhat orthogonal but also quite related, is: Is the future of the world going to be one or a small number of companies—or, for that matter, governments or super AIs—that own and control everything?

Is all the value going to roll up into a handful of entities? You get into the hyper-capitalist version, where a few companies make all the money, or the hyper-communist version, where you have total state control or whatever. Or is this a technology that’s going to diffuse out, be in everybody’s hands, and be a tool of empowerment, creativity, individual effort, and expressiveness—a tool for basically everybody to use?

I think one of the really striking things about this period of time, with you being in this role, is that scenario 2 is very clearly playing out. I think AI is actually hyper-democratized. It has spread to more individuals, both in the country and around the world, in the shortest period of time of any new technology in history, I think.

We’re at something like 600 million users today, rapidly on the way to 1 billion, and rapidly on the way to 5 billion across all the consumer products. The best AIs in the world are in the consumer products, right? If you use current-day ChatGPT or Grok or any of these things, I can’t spend more money and get access to a better AI. It’s in the consumer products.

In practice, what you have playing out in real time is that this technology is going to be in everybody’s hands. Everybody is going to be able to use it to optimize the things they do, have it be a thought partner, have it be an assistant for building companies, starting companies, creating art, or doing all the things that people want to do.

My wife was just using it this morning to design a new entrepreneurship curriculum for our 10-year-old. She spent a couple of hours on it and now has a full curriculum for him to be able to start his first video game company: all the different skills he needs to learn and all the resources.

That’s just the level of capability. To have done that without these modern consumer AI tools, you’d have to go hire a special education specialist or something, which is basically impossible for that kind of thing. Everybody has these stories now in their lives and among people they know.

I think we have a lot of proof that the track this is on is that this is going to be in everybody’s hands, and, in fact, that is going to be a really good thing. I think, David, you guys are really playing a key role in making that happen.

I think it’s so important that this technology remain decentralized, because the Orwellian concern is the ultimate centralization. Fortunately, so far, what we’re seeing in the market is that it’s hypercompetitive. There are 5 major model companies, all making huge investments.

11. The coming wave of open source competition

The benchmarks, the model performance, and the evaluations are relatively clustered. There’s a lot of leapfrogging going on. Grok releases a new model, it leapfrogs ChatGPT, and then ChatGPT releases something new and they leapfrog each other. They’re all very competitive and close to each other, and I think that’s a good thing.

It’s the opposite of what was predicted through this imminent AGI story. The storytelling there was that one model would get a lead, direct its own intelligence toward making itself better, and then its lead would get bigger and bigger. You’d get this recursive self-improvement, and pretty soon you’d be off to the singularity.

We haven’t really seen that. We haven’t seen one model completely pull away in terms of capabilities, and I think that’s a good thing.

Erik, to your point about this narrative about the virtual AI researcher, that was one variant of this imminent AGI narrative. The steps would be: models get smarter, the models create a virtual AI researcher, then you get 1 million virtual AI researchers, and then you get the singularity.

I think the sleight of hand in that is: What is a virtual AI researcher? It’s a very easy thing to say, but what does that really mean? To Balaji’s point, AI is still middle-to-middle. It isn’t end-to-end.

If an AI researcher is end-to-end, there are things it has to do—things the person has to figure out. They’ve got to set their own objective. They’ve got to be able to pivot in ways that AI can’t. Is it really feasible to create a virtual AI researcher?

I think there are parts of the job that AI could get really good at, or even better than humans, but probably that tool has to be used by a human AI researcher. I suspect the argument could be teleological, in the sense that you might need AGI to create a virtual AI researcher, as opposed to the other way around.

If that’s the case, you’re not going to get the singularity. So I’m a little bit skeptical of that claim. We’ll see. Sam says he could do it in 2028. I guess we’ll see in 3 years.

I think all those claims tend to be recruiting ideas as opposed to actual predictions. He’s not the first to mention that idea. Other model companies have been promoting it, but Leopold’s mentioned that, too. We’ll see.

I suspect that’s what’s wrong with that argument: A virtual AI researcher requires AGI. The idea that you’re going to get AGI through a virtual AI researcher is backwards. But we’ll see.

David, you and the administration have also been very supportive of open-source AI, which I think dovetails into this in terms of the market being very competitive. It’s been a moment, considering what you guys have been able to do on that and how you think about it.

Yeah, so open source is very important because I think it’s synonymous with freedom—software freedom. You basically can run your own models on your own hardware and retain control over your own information.

By the way, this is what enterprises typically do all the time. About half the global data center market is on-premises, meaning enterprises and governments create their own data centers. They don’t go to the big clouds.

I’ve got nothing against the hyperscalers, but people like to run their own data centers and maintain control over their own data and that kind of thing. I think that will be true for consumers to some degree as well. So I do think it’s an important area that we should want to encourage and promote.

The irony right now in the market is that the best open-source models are Chinese. It’s sort of a quirk, right? It’s the opposite of what you’d expect. You’d expect the American system to promote openness and somehow the Chinese system to promote closed systems. That has ended up being a little backwards.

I think there are good reasons for it. It could just be a historical accident—the fact that the DeepSeek founder was very committed to open source and that just got things started that way. Or it could be part of a deliberate strategy.

If you’re China and you’re trying to catch up, open source is a really good way to do that because you get all the non-aligned developers to want to help your project, which they can’t do with a closed project. So it’s a great strategy for catching up.

Also, if you think that your business model, as a company or as a country, is, let’s say, scaled manufacturing of hardware, then you would want the software part to be free or cheap because it’s your complement, right? So you try to commoditize your complement.

I don't know whether it's by accident or part of the design, but that seems to be what the Chinese strategy has been. I think the right answer for the U.S. in this is to encourage our own open source. I think it'd be a great thing if we saw more open source initiatives get going. I guess there's one promising one called Reflection, which was founded by former engineers from Google DeepMind.

I hope we see more open source innovation in the West. But look, I think it's very important. It's critical. Like I said, in my view, it's synonymous with freedom, and it's definitely not something we want to suppress.

Now, just back to the closed ecosystem for a second. It's true that we have 5 major competitors there, and they're all spending a lot of money. I do worry a little bit that, at some point in time, the market consolidates and we end up with a monopoly or duopoly, or something like that, as we've seen in other technology markets. We saw this with search, and so on down the line, and I just think that it would be good if this market stayed more competitive than just 1 or 2 winners.

I don't really know what to do about that. I'm just making that observation. I do think that having open source as an option always ensures that, even if the market does consolidate, you have an alternative. It's an alternative that's more fully within your control, as opposed to a large corporation or the deep state working with that corporation.

As we saw in the Twitter Files, the deep state was working with all these social media companies in implementing much more widespread censorship than I think any of us thought possible. We've seen evidence in the past, and again in the social networking space, about how the government could get involved in nefarious ways. It would be good to have alternatives to prevent that, or to make it less likely that that scenario comes about with AI.

Yeah. Well, as you know, we and others are very aggressively investing in new model companies of many kinds, including new foundation-model companies. As you're probably aware, there are a whole bunch of new open source efforts that are not yet public that hopefully will bear fruit over the next couple of years.

12. The AI race with China: Policy, energy, and exports

I think that, at least in the medium term, we're looking at an explosion of model development as opposed to consolidation. Then we'll see what happens from there.

Yeah, that's really good to hear. I think if we assess the state of the AI race vis-à-vis China, this is the only area where we appear to be behind: open source models.

If you don't care whether it's open or closed, I think we have the lead. I think our top model companies are ahead of the top Chinese companies, although they're quite good. But this area of open source seems to be where they have an advantage. It's great to hear that you guys are seeing a lot more efforts coming to market.

Yeah, there's more coming. Good.

Yeah, definitely more coming. Peter Thiel quipped many years ago that he thought crypto would be libertarian or decentralizing, and that AI would be communist or centralizing. I think one thing we've perhaps learned is that technology isn't deterministic, and that there are a set of choices that determine whether these technologies are decentralizing or centralizing.

Maybe we could use that as a segue to go deeper into the state of the race with China. Maybe, David, you could lay out what's most important to get right. You've already indicated that open source is one example. You alluded earlier to our strategy as it relates to chips. Some people say that, yes, it's a good idea to do what we're doing because it'll limit domestic semiconductor production. Other people say, "Well, some of these companies say chips are the biggest limiting factors, so are we enabling them in some way?" Why don't you talk about the state of play and then our strategy?

When we talk about winning the AI race, sometimes we say we're in a race against China, and sometimes we leave it a little bit more vague. I don't think we should become overly obsessed with our competitors or adversaries. I think whether we win or not will mostly have to do with the decisions we make about our own technology ecosystem, not about what we do vis-à-vis them.

The president, in his July 23 speech on AI policy, mentioned a few of the key pillars of how we win this AI race. By the way, I'm not saying it ever ends. This might be an infinite game, but we want to be in the lead, at least. I do think that there could be a period of time where—take the internet, where the internet's still going on, but we understand that who the winners are is kind of baked now. There could be a period of time in which it's kind of baked who the winners in AI are.

In any event, in terms of how we win this race, I mentioned a few of the key pillars. Number 1 is innovation. It's very important to support the private sector because they're the ones who do the innovation. We're not going to regulate our way to beating our adversary. We just have to out-innovate them.

I think right now the biggest obstacle is the frenzy of overregulation happening in the states. I desperately think we need a single federal standard. A patchwork of 50 different regulatory regimes is going to be incredibly burdensome to comply with. I think even the people who support a lot of this regulation are now acknowledging that we're going to need a federal standard.

The problem is that when they talk about it, what they really want is to federalize the most onerous version of all the state laws. That can't be allowed, either. As the states become more and more unwieldy—as it becomes more of a trap for startups that they now have to report into 50 different states, at 50 different times, to 50 different agencies, about 50 different things—people are going to realize this is crazy, and they're going to try to federalize it.

The question, I think, is whether we get preemption-heavy or preemption-light. Do we get a—

I think everyone's ultimately going to be in favor of a single federal standard. One of America's greatest advantages is that we have a large national market, not 50 separate state markets. Europe, before the EU, wasn't competitive at all on the internet because it had 30 different regulatory regimes. If you're a European startup, even if you won in your country, it didn't get you very far because you still had to figure out how to compete in 30 other countries before you could even win Europe. Meanwhile, your American competitors won the entire American market and were ready to scale up globally.

The fact that we have a single national market is fundamental to our competitiveness, and it's why winners in America then go on to win the whole world. We have to preserve that, and I think we will eventually get some federal preemption. The question will just again be whether we preempt heavily or lightly.

The second big area is infrastructure and energy. We want to help this amazing infrastructure boom that's happening. The biggest limiting factor there is going to be energy.

I think President Trump has been incredibly farsighted in this. He was talking about "drill, baby, drill" many years ago. He understood that energy is the basis for everything. It's definitely the basis for this AI boom. We want to basically get all of these unnecessary regulations, permitting restrictions, and a lot of the NIMBYism out of the way so that AI companies can build data centers and get power for them.

We can talk about that more if you want, but I think that's a really huge part of what's going to be required to win the AI race. The third area is exports, and maybe this has been the most controversial one. It really speaks to the cultural divide between Silicon Valley and Washington.

13. Federal vs. state overreach and regulation chaos

All of us in Silicon Valley understand that the way you win a technology race is by building the biggest ecosystem. You get the most developers building on your platform. You get the most apps in your app store. Everyone just uses you. Those are the companies that typically win: the ones that get all the users, all the developers, and so on.

We in Silicon Valley have a partnership mentality. We want to publish the APIs and get everyone using them. Washington has a different mentality. It's much more of a command-and-control mentality: We want you to get approved, and we want to hoard this technology. Only America should have it. This was really fundamental, I think, to the Biden diffusion rule, where the point of that rule is to stop diffusion. Diffusion is a bad word.

But in Silicon Valley, we understand that diffusion is how you win. I don't think we ever called it diffusion before; that was a new word for me. We just called it usage.

Yeah, but we understand that getting the most users is how you win. So there's a fundamental cultural clash going on right now. The way I parse it is that what we decide to sell to China is always going to be complicated because they're our competitor and adversary, and there's the whole potential for dual use.

But what we sell to the rest of the world should be an easy question: We should want to do business with the rest of the world. We should want to have the largest ecosystem possible. Every country we exclude from our technology alliance, we're basically driving into the arms of China, and it makes their ecosystem bigger.

What we saw under the Biden years is that they were constantly pushing other countries into the arms of China, starting with the Gulf states in October 2023. Basically, the Gulf states—countries like Saudi Arabia and the UAE, long-standing U.S. allies—weren't allowed to buy chips from the U.S. In other words, they weren't allowed to set up data centers and participate in AI.

Here we are telling all these countries that AI is fundamental to the future, that it's going to be the basis of the economy, and yet we're excluding them from participating in the American tech stack. It's obvious what they're going to do. The only play we're giving them is to go to China.

All of these rules basically just create pent-up demand for Chinese chips and models, and they create a Huawei Belt and Road. We're hearing that Huawei is starting to proliferate, or diffuse, in the Middle East and Southeast Asia. I just think it's a really counterproductive strategy where we're completely shooting ourselves in the foot.

The greatest irony is that the people who've been pushing this strategy of driving all these countries into China's arms have called themselves China hawks, as if what they're doing is hurting China. No, it's helping China. It's basically just handing them markets.

Our products are better, but if you don't give these countries a choice to buy the American tech stack, obviously they're going to go with the Chinese tech stack. China is out there promoting DeepSeek models and Huawei chips, and they're not wringing their hands about whether exporting chips for a data center in the UAE is going to create the Terminator, or about all these ridiculous narratives and reasons we've invented not to sell American technology to our friends.

14. Infrastructure NIMBY-ism, energy and data centers

That has ended up being, I think, surprisingly, maybe the most controversial part of what we've advocated for. But there you have it. In any event, I'll stop there. Those are some of the major pillars of what we've been advocating.

Should we go deeper on the infrastructure and energy point—what is really going to be required to get enough capacity, or what's most important in that second bullet you were talking about?

Yeah. There are definitely people who are much more knowledgeable about energy than I am and are experts in the space. Here's what I've been able to divine.

First of all, the administration—President Trump—has signed multiple executive orders to allow for nuclear power and to make permitting easier. We've even freed up federal land for data centers, hopefully to help get around some of these state and local restrictions. Obviously, the president has made it a lot easier to stand up new energy projects and power generation.

I still think, though, that we have a growing NIMBY problem at the state and local level in the U.S. that's becoming a little bit worrisome. If we don't figure out a way to address it, then it could really slow down the buildout of this infrastructure.

In terms of power, my understanding is that nuclear is going to take 5 or 10 years. It's just not something we're going to be able to do in the next 2 or 3 years. In the short term, it really means that gas is how these data centers are going to get powered.

The issue with gas is not a shortage of natural gas. America has plenty of natural gas, and it exists in enough red states that you could just build out data centers close to the source, which would be smart. The issue is that there's a shortage of gas turbines. Only 2 or 3 companies make these things, and there's a backlog of 2 or 3 years. I think that's probably the immediate problem that needs to be solved.

However, I do think that in the next 2 or 3 years, we could get a lot more out of the grid. I've had energy executives tell me that if we could just shed 40 hours a year of peak load from the grid to backup generators, diesel, and things like that, we could free up an additional 80 gigawatts of power, which is a lot.

The way it works is that the grid is only used at about 50% of capacity throughout the year because they have to build enough capacity for the peak days—the hottest day in summer or the coldest day in winter. Those are your peak days. They don't want to commit to a bunch of the capacity being used and then find out that you have a really cold day in winter and people can't get enough heat for their homes.

They can't overcommit to contracts for data centers and things like that. But if you could shed that 40 hours a year of peak load to backup power, then you would be able to free up 80 gigawatts, which is a lot. That would definitely get us through the next 2 or 3 years, until the gas-turbine bottleneck has been alleviated. Eventually, you get to nuclear. That would be very good.

I think the issue there is just that there's a whole bunch of insane regulation preventing load shedding. For example, you can't use diesel.

The EU just announced a big new growth fund, a big new public-private-sector tech growth fund to grow EU companies to scale. I shouldn't be too down on them, but I would literally think, “They do everything they can to strangle them in their crib.” If they make it through a decade of abuse as small companies, then they're going to get the money to grow.

Well, Ronald Reagan had a line about this: “If it moves, tax it. If it keeps moving, regulate it. If it stops moving, subsidize it.”

Yeah. The Europeans are definitely at the “subsidize it” stage.

Yeah, and I shouldn't be too down on them. I've always been proud to be an American, but particularly now, because it really feels like we're recentering on core American values in a lot of the things we're talking about, which is great.

Yeah. Again, our view is that, first of all, we have to win the AI race. We want America to lead in this critical area. It's fundamental to our economy and our national security.

How do you do that? Our companies have to be successful because they're the ones who do the innovation. Again, you're not going to regulate your way to winning the AI race. I'm not saying we don't need any regulations, but that's not what's going to determine whether we're the winners or not.

David, you recently tweeted that climate doomerism is perhaps giving way to AI doomerism, based on Bill Gates' recent comments. What do you mean by this? Do you mean it's going to be a major plank of the U.S. left?

I think the left needs a central organizing catastrophe to justify its takeover of the economy, to regulate everything, and especially to control the information space. I think the allure of the whole climate-change doomer narrative has kind of faded.

Maybe it's the fact that they predicted 10 years ago that the whole world would be underwater in 10 years, and that hasn't happened. At a certain point, you get discredited by your own catastrophic predictions. I suspect that's where we'll be with AI doomerism in a few years.

15. AI Doomerism, Pseudoscience and Existential Risk

In the meantime, it's a really good narrative to take the place of climate doomerism. There's actually a lot of similarity. You've got a lot of preexisting Hollywood storytelling and pop culture that supports this idea. You've got The Terminator movies, The Matrix, and all this kind of stuff, so people have been taught to be afraid of it.

Then there's enough pseudoscience behind it. You've got all these contrived studies, like the one where they claimed that an AI researcher got blackmailed by his own AI model or whatever.

Look, it’s very easy to steer the model toward the answer that you want. A lot of these studies have been very contrived, but there’s this patina of pseudoscience to it. It’s certainly technical enough that the average person doesn’t feel comfortable saying, “This doesn’t make any sense.” It’s more like, “You’re a nonexpert. What do you know?” Even Republican politicians, I think, are kind of falling for this.

So, yeah, it’s a really desirable narrative. And, of course, as AI touches every business, everyone is going to use it to some degree. If you can regulate AI, that kind of gives you lots of other things. And, like I mentioned, AI is kind of eating the internet. It’s the main way that you’re getting information. So, again, if you can get your hooks into what the AI is showing people, now you can control what they see and hear and think.

That dovetails with the whole left censorship agenda, which they’ve never given up on. It dovetails with their agenda to brainwash kids, which is kind of the whole woke thing. So, I mean, this is going to be very desirable for the left.

And this is why—I mean, look, they’re already doing this. It’s not some prediction on my part. Basically, after Sam Bankman-Fried did what he did with FTX and got sent to jail, he was a big effective altruist, and he had made pandemics their big cause. They needed a new cause, and they got behind this idea of X-risk, which is existential risk.

The idea is that if there’s a 1% chance of AI ending the world, then we should drop everything and just focus on that because you do the expected-value calculation. If it ends humanity, then that’s the only thing you should focus on, even if it’s a very small percentage chance.

They really reorganized behind this, and they’ve got quite a few advocates. Actually, it’s an amazing story about how much influence they were able to achieve, largely behind the scenes or in the shadows, during the Biden years. They basically convinced all of the major Biden staffers of this view: that imminent superintelligence is coming, we should be really afraid of it, and we need to consolidate control over it.

There should ideally be only 2 or 3 companies that have it. We don’t want anyone in the rest of the world to get it. What they said was, “Once we make sure that there are only 2 or 3 American companies, we’ll solve the coordination problems.” That’s what they consider to be the free market: we’ll solve those coordination problems for those companies, and we’ll be able to control this whole thing and prevent the genie from escaping the bottle.

I think it was this totally paranoid version of what would happen, and it’s already in the process of being refuted. But this vision fundamentally animated the Biden executive order on AI. It’s what animated the Biden diffusion rule.

And, Marc, you’ve talked about how you were in a meeting with Biden people and they were going to basically ban open source. They were going to anoint 2 or 3 winners, and that was it.

Yeah, they told us that explicitly. They told us exactly what you just said. They told us they were going to ban open source. When we challenged them on the ability to ban open source because we’re talking about math—mathematical algorithms that are taught in textbooks, YouTube videos, and universities—they said, “Well, during the Cold War, we banned entire areas of physics and put them off-limits. We’ll do the same thing for math if we have to.”

And you'll be happy to know that the guy who actually said that is now an Anthropic employee.

No, that’s exactly right. Literally, the minute the Biden administration was over, all the top Biden AI employees went to work at Anthropic, which tells you who they were working with during the Biden years.

This was very much the narrative. You had this imminent superintelligence, and one of the refrains you heard was that AI is like nuclear weapons and GPUs are like uranium or plutonium. Therefore, we need an international atomic energy commission. The proper way to regulate this is with an international organization like that. Everything would be centralized and controlled, and they would anoint 2 or 3 winners.

16. The DeepSeek Moment and China's Relative Progress

This narrative really started to fall apart with the launch of DeepSeek, which happened in the first couple of weeks of the Trump administration. If you asked any of these people what they thought of China during the time when they were pushing all these regulations—specifically, if you asked, “Wait, if we shoot ourselves in the foot by overregulating AI, won’t China just win the AI race?”—they would have said that China was so far behind us that it didn’t matter.

Furthermore, they said, completely without evidence, that if we slowed down to impose all these supposedly healthy regulations, China would just copy us and do the same thing. I think it was an absurdly naïve view. If we shoot ourselves in the foot, China will just be like, “Thank you very much. We’ll take leadership in this technology.” Why wouldn’t we?

But this is what they said. And when the Biden executive order on AI was crafted, there was no discussion whatsoever of the China competition. It was just assumed that we were so far ahead that we could basically do anything to our companies and it wouldn’t really affect our competitiveness.

I think that narrative really started to fall apart with DeepSeek at the model level. Back in April, Huawei launched a technology called CloudMatrix, in which it compensated for the fact that its chips individually aren’t as good as Nvidia’s chips by networking more of them together. It took 384 of them and used its powerful networking to create this rack system, CloudMatrix.

It was demonstrated to show that, yes, Nvidia chips are better and much more power-efficient, but at the rack level, at the system level, Huawei could get the job done with these Ascend chips and CloudMatrix. Again, I think that showed that we’re not the only game in town on chips, which means that if we don’t sell our chips to our friends and allies in the Middle East and other places, Huawei certainly will.

So, I think it’s been one revelation after another in which we’ve learned that a lot of their preconceptions and beliefs were wrong. We’ve talked about the fact that the markets ended up being much more decentralized than they ever could have predicted.

I would also say one other thing: They also believed there would be imminent catastrophes that haven’t happened. This is kind of the equivalent of the global-warming thing, where we’re all supposed to be underwater by now. They were saying that models trained on, I think, 10²⁵ FLOPs or whatever were way too risky.

Every model now at the frontier is trained on that level of compute. They would have banned us from even being where we are today if we had listened to these people back in 2023, just a couple of years ago. It’s really important to keep in mind that their predictions of imminent catastrophe have already been refuted.

Things are moving in a direction that I think is very different from what they thought in the first year after the launch of ChatGPT.

Right. So, David, just to come back real quick while we still have you on crypto: The administration—and I think the country—had a significant victory earlier this year with the president signing the stablecoin bill into law, which was the GENIUS Act.

I’ll just tell you what we see as the positive consequences of that law: They’ve been even bigger than we thought. I would say that’s true both for the stablecoin industry—you now see a lot of financial institutions of all kinds embracing stablecoins in a way that they weren’t before—and, more broadly, as the phenomenon spreads in America, which is, by the way, in the lead and doing very well there.

17. Crypto clarity: Genius Act and next-gen regulation

It’s also a signal to the crypto industry that this really is a new day. There really are going to be regulatory frameworks that make these things possible, that are responsible, but also make it possible for this industry to flourish in the U.S.

As you know, there’s a second piece of legislation being constructed right now: the market-structure bill called the CLARITY Act, which is phase 2 of the legislative agenda. I wondered if you could tell us a little bit about your view of the importance of that bill and how you think that process is going.

I think it’s extremely important.

As you mentioned, we passed the GENIUS Act a few months ago, but that was just for stablecoins. Stablecoins are about 6% of the total market cap in terms of tokens, so 94% are all the other types of tokens. The CLARITY Act would apply to all of that and provide the regulatory framework for all those other crypto projects and companies.

If we could be sure that Paul Atkins, or a person like Paul Atkins, was always at the SEC forever, then we wouldn't necessarily need legislation because they're already in the process of implementing much better rules and providing regulatory clarity. But the truth is that we don't know for sure. And if you're a founder who's trying to make a decision now about where you're going to build your company, you want to have certainty for 10 years out, 20 years out.

We want to encourage long-term projects. And so, again, I think it's very important to codify the rules that first provide the clarity and then make sure there's enough stability around them, and sort of codify those rules in legislation. That's the only way that you provide that long-term stability.

I think that we will get the CLARITY Act done. Like you mentioned, it passed the House with about 300 votes, including about 78 Democrats, so it was substantially bipartisan. I think it will ultimately get done. It's now going through the Senate.

We're negotiating with a dozen or so Democrats. We have to get to 60 votes, so that's the hard part. Under the filibuster, we've got to get 60. But we're negotiating with about a dozen Democrats, and I do think that we will ultimately get to that number.

By the way, we ended up having 68 votes in the Senate for the GENIUS Act, including 18 Democrats. So I do think that even if we just get two-thirds of the number of Democrats that we got for GENIUS, then we'll be fine on CLARITY.

This will provide the regulatory framework, again, for all the other tokens besides stablecoins, and I think it's just a critical piece of legislation. This would ultimately complete the crypto agenda, where we've moved from Biden's war on crypto to Trump's crypto capital of the planet.

I think the industry will have the stability it needs and can just focus on innovating. There'll be rule updates and things like that, but we'll fundamentally have the foundation for the industry in place.

On the GENIUS Act, President Trump really made that bill possible. First of all, it was his election that completely shifted the conversation on crypto. If a different result had been reached, we would still have, again, a Gary Gensler at the SEC. The founders would still be getting prosecuted, we wouldn't know what the rules are, and Elizabeth Warren would be calling the shots.

President Trump's election made everything possible, and his commitment to the industry and to keeping his promises during the election has made all of this possible. But also, he got directly involved in making sure the GENIUS Act passed. The legislation was declared dead many times.

I saw with my own eyes that he was able to persuade recalcitrant votes, twist arms, cajole, and charm. He ultimately got it done, and I think the CLARITY Act will have a similar result. People are always prematurely declaring these things to be dead. There are a lot of twists and turns in the legislative process.

It's definitely true that you don't want to see the sausage getting made, but I think we're on a good track right now.

18. The Evolving Democratic Party

Good. Fantastic. Great. Pete Buttigieg went on All In recently, and you guys talked about the left's identity crisis. He's hoping for more of a moderate center-left to emerge.

At the same time, we see Mondaire in New York. I'm curious what you think: Is there a more moderate presence in the future of the Democratic Party, or is it this Mondaire-style woke populism?

It certainly seems to me that Mondaire and woke socialism are the future of the party. That's where all the energy is in their base. I don't want that to be the case. I'd rather have a rational Democratic Party, but that seems to be where their base is, where the energy is.

You don't really hear Democrats within the party trying to self-police or distance themselves from that. You saw all the major figures in the Democratic Party endorse Mamdani. So, yeah, that's where that party seems to be headed.

I think partly it might be a misread. It's sort of a partial reaction to Trump, where they feel like establishment politics has failed, and so they need a populism of the left to compete with a populism of the right. I think that's maybe part of the calculation for why they're going in this direction.

But I firmly don't think it works. I don't think socialism works. I don't think the defund-the-police and empty-all-the-jails policies work. I think we're about to get another case, a teaching moment, in New York. Unfortunately, it's not going to be good for the city. We've seen this movie before.

That's where the Democratic Party appears to be. I don't completely get it. Other people have made this observation, but they do seem to be on the 20% side of every 80/20 issue: opening the border, being soft on crime, releasing all the repeat offenders, and just this anti-capitalist approach, which I think will be disastrous for the economy.

This is kind of where the party's at right now. It's a little scary because it means that if we lose elections in places where we do lose elections, you could end up with something really horrible. We're not just playing between the 40-yard lines anymore in American politics, and that is a little bit scary.

19. The Future of San Francisco

If it weren't for Donald Trump, I think in a way we might already be there. We have to make sure that this Trump revolution continues.

Lastly, we just talked about New York. Recently, on an episode of All In in San Francisco, you endorsed bringing in the National Guard. Marc Benioff had his comments; he sort of went back and forth on those comments.

I'm curious, speaking of teaching moments, if you see San Francisco as savable in some sense, and what needs to be true to get there, if so.

Daniel Lurie is the best mayor we've had in decades, so I think he's doing a very good job within the constraints that San Francisco presents. Unfortunately, we have a weak mayor in San Francisco. I don't mean him; I just mean the way it's all set up.

The Board of Supervisors has a ton of power, and over time they've been able to transfer power from the mayor to themselves. Then, of course, you've got all these left-wing judges.

It's amazing to me that there's a case right now. This is a case that galvanized me several years ago: the case of Troy McAlister, who was a repeat offender who killed 2 people on New Year's Eve, I think it was 2020. He was arrested 4 times in the year before that, and he ended up killing these 2 people. He had a very long criminal history.

He had committed armed robbery before and had stolen many cars. He should have been in jail. He should not have been released, but he was basically released thanks to the zero-bail policies of Chesa Boudin, who was then the district attorney and whom we got recalled.

There was a huge outcry. Even in San Francisco, for there to be a recall of a politician, you've got to be seriously left-wing to alienate San Francisco. Chesa Boudin managed to be so far out there that he alienated even San Francisco.

Yet, I don't know why Troy McAlister isn't sentenced already to 20 years or more in jail. His case is still pending through the courts; it's never-ending. There's a left-wing judge who's considering just giving him diversion. Basically, that means you just get released, maybe with an ankle bracelet or something.

That's insane. That's what we're dealing with in San Francisco: crazy left-wing judges who want to release all the criminals. So I just wonder: Is Daniel up against too many constraints?

I know he doesn't want the president to send in the National Guard, but maybe ultimately it would be helpful. In any event, I think the president has agreed to hold off on that. Daniel had a good conversation with the president and asked him to hold back, and the president agreed and is giving him time to implement his solutions.

If Daniel and his team can keep making progress and fix the problems without the National Guard having to come in, then so much the better. We'll just see. I know he wants to fix the problems, and like I said, he's the best mayor we've had in decades.

It's just a question of whether he'll be too constrained by the other powers that be in the city.

David, thank you so much for coming on the podcast.

Yeah, good to see you guys.

Fantastic. Thank you, David. It was a great year. And thank you for the work. We, as much as anybody, appreciate the work that you've done to fix the things in the past and to put us on a great road to the future.

Yep. Well, thanks. I appreciate what you guys have done as well. So thank you for your support and everything you're doing. I appreciate it.

Definitely.