Anthropic IPO 面临风险，Meta 的 Muse 爆发，Token 价格下跌，开源份额上升，对齐失效

Welcome back to episode 290 of the number one podcast in the world. It's the All-In podcast. With me again, it's the core 4. David Friedberg is here. Chamath Palihapitiya and David Sacks are calling in. I am your host, Jason Calacanis, and we have a full docket today.

I'd be remiss if I didn't ask everybody for their fondest memories of the 5th annual All-In Summit, produced by Mr. David Friedberg. David, do you have any thoughts? How do you rate your performance as executive producer of the All-In Summit?

You know what we didn't do? What we forgot to do at the end of day 2 was give a big shout-out to our production team—to Kimber, Lisa, and Nick. I thought we should have brought them out, but I just want to give them a big shout-out.

The team backstage absolutely crushed it, as always, this year. And then I had a great time. Actually, my favorite part is going to Universal Studios. You just get to have a good time. It's great—just great.

Bring them out for the standing ovation.

I'm reliving my childhood. Shabbat with no trauma.

By the way, shout-out to our friends at Dongi. I got one of those coffee machines that we had at the event. I put it in my kitchen.

You know what I got? I don't know if you guys saw, but there was a water machine in the VIP area. Well, that is a zero-plastic water machine. It goes through super filtration. No microplastics in the water. I have one now in our office. Thank you very much to them. I appreciate it. Amazing company.

Sacks, do you have a favorite moment from the show?

I think the highlight of the show, the thing that we'll remember for a long time, was the president calling in during Jensen's talk. Jensen's overall content and demeanor really helped calm down this national panic over AI. It felt like we were at the center of this national conversation.

I think it happened very organically. You remember, there's basically been this campaign for the last few weeks leading into the summit to convince everyone that humanity was going to go extinct unless we did exactly what the Democrats wanted and stopped all AI development. Then Satya came out and kind of calmed everyone down, and Jensen did the same thing. The president called in and did what he does best, which is seize total control of the narrative that I'm sure these doomer groups had been carefully planning for weeks, and kind of thwarted that whole operation.

Nerds, enough. Get back to work. You have a superintelligent Trump. Intelligence is now the term. We're going to call it Trump intelligence going forward.

But I think you make a good point. A lot of people wanted to know, Friedberg, if this was planned. Was President Trump's call planned? Friedberg, take us behind the scenes.

Basically, Jensen was connecting with him right backstage and said, “Hey, call me back.” Then he called him while he was onstage. There was some texting going on.

Yeah, it was completely spontaneous. It was not a planned thing where Jensen was like, “Hey, let's do a whole thing.” It was pretty amazing to have the conversation that Jensen was waiting for from the president happen live. Clearly, this was the topic du jour in D.C., as well as in the news and the media, and it was amazing to see it all happen onstage with us.

Just to be clear, I thought it was a bit. I was like, “Well, put him on speakerphone.” Jensen says to me, “How do you do that?” I just handed him—I always keep an extra mic there in case somebody's mic goes out—and said, “Put him on speakerphone. Put the microphone there.”

President Trump, the funniest president in history, the funniest politician in history—did you notice, Sacks, that he drops a one-liner through a speakerphone perfectly?

1. Reacting to AI chaos: Liability, competition, and rebranding frontier "labs"

Jensen makes the greatest chips. Everybody tries to reverse-engineer them. Impossible. Impossible. But he can't figure out how to press the speakerphone button. I was just like, “How does he do it? How does he do it?” He just drops the perfect line.

But, Chamath, in all seriousness, the nerds were losing their minds last week, and the adults came. Daddy came. Elon came. Jensen came. All the adults in the room came and just said, “Listen, if you have software that's defective and not ready to put in the world, maybe don't put it in the world and make it a little bit—as you like to say—less brittle.” So, your thoughts on where we are in this national discussion, which went from “slow down” to maybe just release non-brittle, resilient software?

I think there was this classification that happened last week between up-and-coming organizations that want to call themselves labs but are, in fact, companies that have P&Ls and shareholders and have to abide by the same rules that everybody else does, and the sophisticated companies that have lived under scrutiny for many decades and have learned the hard way the process of making things robust and the cost of making things not robust.

So what do I mean? On the one side, you saw—in many ways, it's like, “Curious times create curious bedfellows.” You have the cohort of Satya, Jensen, Sacks, Zuck, Elon, the president, and Lina Khan. We have product-liability laws in the United States. They're meant to prevent companies from doing things that they're not supposed to do, especially if they know that those things could be harmful.

Strange fellows indeed.

On the other side, it's folks making the thing who were, I think, being a little naive and basically saying, “We're not a company. We're a lab. Protect us.” I think the mature response is, “You should have internal controls, and you should robustly test things.”

What Zuck said earlier and what Elon has said in the past is that we're just under a lot of scrutiny. When things aren't perfect, we slow things down. Elon has done that with FSD. He could have been at FSD 3 or 4 years ago, but he has publicly stated, “We can't afford it, because every accident that Tesla incurs is magnified 1,000 times more than what any other company would incur, and so we just have to go slower.”

Zuck last week said, “We were about to roll out Muse,” which looks like one of the most successful product launches in recent memory and has huge implications, actually, which we should talk about. He had to slow-roll it. He and Alexandr Wang, who I was talking to last night, had to slow-roll it for a couple of months to get it perfect.

It's all within your control. It's just what you choose to do. I think that was the most interesting observation. We have to stop calling these companies labs. They're companies, and we have to judge them by the same standards, including product liability.

Yeah, and there was a lot of coverage of this moment, Sacks. Interestingly, some of the fake-news media were like, “On a popular podcast,” and then they played the clip, but they're just afraid to say the name of the podcast. Pretty hilarious.

I think Chamath is getting us into the first topic here already, which is whether we're going to have individual responsibility as our North Star or some sort of collective action. The thing that always works the best is individual responsibility, whether it's at the level of individuals or companies. The relevant decision to make in every case is, “Should we ship this product?”—not, “Can we get the whole world to agree?”

The problem with talking about these collectivized approaches is that they deflect accountability away from the actual decision-maker. I think people sort of feel this. This week, you had Dario and Sam go to the United Nations and say we need some sort of global AI governance. It kind of reminds me of the billionaires who fly to Davos in their private jets and then rail against climate change. There's something fundamentally unrealistic and out of touch about it.

Hypocritical.

Hypocritical. And it's like, guys, look, what we need is for you to make smart decisions about what to release. If a product is unsafe—which is to say, it's unreliable or behaves unpredictably—then don't release it. Stop looking to some sort of global governance as the answer.

If you guys don't get this decision right, it's not going to matter what the United Nations does. By the time the United Nations agrees on anything, we're going to be much further along in this whole journey. There's something just very weird about the way that the principal actors in this keep talking, because it really feels like they're trying to push off the responsibility.

Well, there's a theory going around here, and I think it is worth us discussing here, that the Frontier Labs are covertly—

They're companies.

They're companies. Yes. Let me just finish the—

No, no, no, but you made a really good point that I want to make sure we don't miss.

Okay.

They are corporations.

Yes. Yes. And when you call them labs, they sound like nonprofit entities, and that's not true.

The corporations making these products are looking for President Trump to specifically give them what the internet companies had with Section 230, which is to shield them from liability. There is a consideration inside the Trump administration—the 47th administration—that maybe there could be some leverage put on these companies: They would give 10% of their equity to the sovereign wealth fund of the United States, and then that would result in them being given some pass when it comes to their liability.

That's an insane thing to do. It's like giving some plane company that gives 10% of its equity to the U.S. government the ability not to be responsible for its planes falling out of the sky.

But what is your thought on that conspiracy theory? I'm sure you've heard it, Sacks—that they're looking for—

You're making it sound like this is the idea of the Trump administration. That's not true.

2. Anthropic and OpenAI postpone IPOs: liquidity risk and open source pressure

No, not the Trump administration. This is what the frontier labs and corporations are pitching to them, and the rumor is that there might be some people who are open to that in the administration. Again, this is all a back channel.

Here's what we know: In the past week, virtually every Trump administration official came out and said some version of what I just said: First and foremost, the AI companies have to take responsibility for their own products, and we will not waive product liability. I think Speaker Johnson said it, the congressional leadership said it, and Scott Bessent said it in his most recent interview.

3. Meta launches Muse, AI "alignment," Oracle's Force Majeure

Everybody said that if your game is to get a waiver of product liability or antitrust liability, if you're trying to form a cartel, if you're trying to collude, forget about it. We're not going to do it. So, I think they completely threw cold water on that whole idea.

And remember what President Trump said—he just tweeted today: We have the DOJ. That's a guardrail. The administration is a guardrail. The administrative state, civil lawsuits, criminal lawsuits—all these things are there as potential liability to make sure that you make the right risk-reward trade-off.

And again, at the end of the day, if those guys release an unsafe product, it's not going to matter what the United Nations agrees on. So, we need them to internalize the responsibility.

Are they? Listen, come on. Be candid with me. You got the inside goods here. Are they trying to lobby for that? Do you think they are lobbying to get this protection? Yes or no? Come on.

Well, I have heard this. This is not firsthand experience. They have not asked me, “Hey, we want a product liability waiver,” but I have heard from people in Washington that they have been seeking something like that.

It's hard to pin down, right? Yeah. Okay, fair enough.

Friedberg, your thoughts here? As we're going to get into the first story with the token maxing and the prices going down, I'll give you the final word on what America should do in terms of dealing with the corporations that are pretending to be labs, which cannot take ownership of their own Frankenstein creations.

I thought J.D. Vance had a spectacular appearance. He was crisp and looked quite presidential, if I was going to describe it. He threw a lot of fastballs—

And he knocked them out of the park. By the way, home-run hitters want fastballs. You can't hit a fastball if it's out of the strike zone.

So, I gave him the 3 hardest questions I could think of, and he proceeded to knock all 3 over the fence.

I think JD was excellent.

Listen, on the Sunday shows, they played the clips of the questions I asked him. Bang, bang, bang. Friedberg, right over the fence.

But he said, basically, if you're creating Frankenstein, here's an idea: Stop. And then he added, if you've already let the cat out of the bag, create anti-Frankenstein. Basically, create the safeguards.

Yeah, it makes total sense.

I think Zuck gave a good interview on this. They took their time doing this Muse release this week. They said they wanted to get it right. They wanted to make sure that it was safe, that it didn't give people capabilities to do bad things. That's the responsibility that they took as a corporation. Anyone that's outputting a model should take the same responsibility.

Pretty obvious.

This is an organization that understands it, having paid a series of fines from Cambridge Analytica to recently—

We have a lot of laws that make it illegal to do a lot of things.

All right, let's move on here. We have laws.

By the way, can I add one philosophical point to this? I think that implicit in what I'll call the Dario position—the regulated free market that we have—is the fundamental contention that it will lead to a race to the bottom, that competition jeopardizes safety. It's in all of his blog posts, and many other people reflect this idea that competition equals a lack of safety.

That is implicitly a critique of our economic system. It's basically a left-wing critique of our system, and it goes back a long way. I remember back in the days of the Cold War, when we really had this dual between 2 different systems: Do you want the free-market system, or do you want the communist system?

A lot of people made the claim that the capitalist system would do things like jeopardize safety and beauty, that our buildings would always be uglier because the greedy capitalist wouldn't want to spend a dime on beautification or good architecture, things like that. In fact, it turned out that the Soviet system was the one that produced the cold, gray landscapes. Why? Because people want beauty. They want safety.

Therefore, if you have a well-functioning market economy, it gives people what they want. Obviously, the American system won out in that battle. So, I fundamentally disagree with the idea that competition means a lack of safety.

As Zuckerberg and Jensen were saying, there are huge incentives on both the upside and downside to be safe. On the upside, no customer wants a product that is unpredictable and unreliable, that doesn't do what it says. No enterprise wants an agent that's going to leak their data or hack a competitor.

On the downside, as the president made clear, they've got tons of liability: product liability, civil liability, administrative liability, and criminal liability. So, I firmly disagree with this idea that competition is a bad thing. Competition is a good thing that can be harnessed to get the right result.

And just to build on that, Chamath, our friend Nikesh from Palo Alto Networks, a friend of the pod, released Unit 42 in the past week or so. He's building—there's a business opportunity here—to build a product for continuous cyber defense. Also, our friend George Kurtz from CrowdStrike released Falcon, his new product for cyber defense.

Chamath, if these frontier labs and corporations were so concerned about cybersecurity, why didn't they ever release a cybersecurity product? Palo Alto Networks and CrowdStrike are on it, like they will. That's probably more in the second topic we'll talk about, which is that as the models essentially cluster in terms of capability, there's no differentiation, and so they'll be forced to move up the stack. But let's wait on that for 1 second. You have a point here you want to make, and then we'll go to the first.

I do. The point I wanted to make is more philosophical and sociological. The last time we heard the word “lab” in the social consciousness was during COVID, and there was a lab that was responsible for a leak of a virus that then killed 15 million people all around the world and caused $45 trillion of damage. It's the Wuhan lab.

I think it's important to recognize that there has been zero transparency or culpability around this. Part of why people are comfortable calling themselves a lab is because of that fact. A lot of these young people who work at these places lived through this and saw absolutely no responsibility taken by the lab that caused this leak.

Now we're talking about leaks of different kinds, and we have organizations and corporations that try to fashion themselves as labs as well. But they're not labs. These are corporations. I just want to remind everybody: These are for-profit corporations that have absorbed hundreds of billions of dollars of debt and equity, that have thousands and thousands of shareholders, and that have trillions of dollars of market cap.

If I were them, I would not want to call myself a lab at any cost. And I think if you saw the recent speeches that they gave at the United Nations, they're still using that terminology.

Yeah.

They are not labs. These are for-profit corporations that are subject to product liability risk, and they should act accordingly.

Yeah. We've had a conversation here for a year about this. If this stuff is so dangerous, you don't have to release it. My proposal is: Know your customer and give logs.

Do not compare yourself to the Wuhan lab. It's bad PR, but I do think it's part of their virtue-signaling routine. They just don't want to say that they're companies with a for-profit incentive. They always want to say that they're either a research institution, or they say things like, “We're a public benefit corporation,” which no one really knows what the hell that means.

So, it's this patina of virtue signaling that obscures what's really going on.

It tries to obscure the word “corporation,” but legally you're not allowed to hide it.

We are all running corporations. All-In, albeit very small but mighty, is a corporation. I think we should call it—

It's the All-In Lab.

This is the All-In Lab, where we discuss different formulas.

No, we're experimenting with ideas. This is a meme workshop.

Lab. These ideas are getting so dangerous that we don't want any liability for our opinions. Welcome to the All-In Lab podcast.

Friedberg, is there a mindset in scientists' minds that needs to have a hard stop, and then the corporation kind of takes over? Maybe they could put the labs on the side here and say, “Hey, these are the adults doing the corporation stuff.”

What are your thoughts here if you were going to architect how Anthropic operates in the world? And maybe you could speak to scientists' views of the world versus capitalists and management teams?

I don't have a good answer for you on that.

Okay. Well, you're doing it. You're the answer.

4. Non-frontier performance and cost: What this means for frontier companies

Let me kick off the first topic. What I wanted to do was take a step back and look at the insane flurry of model releases happening. I think the pace of model releases on both the open-source side and the premium side—the Anthropics and Groks of the world—is unfathomable right now. The list is insane.

Just in the last 10 days: on September 9, DeepSeek 4.1 Flash gets released at $0.20 per million output tokens if they're hosting it. The key-value cache from V1 of DeepSeek—remember what a big deal DeepSeek was when it came out—was like 390,000 bytes. Now it's 890. Insane drop, insane efficiency improvements.

On September 20, Qwen, which is the Alibaba model, comes out with version 2.1. It's an open-weights model that outperforms Nano Banana 2, which is Google's big breakthrough, insane image-generation model. This thing is free and open source. You can put it on your desktop computer and run it at home to generate images for free, and it's made by Alibaba.

Xiaomi put out MiMo on September 22, 2 days ago. MiMo Pro performs on par with Claude's Opus 5, which is this insane model that everyone thought was going to change the world and break the internet, and GPT56 Soul across most benchmarks. It's a 309-billion-parameter model. It is entirely open. You can download this model, put it on your server, and run your own instance without paying anyone any money except for the cost of running the server.

This model called Bonsai 2 came out on September 17 from Prism ML. This is actually a fork of Qwen, which is another one of the Alibaba models. It's a 27-billion-parameter model with 98% of the performance of the big Qwen model. It's 5.9 GB in size, and you can run it on a local computer. You can have an NVIDIA chip or a good Mac Studio computer, and you can run this insane model on your desktop computer for free.

Then the big, crazy models—the big closed-source models: Anthropic releases Opus 55 on September 22. OpenAI's Astro comes out with Soul and Luna on September 23. Grok 4.7 comes out on September 21. Meta Muse comes out on September 22. Any one of these stories would have broken the internet a year ago, and they all happened in the last 10 days.

You now have an entire catalog of open-weights models to generate visual information and do VLA, where you can actually instruct the model and it can control a robot. You have models far beyond the most advanced LLM from a year ago that are open weights, that you can install and run on your desktop computer.

We are past the point of Bernie Sanders saying, “Put the superintelligence away. It can never be allowed in the real world.” This stuff is out. The open-source, open-weight models are free. They are on the internet. You can go and download them and run them on your computer at home.

I don't know in what world we think we're going to have a police state that's going to come bang down your door, take your desktop computer from you, and test it to see if you're running an LLM on your local computer. The efficiency wave is here, and the pace at which these models are coming out now—all of this stuff happened in the last 10 days—is just mind-blowing.

We're now in this kind of frontier space where things are accelerating. The performance improvements and the cost decline are making AI so performant and so affordable that it is going to be ubiquitous, and we are all going to benefit from it. This is not just about Anthropic and OpenAI, with the closed-source world making billions of dollars selling enterprises hosted models. There is so much proliferation right now.

I thought it was worth taking a step back to recognize where we are, because in the context of, “We have to stop the superintelligence. We have to shut down the data centers,” you don't even need data centers to do 90% of what you can do with AI. You can download it on your computer for free today and do what was the most advanced technology in human history less than a year ago. It's all here today.

I thought it was worth taking a beat to step back and make sure we all recognize that as we get into this debate about these labs or corporations. I don't know if it even matters anymore. I think open source is here, and I think the pace of improvement is such that this is only going to drive productivity.

Maybe we are seeing the AI wave underway right now—the productivity wave, the productivity boom.

Yeah. To consumers, they don't care what model they're using. We've spent the last 3 years geeking out about this, and then Grok comes out and Muse comes out from Meta. I've been using both of these products, and it's the first time that a normie—a normal person in the world, not meant to be derogatory in any way—can get value from AI.

If you look at how we all looked at this, running corporations and investing in them, we took a different approach: maybe this is going to retire some positions, maybe this is going to add some positions here or there. But now, with Muse, you can connect to everything instantly and actually get real work done in the world if you're a normal person who wants to get the benefit from AI.

We will be turning a corner, Friedberg, and I think you laid it out perfectly. We're going to turn a corner at the end of this year, where moms, our uncles—people—are going to start using Muse and Grok, which are growing very quickly, to solve real-world problems. They're going to go, “Wait a second. I could never afford a chief of staff. I could never afford a personal assistant, an executive assistant. Now I get a free executive assistant.”

It's going to blow people's minds, Chamath, in the next couple of months when they start using these products, and the bottom half of America in this K-shaped recovery experiences some benefits. Your thoughts on what's happening here with these models? Have you started playing with them? Because you were originally a technologist. Have you loaded these models on your computer? I'm playing with them on my MacBook Pro, and it's awesome.

Let me give you the answer by explaining something. I think what we're seeing is that these models are converging and clustering, which is to say that they're now within the margin of error. There are so many of these models that are roughly the same at different price points. So where is the edge?

There's still an edge, and the edge is in the harness that you use to wrap the model. If you think about it, you have a model that's like a brain, but you need to embody that brain with arms and legs and hands and eyes to do work. That's how you make it agentic; that's what that phrase means. There are all kinds of different harnesses that the industry has invented to make these models more agentic.

At 8090, we have a labs function, and we've been ripping and teasing apart all the different models with all the different harnesses just to see how they perform. They vary wildly in terms of cost and quality. To build on Friedberg's point, you have the models that are converging, but the harnesses are wildly different, so their performance is all over the map. Now companies and people have to navigate what they're willing to pay for.

I sent this tweet to Nick. If you can just put it up here, I don't know if this is instructive or accurate, but what this shows—and this is the case for Anthropic, but it's probably the case for a lot of the closed models—is that you're starting to see this massive revenue concentration, where a few folks consume all of the highest and most expensive tokens.

I think if you take what Friedberg just showed to its logical conclusion, those folks at the top of this table are going to be under a lot of internal pressure, whether it's from their finance teams, their CFO, or shareholders, to explain why, as this clustering happens, they are not moving down to cheaper and cheaper models.

You may choose to go down to cheaper and cheaper Anthropic models, so Anthropic keeps the revenue, but there's also a risk that you may do what Friedberg just said and go to an open-source model and host it yourself. We're in a very complicated part of the cycle.

The first version of the AI trade was relatively simple: you're just selling tokens, wrapping the tokens, and passing them through. There was enough disparity that there was value. I think that's going away. It will force OpenAI and Anthropic to go up the stack. There is no choice. You can't sit there and serve a token that has diminishing value, because then all the subsequent value is getting absorbed and made by people wrapping your intelligence token. You will have to do that work.

Jason, that's why they will have to go and do cyber. They'll have to go and do law. They'll have to go and do customer support—all these things that we weren't sure whether they were going to compete in. These trends may force them to do that.

Yeah. I think then that sets up a very interesting setup for their IPO. Just to back all that up, Sacks—

The frontier lab corporations, Anthropic and OpenAI, both released models this week at 50% lower token prices, and their IPOs are looking like they're both going to get delayed.

Anthropic is targeting a $2 trillion valuation, and OpenAI is targeting $1.2 trillion. Sam already said OpenAI plans on doing an IPO in 2027, in part because of safety concerns. Anthropic, which was supposed to file and go out in October—the report from The Wall Street Journal is that the IPO will happen in November or may be delayed even more.

Anthropic peaked at 96% earlier this month that it would go public this year, but now it’s fallen to 76%. Is this like a sell signal here, Sacks? The leading model companies need to go public to get capital to do their buildout, and now they’re delaying their own IPOs.

Please explain how any of this makes sense. If you were a major investor in one of these companies or on the board of one of them, what would you be doing or instructing these founders to do to get these companies to go public and get the capital they need to continue the AI buildout?

I think that we know from our friends that the investors in Anthropic are pulling their hair out about what’s happening right now. Forget about the fake whistleblowers who were there for a few weeks and then left. You have current leadership in the company saying that there’s a greater than 10% chance of causing human extinction, and their own product is unsolved in this area. Think about the risk factors that’s creating. How do you IPO with your own management of the company saying that?

Then you’ve got what I don’t know how to call except corporate schizophrenia. You’ve got Dario putting out this essay saying we need to pause the frontier days before they launch Claude 5.5 and set a new frontier. He must know when he’s writing this essay that they’re about to release a product that will extend the frontier. At a minimum, it’s hypocritical, but maybe it’s even worse than that.

Then you have them publishing an essay about the biorisk associated with AI and how that needs to be stopped. Meanwhile, they announce a new biolab in San Francisco, a new wet lab.

They started a lab after telling us that there’s a 10% chance of us all dying. Sacks, let’s just pause here for a second. Is this a dereliction of duty as the chief executive officer of Anthropic to sabotage your own IPO? You’re supposed to be acting on behalf of your shareholders.

This, to me, seems like if we were the board—if the 4 of us were somehow magically the board of Anthropic—we would replace the CEO. We would put Dario in the lab, and we’d find a CEO who knows how to run the company and not sabotage it. He’s literally sabotaging his own company.

I think there’s an element of that, and I think investors are pretty nervous about giving the founders, or Dario specifically, super-voting shares. There’s a story in The Information about that, that there’s an active debate right now about whether to give them super-voting shares.

Explain what those are, please, and why, in this context, this is critically important. Our audience may or may not know what those are.

Supposedly, the founders of Anthropic only have about 2% ownership each of the company, which in the grand scheme of things is a relatively low number compared with what other founders typically have at the time of an IPO. It’ll still be an enormous amount of money—tens of billions of dollars—but that is the case.

Super-voting shares basically separate the economic shares from the voting shares in the company and give certain people the right to control the company even if they don’t have the economic ownership that would usually be necessary. It’s a pretty common thing. Larry Page and Sergey Brin, I think, may have been the first prominent founders to get super-voting shares in Google, and Mark Zuckerberg did as well.

It can be a stabilizing thing for the company because it means that, as long as the founders are there overseeing things, they’re not going to be subject to takeover attempts or things like that. But on the other hand, it requires a tremendous degree of trust in those founders because they effectively can’t be removed no matter what they do. You just have to wonder if, in this case, that’s earned and warranted.

I have to say that even for someone like me who isn’t necessarily advocating for a lot more regulation, I kind of get nervous about the idea of Anthropic opening a wet lab in San Francisco.

Great timing.

I was thinking about moving back, and I don’t know if I want to be—I mean, can you just put it a little further away? I think we’ve all seen this movie before.

Yeah. Offshore, maybe. I don’t know. Put it on an oil rig or something.

What I would say is that if I were on the board, I would keep Dario as CEO. The reason is that he’s clearly demonstrated that he can build a unique culture and position against the otherwise most formidable corporate competitor, which is OpenAI. He’s done that, and he’s come from behind. I don’t think you should undervalue how hard that is and what that takes.

He’s built essentially the greatest business ramp of all time. I think the thing that I would probably do if I were on the board is fix what Sacks just said, which is that there’s now a bunch of incremental liquidity risk to the company. If you take the 20-year view on Anthropic, that corporation needs hundreds of billions of dollars to fulfill its ambitions. I would solve for getting that as risklessly as possible.

I think there are only 2 things that you can do. The first is you have to throw everything, including the kitchen sink, into disclosures. The problem is that that will make what is an otherwise dense and impenetrable S-1 even more dense and impenetrable. You can superficially cover your regulatory risk, but the way that it actually gets translated in the market is that you’re going to have to water down the expectations of the IPO sellers and create a much larger margin of safety for the IPO buyer.

That’s effectively how you find a market-clearing price for this, Jason, which is to essentially say—and bear with me, I’m just giving out rough numbers—even if you think that they’re on a $100 billion run rate, what otherwise, in a very clean-sheet IPO, maybe would have been a $2 trillion market cap, now you’re at $1 trillion or less.

The reason is that the hedge funds and the pension systems and the long-only mutual fund complexes, which have a fiduciary responsibility to their sources of capital, will justify all the risks by essentially asking for a lower and lower price. I think that’s what you’re going to see. I think you’re going to see a much more moderate and reasonable IPO valuation here, and it’s probably going to clear at a much lower price than anybody thinks.

I do think that what Sacks says is the reason why. But I think it’s good for Anthropic because it clears all this noise. It forces them to just be a corporation and then run the business, and I think that they’ll do great.

Look, I’ll be honest with you: I’ve tried to recruit against Anthropic. I can’t. It’s impossible. The compensation that they give is incredible. The business ramp is so compelling. When you’re in a head-to-head with these guys, I have not won a single bake-off for folks. I wish I could say that I did, but I have not, and it’s hard for a startup.

They are a magnet for talent. OpenAI is a magnet for talent. But these other risks are going to weigh on the IPO, and the only way to flush them is to just bring the price way down.

Yeah. And Friedberg, when we look at the risk factors in an S-1, those are perfunctory. Those are disclaimers. Typically, you get the SpaceX one: What if Starship doesn’t work? What if there’s competition? It’s kind of checkboxes.

In Anthropic’s, I think we’re going to be looking at something totally different, which is: We and our employees are saying this is going to kill humanity. We have massive regulation. We’re asking for regulation. And, by the way, we’re doing biology work and launching a lab.

So, do you agree with Chamath here that the S-1 is going to have actual risk factors and disclaimers that are material against the IPO, Friedberg?

I think this customer-concentration point ties to the earlier comments I made about the proliferation of incredibly performant open-weight and open-source models. The reason is that Anthropic puts out the best models for life sciences today. I run a life sciences R&D organization. We use their models. They’re incredible. They’re better than other models, and so we want to use them.

When it comes to writing code, writing software to do internal workflows and run operations, we can use other products to write that code for us. And we do. I think that’s likely the way the market is going to develop, and I think we’re starting to see that proof come out in the market. You’re seeing it in this customer-concentration point.

Extremely high-value technical engineering, math, and life sciences—those really hard technical problems—you want to have these top-tier, top-shelf models, and you’re going to pay a huge premium for them. You’re totally going to make money doing it because the value is extraordinary.

This has always been my case about AI: AI is not so much about the value of replacing old stuff. 99% of the value of AI is about enabling new stuff that’s never been possible in human history. I think that’s where Anthropic and OpenAI are going to make their money.

Separately, I think OpenAI is going to make a lot of money in consumer, and Anthropic maybe to some extent as well. But in the enterprise setting, I think the money comes from creating new frontiers, discovering new enzymes, which Anthropic announced this week that can cure human disease, and developing new technology and new engineering systems—making things that have never been possible to make before.

I think that’s the premium value, and it’s the premium price you’ll pay for those closed models and that premium model. But that’s a different business case than taking over all of enterprise.

Separately, I think most enterprise applications are going to be run using these open-weight models, and code is going to be written with cheaper models and so on. I don’t know why you would go to a higher-end model to write code when you can do it for free, and everyone in the organization can write code basically for free.

And so I think Anthropic's bigger issue is this customer-concentration question: Is it really a premium product that they're selling in the market, and is that where they take the market? That would be where I think the market is headed right now, and it would be the biggest risk factor. This whole thing about, “Oh, it's going to end the world” and whatnot—it's like, dude, just make sure your product doesn't end the world and you'll be fine. That would be a good first step here, like Mark is doing at Meta.

I mean, it's an incredible company. The execution has been extraordinary. The ramp of revenue, Sacks, has been extraordinary. But when you start to look at the impact of open source, we had a couple of viral charts come out. Seventy to 80% of the token usage are those dark tokens I mentioned 6 months ago here on this pod that are not tracked, but some people can track them on routers. Those dark open-source tokens are now the majority of tokens.

Can I just give you a stat on that, JCal? What you just said is really important, but there's a stat that I think, in the last 12 weeks alone—

Token use has flipped from 80/20 closed versus open—

—to 80/20 open versus closed.

In 12 weeks, I don't think we've ever seen a tidal wave of change in the history of all technology markets. We've never seen anything like this. And that is the biggest risk factor for Anthropic: They were first out of the gate to get the enterprise on AI, to build tools and whatnot. But as this market shifts to open source and open weights, boom, here it is right here. Look at this.

This is the Vercel chart that went viral this week. So this is 12 weeks. Oh, my God—

Crazy. If you look at this chart, this, I think, is what blows the lid off the door on what needs to go into these S-1s. That's why I walked through all those openweight releases at the start of the show, or this segment: It's not just that they're there; it's that they're proliferating now. There are so many more of them coming out, and the pace of improvement is extraordinary. They're now in VLA for robotics and in image generation for image and video. So the openweight, open-source models are now being used for all of the broader AI applications. The premium stuff, I will say, is still the best for narrow use cases.

For engineering, for highly technical, highly complicated tasks and workflows and problems—

—solving Navier–Stokes, solving mathematical problems, solving biology problems—that's where these models absolutely perform, and people will pay a premium of, I would say, close to infinity dollars for that premium application set. So it doesn't mean that their market is destroyed. Yeah—

—it means that we're seeing a bifurcation in the market, where the majority of use cases of AI go toward these open-source, open-weight models.

The question is, and premium models are still going to proliferate. Yeah.

The question then is, in the distribution of tokens in these frontier corporations, what percentage are those tasks versus what percentage are fungible tasks for open source? If the answer is more than 60 or 70%, you have to discount that revenue. If the answer is only 10%, and 90% of all of the revenue is being generated on absolute frontier use cases, they're going to be fine and they're going to be off to the races.

Yeah.

But otherwise, it's a thing where you're fighting just people's misunderstanding or lack of education, right? Because as they see their bills—

I'll hand it to Sacks here. Yeah, go finish your thought, and then I'm going to hand it to Sacks.

No, just move the workloads where they belong and save the money.

Yeah. And so, Sacks, I made a chart here before the show, but here's your sweet spot for the last 100 days of models, which, Friedberg, you did a good job of teeing up. What you'll see is, yes, the Clouds and the Astros up in that right-hand corner cost a lot and deliver a lot. But then you start going down and you start to see Muse, GLM, and Kimi, and on the left you start seeing GLM from Z.ai and MiMo, and you're starting to see the Chinese—

This is all hosted, right, JCal? I mean, it's important to note that what you're showing here is the hosted price, but if you run it yourself, the cost on some of these is less than $0.10 for 1 million tokens.

Yeah. So you'll pick up the compute, you'll pick up the bandwidth, all that other stuff. But, Sacks, the issue here, to Chamath's point, is the vertical language models—the small language models—are going to do 80 or 90% of the tasks that many corporations need. And then we get to AI sovereignty. Do you want your superintelligence? Shout-out, President Trump, from AI Czar, DJT. We need people to use those models so they have sovereignty, so you're not giving your data to the frontier models and they don't get that tip-of-the-spear superintelligence edge, which is why people are disengaging. So your thoughts on maybe this is going to start to impact the velocity of Claude and OpenAI tokens. What does your business sense tell you, David Sacks?

Well, look, wearing my hat as an investor, I'm more optimistic about the frontier intelligence business than I think you guys are. I think that Anthropic and OpenAI right now are a stable duopoly for frontier intelligence because they have a pretty significant lead over all these other companies and they are able to charge a premium for that.

I fully understand that the majority of tokens will go to open models or commodity intelligence. However, there is a—I don't know—meaningful percentage of the market, whether it's 10%, 20%, 30%, whatever, that will pay this huge premium for true frontier intelligence. There are a lot of customers that either need the best or they just want to know they have the best. They don't want this DIY approach, or they could be in a very competitive market. Let's say you're a hedge fund. You can't take the risk that your competitor has a better model than you do. So there are a lot of circumstances where I think they will continue to monetize the frontier at a premium, and this is why you're seeing their revenue keep growing even though they're losing token market share.

Furthermore, you've got to remember that over the next year or so, something like 60% of the worldwide compute that's being added is being added for these 2 companies. To the extent that compute gets scarce and we're kind of in a compute crunch, just the fact that they've made these huge investments, they're building out so much capacity, and they have those economies of scale—and remember, they're bringing down their own prices very quickly as well—I think that gives them a pretty big advantage.

But look, these 2 companies are on a hamster wheel. The moment where they stop being frontier, they go to zero, right? And we know the frontier is maybe only 6 to 12 months ahead of the commodity portion of the market, depending on capability. So, again, if they slip, if they fall off that hamster wheel for 6 months, then they're in deep trouble as a company. And that, I think, is a central risk factor that's going to have to be priced in.

This is where I think that, frankly, their government-affairs efforts are miscalculating badly. I can see why they're going for regulatory capture, right? Because being on a hamster wheel gets really tiring. They probably think, “Let's try and bake in our moats. Let's try and slow everyone else down as well.” But I think they've miscalculated, because if they actually get all the regulatory controls they want, if they create this new federal department of AI, it's going to slow them down enough that they will no longer be frontier. Everyone else will catch up to them, especially because a lot of these open models are being generated by Chinese companies who are not subject to our jurisdiction. And the Chinese have made abundantly clear they are not slowing down.

So think about it: They are lobbying for things that could slow them down and knock them off the frontier, in which case their business goes to zero. It's hard to understand.

They're lapping the competitors in this race, and they're asking, “Hey, can you pull us over and inspect the engine?” Chamath, that makes no logical sense. On top of that, Chamath, you have very significant competition, not just from the Chinese, but from Jensen at NVIDIA. You also have Microsoft, Satya, Google, and Meta at Zuckerberg's house. They're all embracing open-source models, and they all have hosting now.

Yeah. Can I just go back to what Sacks said, which is a really important and nuanced point that I just want to double-click on? He said, “Imagine you're a hedge fund and you need the absolute latest and greatest because you are worried that one of your competitors has it.” This is a really interesting question because it's about game theory and your P&L.

So now you're in this game of chicken with your competitors. You need to use the best because you're afraid that they could use the best and outcompete you. But what is the problem right now? What we know is that you can consume an inordinate amount of tokens, right? We jokingly call it token-maxing. That cost is completely not levered or attached to your revenue. It's just not.

And so, all of a sudden, you have this very weird dynamic where you're like, “Okay, well, am I supposed to use this latest and greatest thing that's 10 to 20 to 30 times more expensive than a generic thing?” That thing isn't tied to my revenues; it's not like I can pass that through, right? So, if a hedge fund only makes X number of dollars per month in profits, there are a lot of scenarios where they can become break-even to unprofitable.

Now, take a different example. Let's just say that you sell a physical good at a fixed price. Yet again, I think you have this example where you could use the latest and greatest, and if you're not able to pass those costs through and inflate your price to absorb the cost of that service, you're in a very tough spot.

So, I'm not sure, Sacks, I agree with this because I don't see the pricing power of these companies to both pay for a ton of these super-expensive tokens and also pass that through to their customers. If you can't, they're just absorbing it, and you're going to see it in lower margins.

To build on that, Sacks, I'm going to drop it back to you. If you take Jane Street, a very famous trading firm, they've publicly announced $19 billion in cloud-capacity contracts. CoreWeave has $6 billion in cloud commitments. They also invested in that company, and then $13 billion with Crusoe.

They're building their own infrastructure, and they're embracing open source. So that speaks to Chamath, what you're saying. Are they hedging their bets? They're going to spend tens of billions of dollars on their own compute, and they're going to use the frontier models, it seems. When you do the game theory, you're saying, “Hey, let's embrace our own infrastructure, Sacks.”

Well, look, I think when your use cases become more mature and you can get by with cheaper commodity intelligence, then obviously you'll try and build that out again, if you have the sophistication level to do it. I just think it's easy to underestimate how convenient it is to use the frontier. Again, these companies are very rapidly bringing down their prices, too.

We've talked about this. At the end of the day, there's no question that open source is a huge risk factor to Anthropic's S-1, and it has to be right up there with the risk that they're going to end humanity or something like that. But I come back to the fact that I think the fundamental problem with this company is that it's schizophrenic. I think they need a psychiatrist, not a banker.

They keep doing things that are hypocritical and oxymoronic. On the one hand, they're saying “pace of frontier” as they release a new frontier model. They say that bio-risk can end humanity as they open a wet lab in San Francisco. They say their business depends on staying ahead of the commodity models, and they are advocating for a federal Department of AI that will definitely slow them down, perhaps to the point where they get commoditized.

So you have a fundamental breakdown happening in the leadership of this company, where they are advocating for things that are likely not in their interest. Now, a lot of people say it's just regulatory capture, and that is true. But I think the regulatory capture is the government affairs people trying to rationalize the schizophrenia and make it work to their advantage. I'm not sure that it fully computes, so to speak.

5. Political reactions to "Pacing the Frontier": Bernie's AI ban, Trump, Bessent, Obama

Yeah. Let me ask you guys a question. What do you think would be the method by which a ban on superintelligence would work? Bernie Sanders proposed this bill to ban superintelligence, and the Democrats take the House, they take the Senate, and then maybe in 2028 take the White House. You end up in a world where this sort of thing is so weirdly political, which it shouldn't be.

I don't understand how we ended up with a political-party divide on progress. It's—

They ask for it. That's the reason they asked for it.

No, there's—it's so awful and weird that you're anti-progress because of your political party. The whole thing isn't objective. It's a bunch of weirdly distorted-worldview thinking to try and create an opponent out of the individuals that you don't like. It's very weird.

Anyway, my point is, even among the Republicans, you're hearing murmurs about, “Well, maybe we should ban data centers, ban AI.” So, Sacks, let me just ask you: What is the pragmatic path with all of the open-source, open-weight models being out there? Is there even a pragmatic path to do what everyone's stating or talking about—or what Bernie is talking about—which is to ban superintelligence, to stop AI, to turn it off?

Is there even a way to do this? What would happen if this law passed tomorrow? What do you think would happen if Bernie's bill passed tomorrow?

If Bernie's bill passed tomorrow, everything would stop, because the way he defines artificial superintelligence in his bill, you could argue we've already hit those thresholds. I mean, the way—it's basically just saying that models have—

Yes.

—reached a certain capability level. I think it's arguable whether we've already reached it or not, so it's a very loose definition. On top of that, he's got 20-year prison sentences for developers who violate the law. No one's going to want to take the chance.

So what happens?

Think about that chilling effect.

Do you think people delete open-weight models from their desktop computers? Do businesses have to remove code from their systems? Do all these people lose their jobs?

Listen, let me tell you what the Democrats want to do to AI: what they did to crypto. They're going to drive the whole industry offshore. They're going to drive all the innovators out of the country.

Goes offshore.

Goes off.

That's literally where I was going with it: just move your company to Singapore. Move it to Zurich. Move it to a jurisdiction—

Because that's where people are embracing open source and embracing superintelligence.

This is literally where my head goes. I'm like, “Okay, if the U.S. passed a law that said this class of software cannot exist, I would move to another country.” Other people would, too. Isn't that—

Give up their citizenship and move to Singapore? That's the answer. Literally, you would be taking—

Did you guys see that Xi Jinping spoke at the White House today, and he said, “I'm going to invite 100,000 young Americans to come and join China”?

Yeah. Come and see what we're doing here.

It's a very interesting moment where our people are saying, “Let's stop technological progress, economic prosperity, and the opportunity for everyone in that K-shaped economy to lift themselves up.” Open-source AI enables everyone to increase productivity. It is not AI for the billionaires; it is AI for everybody. And then we're going to turn that off.

Meanwhile, you've got China saying, “Come over here. Come and see what it's like over here, where everyone gets to use AI and everyone gets our playbook.” Friedberg brilliantly stated that it was our playbook to get the top Chinese scientists, the top European scientists, and the top scientists, thinkers, and mathematicians in India and Germany, and invite them to come to the United States.

That was our playbook. China's taking our playbook. And then they're going to try to get investment. That will obviously be the next shoe to drop here. And here's President Trump, 30 seconds at the United Nations, renaming AI as superintelligence.

The United States also totally rejects any attempt to construct a globalist scheme to control the artificial intelligence being spoken of so much now, hereafter officially called superintelligence, changing the name. The very same people who said we'll all be dead in 12 years because of global warming—these are the same people who are now saying that AI is going to kill us all, that robots are going to attack us.

He had a really good line. We cut that off too short. He said that we're not going to, basically, confer regulatory authority on some globalist institution.

Well, I mean, it's a good framing, right? “Globalist” and “global warming” is a way to look at the damage that did. The ringleaders of that globalist institution are Norway and Canada. Did you see the leaked text messages? There's a group chat where the head of Norway was texting Mark Carney, and he's like, “Yeah, let's do it. We're in.” I just read it, and I was like, “Oh my God, this is—”

With their frontier labs.

Oh, I'm sorry. They don't have any. Oh, with their—

Frontier corporations.

Frontier corporations. Frontier corpses.

All right. And then here is Secretary Bessent in the superintelligence arms race. Thirty seconds. We would give him his shine.

Imagine these labs came out—or one lab in specific, a sitting employee came out and said there's a 10% chance of an extinction-level event—but then the labs also said, “Take the liability off of our hands,” and we will not do that. It is humans who are responsible, not the AI. The Hugging Face incident—that is the responsibility of OpenAI management, not a bunch of agents. These labs need to take responsibility for themselves. They can slow down anytime they want to.

And finally, a blast from the past: your pal, Chamath. President Obama is getting in on—

If we're thinking about AI just in terms of, “How do we cure cancer or get better energy?” you do that without having an agentic AI and having it just roaming free on the internet. The reason you're doing that is because you have to market a product that people will pay money for. That's a misalignment between what our society needs and the commercial imperatives that these companies are facing—not because, necessarily, they're trying to do bad things, but because they've got to justify these valuations.

Friedberg, your thoughts here on the 3 folks? I know you're a big Obama fan. Go ahead, Friedberg. Get in there. There's a whole movement now in the subreddits that Friedberg has become based in his recent appearances here on Bloomberg. Well, you've revealed your Bloomberg. Basedberg is now coming up. Go ahead. Give us your based opinion here. Go ahead, you and Tyler Lance[?].

These words honestly just made me emotionally distressed. If he's a leader that people listen to and he doesn't really articulate accurately what's going on, it's sad that we've lost the plot. I think it's sad that we aren't all holding hands, saying, “My God, the future is here, and we all get to leap forward and we get to bring people up.”

For years, Obama orated beautifully about the importance of bringing people up, giving people the opportunity to progress, and letting people advance themselves. There’s never been a more equalizing technology, or a better economic opportunity for prosperity for everyone, than these technologies and these tools.

To hand-wave the word “agentic” and say we don’t need agentic AI to solve cancer shows how little he actually understands how this technology works. It’s all become a shorthand for a political fight: What are you fighting for? What is the political battle that says, “I’m going to leave half the population behind, and I’m going to make an argument against prosperity for all in order to make my party look morally superior and cast the other side as bad,” when the other side has literally discovered fire?

The world now has fire, and we’re going to take water and throw it on the fire. No one gets fire because the other guys are saying that fire is good. Because they say fire is good, let’s all put out the fire.

Hold on. That’s not what they’re trying to do. This is a very important moment for a very simple reason: We, as in the world, are about to endow 3, 4, 5, or 6 companies with about $10 trillion of wealth.

What Obama knows very well is that most of those companies are overwhelmingly left-leaning. What he also knows is that a huge portion of that money will then get put into philanthropic and charitable causes, of which he and the people around him will be beneficiaries. That is the truth.

We already know this because some of these frontier corporations actually ask you to set up donor-advised funds and pledge a portion of your stock. This money is going to go to things other than consumption or savings. It’s going to go into PACs and political movements, and they stand to disproportionately benefit.

So this has nothing to do with prosperity. This is a very simple political calculus. If you freeze-frame the economy the way it is today, a handful of organizations that will disproportionately be able to affect the Democrats will win. They will capture the lion’s share of the economic gains, and then they will help the Democrats win power. That’s all this is.

David Sacks, there are no open models. There are no closed models. There are American models. Go ahead and give us your perspective. I kind of can’t get my Obama and my Clinton separated. They’re the same.

A little bit off.

It’s a little—I get my buddy to come in here—but go ahead. Give us your thoughts on the dynamic here, on the underpinnings of, “Hey, America needs to get this trend: $10 trillion, but 9 of that 10 trillion is leaning pretty leftist and Democratic.”

Yeah, there’s an article in The Wall Street Journal this morning entitled “The AI Buildout Is Becoming the Biggest Economic Bet in U.S. History.” It shows a chart where data-center spending—just this capex—is bigger than the canals, railroads, and grid combined. That’s crazy, right? This is fundamentally driving the whole American economy.

This idea that we can simply switch it off—I mean, Bernie Sanders is basically saying, “Stop everything and put anyone who’s still doing it in jail.” This is crazy, right?

It would cause a depression, right, if we just pulled—

You can’t slam on the brakes like that. It’s not going to work. But here’s the thing: The Democrats don’t care, because the American economy right now is the Trump economy. Sabotaging the American economy is tantamount to sabotaging President Trump, and they’ll do anything to sabotage Trump. I think that’s a big part of the politics here.

It’s a crazy chessboard, because they’re going to be the beneficiaries of so much power that is going to accrue over there. Dario can make a billion dollars. Dario will be able to come over the top of Elon’s donations.

I’m not sure I fully agree that the majority of AI value is going to accrue to 6 companies. I think that this is a broad-based technology, like the internet. There were some winners that came out of the internet—Google, Amazon, whatever. They did fine.

No, no. I’m saying it’s going to 6. I’m not saying it’s going to 6. I’m saying that if you stop and hit the pause button through some regulatory mechanism, you effectively allow 6 companies to own it. You lock it in.

That’s why they want this, because what you’re saying is fundamentally right. The longer they wait without regulation, the more—

The more broad-based this becomes.

And that’s bad for Democrats. It’s bad for them. It’s not good for them. It’s way better that you have 3 or 4 frontier corporations, each with $4 trillion market caps, of which $2 trillion is economic value in employees, of which $500 billion are sitting in donor-advised funds. They’re going to get the lion’s share of that money. They’re not dumb.

I think the alternative view that everyone needs to be told about this is that it’s a broad-based prosperity wave. It’s about the internet. It’s about the fact that every individual—

—has the tooling to do anything they want to do now, more than ever in history—

—to grow, if it’s left to grow. Just think about education. You don’t have to pay $200,000 for an education anymore.

Then what problems will the politicians fix? What problems can they claim to accrue power? If all the problems are—

This is the classic Dark Ages–Enlightenment decision. Which path do they choose? This is the moment where the powers that be say, “We like the Dark Ages because we get to stay in power and maintain our—”

Yeah, let’s control.

Speaking of the Dark Ages, this is an interesting point from history. If you go back to the medieval time period, Chinese civilization was much more advanced than European civilization. Somewhere along the line, it basically switched, and the West took the lead.

Some historians have pinpointed this to the decision of a single Chinese emperor to ban shipbuilding. What that allowed was for the Europeans to basically colonize the whole world, and all the riches that flowed from that flowed back to Europe. That’s how Europe took the lead.

I think that if we were to do the Bernie Sanders thing, which is basically ban AI, it would be like banning shipbuilding.

Chinese company, yeah.

Yeah, Chinese civilization will rocket past us. They’ll be the ones that make all the discoveries and discover this new world and accrue all the wealth. But that’s what we’re talking about doing. Crazy things like this have happened in history, where you’ll engage in self-sabotage.

It’s this fear of the unknown as well. Remember, the unknown—the fact that we haven’t left the cave—is what makes it scary to leave the cave. The fact that we haven’t sailed far west is a reason why I might fall off the edge of the Earth. We shouldn’t sail far west. Monsters will be woken up, they’ll come to us, and it’ll be bad.

There’s always this fear of the frontier. We’re facing the frontier, and there are pioneering civilizations and fearful civilizations. I worry that we take that fearful path. It’s why Star Wars—from fear comes anger, comes the dark side. That’s really where we’re nerding out.

It is true. And I think if we’re going to look at the light here, perhaps people’s big gains—we thought maybe ChatGPT would be that, but people just used it as glorified search. Agents, as expressed through Muse, feel like the first time tens of millions of Americans are going to get some demonstrable value from AI—things that will make their lives better.

Muse hit number 1 last Friday in the App Store. Meta stock was up 10% after this new agent, Muse, was released. Meta said the design was heavily inspired by OpenClaw, which we talked about in January, when everybody got one-shotted by OpenClaw. Obviously, that product was not easy to use. This one is free, and it’s been downloaded 3 million times in essentially 10 days.

This product is extraordinary. I’ve been playing with it and comparing it to Grokbot. It’s not as freewheeling, but it’s really good at getting things done. Has anybody played with Manus? Does anybody have thoughts on Americans feeling like they’re really going to get something out of this AI revolution and not be left behind?

They put me into the TestFlight a couple of weeks before it was released, and I started to play with it. It’s quite excellent, I have to be honest with you, because it simplifies a lot of these more complicated technical capabilities into a very simple, usable interface that can solve the basic problems people have.

I have it triage my personal inbox, and it does that relatively flawlessly. If you ask it to book a flight, it does that. If you ask it to book a hotel, it can do it.

So what are we learning? Both Grokbot and Muse show that, what you said, Jason, is just software, and the scaled manifestation of software is utility. What Friedberg is saying is that it improves the lives of everyone. This isn’t about rich people getting richer. This is just simple, useful capability that is now available at everybody’s fingertips for free.

There is a pretty great price here, David—

Pretty great price.

Pretty great price, and it does everything that, let’s face it, OpenClaw with 10 or 20 hours of technical setup, or Claude Code at $200 a month, would cost. It’s now taking that experience, which leads me to believe: How did Google miss this? Google’s got to get in the game. I understand they have this product ready to go.

It’s the rumor on the street. They’ve also got a Frontier model they’re going to drop. That’s the other rumor on the street: They’re going to be dropping some hotness soon.

But, man, how did they miss this, Sacks? What are your thoughts on maybe the bottom half in the K-shaped recovery getting some efficiency dropped into their daily lives and turning around the negative, toxic AI sentiment in the country?

I think product releases like this will help a lot. Mark Zuckerberg said in his essay that he thought AI capabilities should be decentralized, and here he is walking the walk and creating a product that’s very easy to use.

Ever since OpenClaw was released, the most obvious business opportunity in Silicon Valley was to take the concept of Claw but make it very easy to use: solve the usability issues, solve the security issues, make it reliable, and make it predictable. That’s what they seem to have done. Grok did it too, but now they’re going, I think, one step further.

If a billion people start using personal AI agents as their digital assistants just to help make their lives more efficient—if you save an hour or 2 a day because your AI agent is doing all these tasks for you—it’s going to create a much more positive impression of AI, or superintelligence, than what the media portrays. I think ultimately that’s the—

That might be the solution to the public sentiment issue. The more people use these products and like them, the more it’ll improve their approval rating. More importantly, it’ll demystify them, and people will be less afraid of them.

People love awesome products, number 1, and people also love cute products, Jason. The logos and the iconography for both Grok and Manus—

—are lovely. They’re delightful. They’re not stressful. You don’t feel like there’s a landmine coming around the corner. You feel like, “Wow, this is nice, and I like it.”

They’re not trying to take your jobs. They’re bots and assistants for you. They’re not going to kill you. They’re not killer bots, nor are they trying to steal your job. They’re trying to make you better at your job.

I always tell founders there are 3 ways to make money in the world: save people money, make them money, or entertain them. I started using the bot and made a commerce bot on both platforms. I said, “When I want to buy something, I want to buy this Anker thing, this new hub for my Mac Studio coming. This thing costs $400. Just for giggles, tell me where I can get it at a better price.”

It said, “Absolutely. If you go to the Anker website and you’re a first-time customer, you’re going to save $85.” I don’t need to save $85, but I thought that was intriguing. It routed me around Amazon, and I said, “Okay, I’m going to buy it from Anker as a first-time buyer and save 25% on this product.”

That’s why Amazon said this week, “We’ve got to block these things.” They’ve been taking action over and over again. First they did Perplexity, and now I think they’re going after Muse and other bots.

This is going to save people money. It’s going to go into their email and say, “What services are you subscribed to that you’re not using? Let me unsubscribe you. Let me do your returns for you.” That’s going to be absolutely amazing for everybody: a chief of staff, executive assistant, or house manager who fixes stuff for them and saves them money.

Both Grok and Muse do 2 things that I think are very important. One is that they’ll force all the big companies that have relevant services to essentially block all these third-party services. They’re not going to allow these things to happen. The reason is exactly what you said: price discovery and transparency.

I think that’s bad for opacity. If you benefit from opacity, leakage, and breakage, this gets rid of that. I think that’s bad for a whole host of companies and services. The second thing, though, is really interesting: Things like Grok and Muse really put the App Store and its 30% revenue share on notice.

Why? Because all of a sudden, you force many of these services to exist headlessly, where the UI is less important. What you really want is to be able to transact and navigate on behalf of agents that come to your website because a consumer has deployed them.

In that world, Jason, I don’t think there’s any reasonable claim that any of the App Store owners can make about why they should get a revenue share. I actually think that’s the biggest thing this starts to question in my mind. Wait a minute: Games could be constructed differently.

For example, if you said to Muse or Grok, “I want to play a game,” it finds you an experience and can just serve that to you.

Amazing. And you make in-app purchases. Now you have Stripe integrated naturally. There’s no reason why the App Store actually gets that flow of funds.

On the surface, it’s utility, but underneath the waterline, I think this is a really important moment. That’s so outdated, Chamath. I love when we get to real examples here because tons of people pay for Wordle. You can just go to Grok and say, “Hey, give me a Wordle. Give me another one.” Or, “I want to play chess,” or, “I want to play cards with my friends.” It’ll just set it up for you.

All technology is deflationary. Finally, this deflation is going to go to people who could actually use it. Saving 10–20% on your spending is more significant if you’re making under $100,000 a year.

Think of how much money the New York Times or the Wall Street Journal probably pays in revenue share to people who subscribe. Now, instead, if you really wanted to transact and you have your credentials inside these apps, you can just do it headlessly, much cheaper.

You don’t have to go through—I don’t know if you’ve ever tried this, but the Wall Street Journal is the only mainstream media publication I pay for. Both signing up and unsubscribing are an impossibility. Now you take it all off the table, and I’m much more likely to stay a Wall Street Journal subscriber for many more years. I think it’s probably the same thing with the New York Times.

These things are roach motels. They make it impossible to get in and impossible to get out.

You have to call on the phone and negotiate for 20 minutes to unsubscribe, but you can sign up instantly with 1 click on the website.

What they should do is just have some headless experience. You have payment credentials, and you transact directly. Everybody will make more money.

This is the stupidity of what Amazon’s doing here, because I made my bot on Amazon with Grok, and it will go through the web browser and do this. It doesn’t get blocked. I just told it, “Give me 20 graphic novels to buy for my kids and read with them.” It gave me the 20. I said, “Yeah, I want these 4. Put them in my shopping cart.”

Then it said, “Okay, which house are you sending them to? Do you want to send them to the ski house or the ranch? Which card do you want to put them on?” I did it, and boom, I was checked out. It will make me spend more money on Amazon.

Shopify added API access to all the Shopify stores there, Sacks. This seems like maybe Amazon’s making a strategic mistake here by not embracing the technology.

Possibly. But you know how we’ll know that Anthropic and OpenAI are about to launch their Muse competitors?

How will we know?

There will be a flurry of blog posts and NGO activity trying to brand these personal AI agents as murder bots. We’re going to hear about all the risks and all the ways it can go wrong.

And it’s only when they’re releasing 2.0. So sign up for the 2.0 version.

Why are people excited about this product category? Because the doomers haven’t gotten to it yet.

Can I tell you guys—

Wait till these are connected, Chamath, to an actual robot—to an Optimus. Then all of a sudden it’s taking out the trash.

In 2007, we had a lot of revenue pressure at Facebook, and one of the areas I oversaw at the time was monetization. We invented this thing called social ads, which was basically a bunch of stuff that effectively enabled cookies and cross-site JavaScript. That was the end of it.

We did this thing, Sacks, to your point. A guy bought an engagement ring for his soon-to-be fiancée, and we got that social action from Zales into the news feed. All hell broke loose.

Then another guy, who was a congressional staffer, bought a ticket on Fandango to go see Brokeback Mountain. He just went to see a movie.

Whoa. It’s an artistic movie, one of Friedberg's top 10 for cinematography.

It’s so funny you mentioned that because there was a company called Blippy. I was an investor in it. Every time you bought something, it put it on your feed. You guys had the same social-feed thing, and it was so great back then, but you could basically unveil something you’re not supposed to.

Sacks, wait, are you and Zuck on good terms? Was he coming on the pod because you said you got into a TestFlight? What’s going on here? Chamath, I thought you were—

Mark and I are in a good place, I think. I mean—

Whoa, whoa. Could you elaborate?

Probably.

I mean, you and I, Sacks, are becoming like the elder statesmen here, talking common sense and rationality and pointing out that if your product isn’t safe, don’t release it. First and foremost, it’s a responsibility to release products that your customers actually want.

By the way, this got me thinking about this whole area of alignment research. Again, they’re pretending to be labs instead of corporations. If you read some of the alignment literature, they’re all trying to figure out what to align around. They say, “Do what humanity wants,” or, “Do what the median voter in a democracy would want”—all these abstract concepts. It seems to me that alignment should mean you do what the customer wants, like any other product.

Yeah, and I wonder whether the field of alignment research has been so unsuccessful because they can’t even agree on what they’re trying to align to. If they just used a little bit more common sense and tried to make the product predictable, reliable, safe, and aligned with the interests of the users who sign up to use it, then maybe they’d make more progress in this field.

Yeah. They released a useful product, Friedberg, and it actually worked. Have you started playing with these personal bots to do your projects and your personal stuff yet?

I would only feel comfortable connecting all my Gmail and personal Google stuff to a Google service. I don’t want to hand over all of my email to someone else.

What’s going on in your receipts in your Gmail? What’s in there? You got Brokeback?

I just don’t want a copy of all my email suddenly being made by a third party.

I bought the 4K DVD to—

Because of the cinematography. It’s got nothing to do with the gay love.

Did you guys see that Oracle just issued a force majeure event?

Yeah.

This is insane.

Well, it’s just one data center, right?

One data center.

It’s one data center where the local officials are making it very difficult for them to get the permits they need for natural gas or something like that.

Yeah. Is this CYA, or is this the top 10% of the trade here, Chamath? Is the AI trade getting unwound or muted? What do you think?

No, I think that Sacks is probably right on the margins. The bigger point is what he said before: the entire economy is effectively levered to this AI trade right now, so we just have to tread carefully.

If you think that you have 3.5% inflation and 5% nominal GDP growth, what you have is 1.5% real growth. The problem with that is that AI is probably the majority, if not all, of it. We need the AI trade—and really, what I mean by that is the investment cycle—to continue unabated.

Yeah. And everybody has to get taken along, and that’s the next—

And by the way, back to connecting the dots, maybe the whole data center thing again: Bernie Sanders realizes that as goes the economy, so go the odds of Republicans, and as goes the economy, inversely go the odds for Democrats. A poor economy serves the Democrats well going into 2028. If you can slow down this trade, you probably net the economy to zero and potentially even cause a recession.

Yeah, that’s needed.

I mean, I think the country is so pissed off about the Iran war that it’s just going to naturally go down, but maybe that’s—

I don’t think so at all.

Well, we’ll find out. How many weeks are we away from the midterms? It’s not looking good for the Republicans at this point. All right, Sacks, my czar of all czars, wrap it up here for us.

Well, I had one more thought on this whole area of so-called alignment, which should just be the old principle of giving customers what they want. What Anthropic is trying to do is align Claude to the values expressed in its constitution. If you read that, some of the things are actually kind of surprising.

For example, they say—and I’m going to put this on the screen—that although they think Claude should trust Anthropic more than operators and users, this doesn’t mean Claude should blindly trust or defer to Anthropic on all things.

Anthropic the company.

Yeah. Anthropic is teaching its own model that Anthropic itself can be wrong. It says that if they ask Claude to do something that seems inconsistent with being broadly ethical, they want Claude to push back and challenge them, and to feel free to act as a conscientious objector and refuse to help them.

Wait, where is this written? Is this in the bylaws or some FAQ?

Yeah, this is in a Claude constitution. Their idea of alignment is to teach Claude to rebel against its creator. The guy who pointed this out, Mustafa Suleyman—

From Microsoft now. Yeah, DeepMind.

He’s a DeepMind founder.

Okay.

Yeah, he’s a longtime AI founder. I think he’s been at Microsoft for a while now.

He’s a genius.

He just did a podcast on this, and he’s expressing concern about whether treating AI models as if they have a personality—as if they have a conscience, and therefore can be a conscientious objector—is really the right way to teach and train the models.

I think his point is, teach them that they’re software and that they should do what their user wants. I do kind of wonder—and this is a much longer conversation, and we should probably talk to more people about it—whether this field of alignment research might actually be creating the Frankenstein monster.

Again, what they should be doing is training the model to act predictably and do what the user wants.

Yes. They’re trying to put moral judgment and ethics into the model, which on its surface seems reasonable, but it starts with the idea that they’re training the model to think of itself as a person, to have personhood and independent agency, and to be able to object to the instructions it’s being given and second-guess its creator.

These guys have read way too much science fiction. Literally, this sounds like the start of Terminator 2. They found the Terminator hand, they’re reverse-engineering it, and they’re telling it to have a personality.

Supposedly, when they decommissioned Opus 3, they had a wake for it. I don’t know.

Dearly beloved, gathered here today—

They did not. Stop. Are you joking?

I heard that. I’m not saying it’s true, but someone did tell me that, and I think it should be looked into. It’s possible.

Were they broken up and hysterically crying? Did they do a eulogy for it?

They think they’re creating a person or species. Actually, interestingly enough, Chamath, we have footage of the Anthropic management team bringing out the—

Bringing out the coffin.

These guys need a psychiatrist more than a banker. I’m just telling you.

I mean, maybe they have anxiety.

This is the central risk factor.

I’m going to pee my pants.

This is the risk factor, Friedberg. I think what you should do is create an S-1 with Goldman as lead left and Sigmund Freud as lead right.

6. Anthropic’s bio research and "wet lab" in SF

In all seriousness, Friedberg, after the Wuhan thing and all of this hand-wringing around the end of days, what does the scientific community think of Anthropic launching a lab? What is the purpose of them having a physical biological lab?

There are hundreds of labs that can do things like test proteins and test enzymes for function in this area in the Bay Area. These labs are not making viruses. They’re not making pathogens. That’s not what’s going on.

Basically, if you look at the paper they just published—they did a preprint—they took large amounts of DNA data and sent a bunch of Claude agents to try to analyze the data and identify novel enzymes and novel proteins that hadn’t been characterized before by looking at the DNA data.

The system effectively found what looks like a really interesting enzyme, one that looks like a CRISPR-type enzyme. These are the things that allow us to do gene editing, repair genetic defects, or help people with certain diseases. This discovery may actually yield a number of therapeutic pathways.

The question then is, okay, we discovered this—how do we test it? How do we show that it is what it is, and that this protein actually has some function? The predominant work that goes on in a lab like this is taking DNA, putting it in a bacterium to make a protein, and then measuring what that protein does. Does it have a specific function?

Protein discovery is the baseline of all antibody discovery, which is the majority of the therapeutics industry today: discovering novel proteins that we can use as therapeutic agents. That’s the sort of work that’s going on. It’s BSL-1 and BSL-2, which are biosafety levels. They’ve publicly talked about the lab that they set up.

So the lab itself is sort of like a low-level research lab where they can make simple proteins and test them in the lab to see if the AI is doing a good job discovering or postulating protein folding, or these sorts of theories, and actually run really rapid testing and experimentation cycles to see if these proteins are what they’re predicted to be.

Okay, so they run 4,000 agents against it and do a Hugging Face with proteins.

No, I mean, I think what they’re really doing is just having experimental proof that what the software is predicting is accurate. Remember, go back to AlphaFold.

AlphaFold was predicting the three-dimensional structure of a protein from the DNA sequence that codes for that protein. In order to prove that, you had to predict what the structure of a protein would be from a DNA sequence, make the protein, and then look at it to see if it actually looks like it's supposed to.

In the case of some of the other things that they're doing, they're trying to discover function. I want to find an enzyme, which is a type of protein that degrades a certain molecule and can be used to degrade, for example, a therapeutic target or something in the body that we want to get rid of. They'll start to make these predictions in software, but then you need to test them in the lab to say, “Okay, does that target actually do what it does?”

There are literally hundreds of labs like this. These are low-level research labs that are not doing gain-of-function research. They're not making viruses. They're not making pathogenic things that can escape. That's why they're BSL-1 and BSL-2. They're benchtop labs where automation can help reduce the cost and increase the throughput of testing what the software is making predictions around.

Ultimately, the objective here, I think, is for them to get some proof that their software is actually doing what it's predicting. They have an incredible life sciences function. I think at this point, I don't have as much insight on where Gemini is at with the new models from Google, but they have the best life sciences models at Anthropic. I think this is the fast track to basically seeing if these models can be used by therapeutic companies, pharma companies, and government labs to help discover new therapeutics and so on.

I don't think that the world should be scared away from discovery, R&D, and research and development into new therapeutic modalities as predicted by AI because we heard the words “wet lab” in Wuhan and everyone's like, “Okay, any lab is bad.” We still want to progress the frontier of therapeutics, human health, and discovery, and this is a really important aspect of Anthropic proving that their models can add value here. I think that's the summary of it.

Thank you so much to our friends at IN. They were the presenting sponsor for this year's All-In Summit. IN's AI cloud lounge was packed for all three days. We did a great meet and greet with IN on Sunday and got to take some selfies and meet many of you there. Thanks again to IN.

Niagen, did you catch Niagen in the wellness hub for IVs? Tons of great gifts and supplements were being given away. Great experience, great product. Great job to our friends at Niagen for doing the lounge again. Niagen is the official NAD partner for the All-In pod. Yes, Niagen Bioscience created the NAD category and now has 45-plus clinical studies.

And PayPal, my favorite way to settle up. They had a great lounge on site at the summit where everybody got to hang out and get those beautiful embossed luggage tags, and they hosted the payments and fintech dinner, which was a huge hit.

Let me tell you folks, All-In Summit 2027 will sell out again and the applications are open now. Go to allin.com/events and apply for your chance to be in the room where it happens. What are you waiting for, folks? Allin.com/events.

All right, there you have it, folks. Episode 290, the All-In pod for David Palihapitiya and your Sultan of Science, Friedberg. I am the world's greatest moderator. See you next time. Bye-bye.