Anthropic提交9650亿美元IPO文件，Trump签署AI行政令，ChatGPT用户突破10亿 | EP #262

Anthropic just confidentially filed IPO paperwork with SEC and could be the first major frontier lab to go public. Polymarket gives it a 60% chance that Anthropic surpasses $1.8 trillion in market cap on its first day. Anthropic is only 5,000 people. You're talking about an insane amount of money divided 5,000 ways. This group of companies can do thousands of billion-dollar acquisitions. You've never seen revenue growth like this. President Trump just signed an executive order that basically says America is not going to slow down on AI. This is the U.S. planting its flag and saying, “We compete. We don't constrain.” I think the U.S. government realized a year or two ago that AI is key to full-spectrum dominance. OpenAI finally passed 1 billion monthly active users. Nothing in history has scaled this fast. It's crazy to think you're just at the start, right? Intelligence is going to go to every single person and will be accessible to them. Who is your Jarvis? That is actually the only game in town. Now that's a moonshot, ladies and gentlemen.

I'm here with my incredible moonshot mates, D.B., the wizard of investment. Hey, buddy, good to see you.

David Blumberg

Good to see you, too. You got a new shirt on today, I see. What's it say?

Every day, a new shirt. Thanks to the team here. I'm heading off to San Francisco tonight or tomorrow morning, too, so I've got to dress the part.

Alex Wissner-Gross

I love it. A.W.G., our in-house polymath, Alex. I see you're in your normal garb.

Thank goodness. Good to be back.

Alex Wissner-Gross

Normal orchid.

Totally original environment. It may be real, it may not be. It's unclear. And a longtime friend and mate, Immad Wak, the founder and CEO of Intelligent Internet. Immad, thank you for joining us. It's late there in the U.K.

Immad Akhund

But hey, you know, sleep is for other people, not for those living through the singularity.

Yes. How many pints into the evening are you, Immad?

Immad Akhund

A couple. A couple. I have to, with everything that's going on.

When you see this news, you'll be glad.

Salem Ismael, of course, is a probability function someplace on the planet. Salim, we miss you wherever you might be traveling. Come back soon.

I'm Peter Diamandis, your host and your abundance amplifier. We've got a loaded show today: 18-plus stories and some of the biggest developments in AI, longevity, and the future of work. This is no doom, no gloom. This is about the science, the tech, and the money accelerating us toward the singularity.

Let me give you a quick preview of what we're going to be discussing. Trump just signed an AI executive order that rejects heavy regulation and asks the labs to voluntarily share their models 30 days before release. ChatGPT finally hit 1 billion monthly active users. Big news: Anthropic filed its S-1, a trillion-dollar-plus IPO. And Bernie Sanders wants a piece of that IPO, suggesting that the public should own 50% of AI companies through a sovereign wealth fund. Finally, we'll close with 3 multibillion-dollar longevity news stories. Gone are the million-dollar stories; we're into billions and trillions now.

All right, let's jump into our first story here from the White House. President Trump just signed an executive order that basically says America is not going to slow down on AI. No heavy-handed regulations, no permission-based frameworks. Instead, the order asks AI labs to voluntarily give the government access to new models 30 days before public release. Sam Altman said the new EO gets the balance right. Anthropic said they're on board, too. Meanwhile, agencies are being directed to deploy AI-powered cyber defense across government systems.

This is the U.S. planting its flag and saying, “We compete. We don't constrain.” Alex, do you want to jump in first?

Alex Wissner-Gross

I think this is a signpost, a canary, if you will, that a lot of previously governmental functions, such as discovering zero-days at scale, which the NSA historically performed, are now de facto privatized, and this is downstream of that.

We have the so-called Mythos moment, which I think was in no small part the inspiration for this executive order. That basically took some of the core R&D activities that would have been in the NSA—discovering breakthrough cyber vulnerabilities—and commodified them to the point where there's a model, and probably GPT-5.5 as well, with Mythos widely reported to be about to be broadly released. Certainly, Mythos has become more and more accessible both across the EU and within the U.S. to the private sector.

You have to ask the question, just as a thought experiment: What's the right way for the executive to respond when there are private-sector capabilities with dramatic national security implications? I'll add parenthetically: Cyber vulnerabilities are just what's possible now. Imagine at some point in the future when there are breakthrough biological, chemical, physical, and other discoveries and inventions that come out of private-sector models and not out of government laboratories. What is the right policy for the executive to have in order to protect national security?

I think this is such a delicate balance. On the one hand, you don't want to throttle American innovation. A 90-day delay could have meant all the difference between U.S. and Chinese models. On the other hand, not being involved enough in the process and not getting pre-release review capabilities, however voluntary or otherwise, could have a profound impact, positive or negative, on national security.

I think the question of whether this strikes the right balance—we'll know probably in the next few months, because that's the timescale that this is operating on. But I do buy the premise that there was a need for some sort of policy, even if implemented via EO at this point.

You remember, it was about 3 weeks ago that Trump was about to sign it, and he had calls from David Saxs, Elon, and many of the heads of the labs saying, “This is overreaching. Don't do it.” Maybe the old order looked more like Europe.

Immad, you're kind of in Europe, in the U.K., which has been much more prescriptive about its policies. What do you make of this?

Immad Akhund

I think the U.S. government realized a year or 2 ago that AI is key to full-spectrum dominance. This is a military concept: You have to have superiority across air, land, and sea. Intelligence is clearly a vector there as well, and that kind of trumps almost anything.

As Alex said, the Mythos moment was a big deal, and that led to saying, “Well, we need 30 days at least,” which is a very short period of time, particularly with how fast the government actually works. In Europe, you don't really have the same impetus, because Europe has never had full-spectrum dominance. On the other side, on security, they feel that compliance is the key.

I think it's a mixture of these 2, but the U.S. has basically said, “We're not going to be left behind on AI.” Even that 30 days is a massive tactical advantage for us. We're seeing, for the first time, an increasing amount of cyberattacks from all sorts of actors.

On the other side, there needs to be that battle-testing as well. The internet is held together by strings and duct tape, effectively, and that's why you're finding vulnerabilities all over the place. They need it for security as well as the dominance side. Again, in Europe, we just can't move that fast.

I mean, Dave, this is voluntary, right? The question is, is this purely a political show move to say, “Hey, yes, we're going to do some regulation, but we're really not”? Labs can continue business as normal and voluntarily show the models. What do you think of this?

David Blumberg

Yeah, exactly right. I was just on a call right before this podcast with the biggest asset manager in the world, and they were asking, “How is our AI going to give financial advice without breaking the law?” And I said, “Have you gone and talked to Donald Trump yet?”

What Sam and Dario did was say, “This regulation is onerous. It's going to slow us down.” They met with the White House, and then you have a complete watering down of the regulation. This is fine: “We're allowing—because I trust you personally—we're going to allow you to self-regulate for the next window of time. Let's meet again in 3 months.”

But that's exactly what Lip-Bu Tan did. You remember, the White House, or Donald, posted on X: “Lip-Bu Tan must go. Intel cannot have him as CEO. He has investments in China. He cannot be the CEO.” He went to the White House. They had a meeting. Now he's made hundreds of millions of dollars.

CEOs, you must go and talk. But I don't think that is a bad thing in this environment. As a way to govern a country in the long run, it's not good. But in the moment we're in right now, the other regulation would have slowed down progress tremendously. They made the right temporary choice, but it doesn't solve anything in the long run.

But that's okay. I mean, it's fine for now.

Meanwhile, OpenAI finally passed 1 billion monthly active users. Let that number land for a moment. ChatGPT launched in November 2022. Roughly 3 years later, it's used by over 1 billion people.

For context, it took YouTube a decade to get to 1 billion, Instagram 8 years, TikTok 5 years, and now we have ChatGPT there in 3. Nothing in history has scaled this fast. Year-over-year growth is at 62% and holding for OpenAI.

Here's the kicker: Claude, Anthropic's model, is now at 56 million monthly active users and is growing at 10 times the rate—640% year over year.

This entire category is going nuts. I mean, you’re playing in this; you’re building models. Do you expect growth like this to continue? It’s crazy to think you’re just at the start, right?

David Blumberg

The Claude thing is what surprises people a lot, because Claude is a bit like Apple and OpenAI is a bit like Microsoft here. Then there are the 7 billion people who don’t use it. Of course, it’s difficult to use in China because you can’t even have access to these models.

Obviously, intelligence is going to go to every single person and will be accessible to them. Sam Altman has said the price of equivalent intelligence will drop by 100 times over the next 18 months. I think when you look at all the chips and everything, that’s correct as well.

So even though you’ve hit 1 billion faster than anything, I think that 62% will actually increase, especially because I think the next wave, as they hit their IPOs and things like that, will actually be customer acquisition. This happened without massive advertising campaigns, without classical cost-of-user-acquisition metrics. Almost all of this was effectively organic.

You see the odd ad, but they’ve just become ubiquitous, right? I think there’s a long way to go from here.

Yeah, Alex.

Alex Wissner-Gross

You remember when Sam said that if he had a choice between 1 billion monthly active users and the strongest model, he’d pick the 1 billion monthly active users? That’s the future.

Did he say that?

He did say that.

Really?

He did say that. What a memory. He thought at the time, or at least what he conveyed was, that in this game between the frontier labs, distribution ultimately was a stronger mode—a stronger advantage and differentiation—than having the strongest model weights.

Ironically, for a brief moment, at least until GPT-5.5, that was basically the world that we found ourselves in, where GPT and OpenAI had the majority, or at least the largest user base, but not the strongest model. Now, fortunately, thanks to that code red, GPT-5.5 is, broadly, as far as I can tell, the strongest model, and it also has the broadest distribution pipe.

Amazing comeback. Amazing comeback. It’s a good day for OpenAI. They have the distribution, the best distribution, and the strongest model.

Yeah.

Well, pretty soon we’ll have public market caps to compare to each other, so we don’t have to debate who’s ahead. We can just look at the stock prices, right? Then we’ll know.

A trillion here, a trillion there. The most amazing thing is that there will be some product somewhere that grows faster than ChatGPT. I don’t know what it is or when it’s going to hit, but we’re going to have that happen. It’ll probably be a product that is distributed through agents around the world.

How soon, Peter, until we’re measuring the time to the first billion agents using you and not the first billion humans?

I agree. I completely agree with that.

I think this is the big strategic thing, because I think what everyone’s going to realize in the next phase is that the most important agent is the agent that’s next to you, coordinating all the other agents. That’s where the cost of user acquisition is going to go and shoot through the roof.

Because everyone will want that agent.

Is it your super app from ChatGPT, or is it your Claude, or is it your—

Jarvis, right?

Who is your Jarvis? That is actually the only game in town.

Agreed.

Sticking with OpenAI, they just launched something called Rosalind BioDefense, named after Rosalind Franklin, a British chemist who helped discover the structure of the double helix. Rosalind BioDefense gives trusted researchers in government and public health agencies access to specialized AI tools for detecting outbreaks, improving disease surveillance, and accelerating vaccine development. Richard Johnson, OpenAI’s national security risk mitigation lead, is running it.

This is AI’s move from chatbots to biosecurity infrastructure. Politically, I think this is the kind of move they have to make—this, and health care and education—to really give themselves a defense against people slowing them down. Alex, what do you make of this one?

Alex Fink

I’ll paint a story about the elephant in this particular room, which is that I think we’re starting to see, for the first time, the generality of AGI start to shrink. I think we’re seeing it for security reasons and security rationales that are probably legitimate. Nonetheless, the G in AGI is starting to shrink.

We’re seeing bio capabilities that would otherwise be built into general models like the GPT series start to get carved out as separate fine-tuned/post-trained models that are only available to exclusive audiences like government agencies and trusted researchers. We’re seeing the same thing happen with the cyber version of models from OpenAI—the GPT cyber series.

I think this is probably going to end up being the tip of an iceberg, where advanced capabilities that could have security implications, at least to start, and maybe other regulatory implications soon, aren’t built into—or at least aren’t exposed from—the main-series model that’s available to the general public. Rather, they get carved out into more secure, more limited models that are only available to government agencies and trusted users.

Part of me wonders whether this is, in the same sense that you would, back in the day, hear Stallman or other evangelists from the open-source movement complain about the closing off of openness in source code, in computing in general, and in the generality of, say, personal computing. It seems to me like we’re starting to see again, for admittedly well-founded security reasons, the beginnings of the closing off of the generality of intelligence in the name of security.

We know that the same technology that enables OpenAI to deliver this for biodefense is the same technology that terrorists can use to develop new viruses for biowarfare. That’s the scary part.

Dave, are you thinking about this at all? I’m curious, Alex and Ahmad. You can never start a question to Alex with “Do you know?” because the answer is going to be, “It’s got to be yes.” But do you know: did they de-guardrail the existing model for this, or did they train a new set of parameters that are specific to biology and deliver it to the government for this?

Or is it distilled? Well, this is OpenAI, not Anthropic, but my understanding—or guess—for both the Rosalind series and also the cyber vulnerability series that OpenAI released is a combination of scaffolding, unshackling, and probably also a bit of post-training and use of tools that otherwise wouldn’t necessarily be baked in at the back end.

The default tools for ChatGPT are usually Code Interpreter and web search. I think when Rosalind was first announced, they announced that there were a number of other databases and tools that would also be built in. If I were OpenAI and trying to make this useful for biodefense, I would also probably include an unshackling at the scaffolding layer, saying, “You are, in fact, allowed to ask questions about smallpox and about a variety of other subjects that probably would just be completely guardrailed out for the general public.”

Ahmad, you and I have spent time talking about this in terms of serving sovereigns with AI. I’m glad OpenAI is doing this. I think all the labs need to be helping us defend. We’re going to talk about that in the next story as well. Any closing thoughts on this one?

Immad Wak

Yeah, in 2020 and 2021, I remember launching at Stanford HAI an initiative to organize core knowledge and preparedness around that. We had a lot of big labs promise a lot of AI, and none of them would give it because they said it was too dangerous at the time, which is why I actually got into open source. That was a very painful process.

I think Dave was asking the right question. This is an unshackled, grounded version that can really analyze knowledge at scale. We see similar things with the science harnesses now from Google and others, because base models are a lot more creative, but then, on the other side, for preparedness, they can’t analyze huge amounts of knowledge and data—like I said, detecting outbreaks, improving resilience, and so on.

The reality is that we’re probably going to see more of these from threat actors, as well as just timeliness around pandemics. The next big one is probably only a few years away, honestly, when you look at a lot of different things. Hopefully, this time we’ll have our act together.

Because if we don’t, then I think it could be even worse than it was before.

I mean, this is probably what the AI labs should be investing in to befriend the public, because the public needs a reason to be supportive.

All right, our next story related here is a serious one, and it should be. Today, I had the pleasure to co-sign, alongside Sam Altman, Dario, Demis, and Mustafa, an open letter to Congress saying we need laws requiring DNA synthesis companies to screen their orders.

Today, you can go online and order custom synthetic DNA sequences. Dozens of companies will print whatever code you send them. Most screen their customers, but a lot of them do not.

Here’s the scary part: back in 2017, as an example, a Canadian researcher spent about $100,000 on mail-ordered DNA and reconstructed an extinct horsepox virus. The same method could theoretically be used to create smallpox. You can layer large language models on top of that, and you can do a lot of damage.

A dear friend of mine, Olivia Scharfman, at the Institute for Progress and the Foundation for American Innovation, put this forward.

As a result of this, there’s a bipartisan Senate bill already in play. This is one of those rare moments where the industry is saying, “Please regulate us.” We talked about this with Eric Schmidt and Dave: There will be, at some point, some kind of a small global emergency as a result of AI that will get everybody to pay attention. This sort of bioterrorism is probably top of the list. Any thoughts?

Alex Fink

Peter, I don’t want to throw this all on your shoulders or anything, but of the people who signed this, most of them are running big AI labs and have a personal agenda, and then they have IPOs coming up. You’re the one signer who totally understands this issue inside and out and isn’t conflicted in some way.

But it makes complete sense. I mean, why would you not want to do this? It’s like—

Alex Fink

No, no. Yeah, I’m not saying that. I’m saying there’s so much more work that needs to be done here.

Yes. And between you and XPRIZE, this is probably the best agency to save the world from biological disaster. Well, thank you for signing it.

Alex Fink

You’re welcome.

Right now, there is a future in which sensors in train stations, airports, bus stations, and malls are constantly sequencing everything through the air-filtration system and looking for novel sequences. If they detect one, they sequence it and send it to the CDC, and then we start a vaccine program instantly. It is possible. The thing people need to realize is that viruses and pandemics can only move at the speed of airplanes, but information can move at the speed of light. You can sort of stop this in the bud if you’re fast enough and you have enough intelligence. Alex, you and I have discussed this a few times.

Alex Rives

Yeah, I think this is probably both inevitable and inadequate at the same time. I think it’s also part of a broader set of initiatives to gatekeep the interaction between superintelligence and the physical world. I want to draw a split screen. On the left hand, we’re talking about gatekeeping DNA synthesis. On the right hand, 3D printing.

The state of California and other municipalities are trying to regulate what can be 3D-printed, again out of concern that people will 3D-print weapons, guns in particular. I think there is a lively debate that has started to happen about whether the right way—the right choke point, if you will—for AI safety is after the fact, after something has happened. Do you hold the labs or the users responsible for it? Or is the right choke point the time of action, when an intelligence, either at the behest of a human user or autonomously, tries to do something in the physical world via DNA synthesis, protein synthesis, or printing a weapon with a 3D printer?

Or is the right time to regulate it at the time of conception, when either a human—

Or all 3.

Alex Rives

—or all of them: the entire end-to-end thought-to-action pipeline. My guess is different countries will choose different combinations and permutations of each stage of that thought-to-action pipeline. I’m not sure that there is a right answer, but I do think that some combination, some permutation, of being able to—although it probably will enable regulatory capture of the DNA, protein, and biological-sequence synthesis step by frontier labs that will then obviously insist, “Well, you have to use Rosalind. You have to use our model to determine whether the sequence is dangerous. You can’t just rely on naive pattern matching, or worse yet, just naive base-pair matching.” That can easily be routed around.

Some amount of intelligence forward-deployed to the time of synthesis, with frontier intelligence, I think is probably inevitable, at least in the U.S. We have previous examples of this where the Treasury required photocopy machines to—

Yellow dots.

Alex Rives

Yes, to prevent people from being able to photocopy the $100 bill or any kind of money. There are companies like Twist Bioscience, which already screen what they’re synthesizing, and they’ve been lobbying for this for a while. But the screening costs money, and if only responsible players do it, then they competitively disadvantage themselves in doing the work. So this does need legislation. This does need pressure, and it makes no sense not to have this in place.

Totally right. I also don’t think it’s super hard to capture this at the point of the prompt and reasoning trace. If somebody’s trying to build a bioweapon, a chemical weapon, or a nuclear weapon, they have to have a lot of interaction with the AI. It’s not super hard for the AI to say, “I need this reasoning trace to be captured and turned over.” There has to be some inspection.

What if it’s on your Mac Studio?

Alex Rives

You can’t really do it on your Mac Studio. If you could, we would have to ban that globally anyway, which I hate. I don’t like that, but I don’t see any other outcome.

Could you do this on an LLM on your machine, on-premises?

Alex Rives

You can. I was one of the authors on OpenFold, the open-source version of AlphaFold, and supported a range of things in this. You can basically run the prompts on the edge, adapting existing models or using some of the new ESM and other models that have actually been open-sourced. It’s very difficult to stop that. You need to be smart, and you need to understand it. But the threat vector doesn’t come there, which is why you need to look at this side of things.

Again, this is a really good step in the direction of regulating synthesis. I think this actually goes to a bigger thing, which is that the total percentage of tokens in the world used by the public sector is probably 0.01%. Total world GDP represented by the public sector is 20%. We have to use intelligence for public good, and one of those public goods is that the government should be subsidizing or paying for the tokens to analyze every single sequence being put into a sequencer, for example.

And so that’s a bit of a phase shift, because governments don’t think that way. But I think these types of things, where intelligence is causing threats or opportunities to go up, mean we need to ramp up public-sector compute to help mitigate and take advantage of those.

Alex Rives

And that’s a great basis for an XPRIZE. Just say, “Hey, let’s have all the ideas related to the way the government can use that 20% to put together AI processes that stop specific threats at specific points.” Let’s list them all out as XPRIZE opportunities—

—and then throw that out for the world to think about. A model could come back with 50 of them right out of the gate, and those all become XPRIZEs.

Alex Rives

I would add that there’s an enormous hole in this, which is directed evolution of environmental DNA that’s already out there in the biosphere. If we really want to take this seriously, the risk of either non-AI-enhanced bioweapons or AI-enhanced bioweapons—if we think that’s the X-risk scenario we care the most about—we’re really going to have to sequence environmental DNA from everywhere, not just the biosphere, but all environmental DNA everywhere.

And as a baseline, to see deviations from the norm. Absolutely.

Alex Rives

Yeah. Actually, I think no one’s articulated that properly. You don’t want to give people ideas. This is always a thing. But I think that should be a really directed paper where you model it, because once you actually model it, it gets scary really, really quickly, and you have to start putting it in place now.

Yes. Yes. Yeah, that one, like a lot of what Alex was saying earlier, you can wait and figure it out after there’s a disaster and then work back. But this one, you cannot afford a global pandemic that could be so much worse than COVID and then try to work on the problem. You can’t do it that way.

Alex Rives

I mean, listen, we all know this is a very serious risk, and it’s been mentioned so many times as perhaps the most likely risk—

—and why we’re not investing heavily in this yet. I think every politician listening to the show needs to be putting this forward. You want to get the public on your side? Do this, because you’re going to be a hero when it eventually happens.

Alex Rives

Peter, that’s why I was coming back to you as the one signer on that document who has a PhD in biology, who also understands AI inside and out, and everyone else is busy with their IPOs right now. I mean, with their trillion-dollar IPOs, to say it clearly.

By the way, I know that you're busy and sometimes these episodes run long and you don't have time to listen to the whole episode, or if on occasion you miss an episode. I now put out a moonshot summary on Substack, which includes a link to all the stories that we cover. The weekly recap covers what I and the mates had to say, what we think is most important, and what we're most excited about. And it's free. You can subscribe at diamandis.com/metatrends. That's diamandis.com/metatrends. All right, now back to the episode.

This next story is fun. OpenAI Robotics is hiring. Sam Altman posted it plainly: “We are focused on robots to support skilled workers to build our future infrastructure,” and imagining everyone having a personal robot doing anything they need.

I find it fascinating that Sam is following lockstep with all the things that Elon’s doing. A few episodes ago, we talked about Sam potentially investing in rocket companies. So OpenAI is hiring for robotics—not a research paper, not a partnership. This is building an in-house robotics team. It’s a signal about where the frontier labs are heading next. Anthropic is the one lab that’s really focused.

Alex Rives

Well, it is the innermost loop, right? The robots that build the data centers and the fabs that build the chips that host the models, which are powered by the energy that hosts the AI…

And doing it all on the Moon and in Earth orbit.

Alex Rives

That's right. Disassembling the Moon to build an SSO Dyson swarm. Exactly. So, yes, of course, I'm excited about it. I think it's instructive, on the other hand, that Stargate—Sam's signature initiative to go build the data centers within OpenAI—has shifted in the direction of leasing rather than building.

When I see Sam and OpenAI announcing that they're focusing on robots to build their future infrastructure, it makes me start to think that either Sam is going to go ahead and spin up an independent-of-OpenAI data center infrastructure initiative, or maybe OpenAI is seeking to revive Stargate's future from a robotics angle. Maybe it'll be OpenAI robots helping to build or maintain data center facilities that aren't owned by Stargate, but rather are maintained by third parties.

But either way, I think the arrow of progress is clear, both from Sam and from Elon and from everyone else: the most productive use of humanoid robots right now is probably just to build out the Dyson swarm in orbit and to tile the Earth with compute here on land. Then we all can get our domestic robots as a secondary afterthought, even though I'd really rather get mine first.

Uh-huh. Another thing we've talked about in the past is that humanoid robots interacting in the real world are going to be an important source of data. Everybody's talking about, "Where do I get new data?" Well, this is definitely one of them. Emad, any thoughts on this?

Yeah, so it's interesting. We remember Sora shut down, right?

Yes.

And I was like, why, when GPT Image was coming out? So Aditya Ramesh, who led the Sora team, is leading the robotics team. So they've taken all their video researchers and moved them onto the embodied robotics team.

When you look at robotics, I did a whole bunch of work recently. I think I've got a paper coming out soon. It is far bigger than the GPU market—actual physical, embodied robots. They will not have the same depreciation schedule either. I think people are becoming wise to this because GPUs depreciate really fast.

I think we've seen GPU prices going up. Old, used GPUs are going up recently.

Yeah, they go up, but they still depreciate in terms of that. But you look at a Unitree G1 two years ago compared to it being on America's Got Talent now, dancing. Once you get the robotics to a certain level, those things will have a massively long life and they'll be incredibly profitable, especially when you go full stack.

I think, again, this is why OpenAI is realizing that they've got their GPU build-out. Maybe that'll continue, maybe it won't, but the market for robots is way bigger than that. They will tile the Earth. They'll build your extensions or pyramids or go across the universe.

Again, I think it's interesting that their video team is now fully on robotics. They decided and made the choice: We have to move that talent over there.

Dave, you've started to invest in robotics, haven't you?

Dave Blumberg

Yeah, I think Sam and I are on the same page on the fact that this is a very, very good 10-year investment theme, and that the battle for self-improving software is well underway. So if you're maybe 16 or 17 years old and you're thinking about what you're going to do post-AGI, working toward getting this job with Sam is a very, very good next move.

What do I have to do? What degree do I have to get? Who do I have to beg to get this job? Because once you're in that group, you look at what happened with all the software people that were original OpenAI people, and they're thriving at 5 different companies now. They're all billionaires.

Yep. So the robotics version of that is just starting, and you can get in early if you pivot in this direction. Also, I think Sam was early to recognize that whoever controls compute controls everything, and robotic construction of data centers is one part of that strategy, one linchpin.

The custom chips—remember, he cut that deal to design AI-specific chips. I haven't heard much about it, but that's in the works, too. So that's the whole thing Alex was describing: the self-improving loop that includes the hardware construction.

All right, big news this week. Anthropic just confidentially filed IPO paperwork with the SEC and could be the first major frontier lab to go public, beating OpenAI to the punch. We've talked about it on the pod before, with SpaceX coming out at $1.77 trillion. I think that's the price right now. We'll be seeing it very shortly, soaking up a huge amount of the capital out there in the world.

Is there enough for 2 additional trillion-dollar IPOs? Polymarket gives it a 60% chance that Anthropic surpasses $1.8 trillion in market cap on its first day. $1.8 trillion—just fascinating. That's the same price as SpaceX's anticipated IPO. We've entered really rarified territory here.

Dave, are we just getting numb to trillion-dollar IPOs? Is this likely to be just the expected, like, "Hey, when's your trillion-dollar IPO happening?"

Dave Blumberg

Well, people are numb, and they shouldn't be. They should be situationally aware and recognize that this flow of money is the biggest in the history of the world by an order of magnitude or more, and that they should try to be involved in it.

These labs—Anthropic is only 5,000 people—you're talking about an insane amount of money divided 5,000 ways. They're going to want a whole ecosystem of other companies to help them build, deploy, and create. They'll have massive amounts of capital to invest in that. I mentioned on one of the prior podcasts that this group of companies can do thousands of billion-dollar acquisitions.

Yes, it's powder. It's—

Dave Blumberg

Dry powder like we've never seen before. So don't get numb. Unnumb yourself and recognize the amount of opportunity that's suddenly available to you. Be part of it.

It's a good thing for them to go public, right? Anthropic is thought to be the safety-conscious AI company, and being a public company forces them to have enough disclosure. So I think that's a good thing. Immad or Alex, any thoughts here?

Alex Wissner-Gross

Should I sing the Magna Mopa song again?

I'm sure you do it beautifully, my friend.

Alex Wissner-Gross

Thank you. I think it's so essential that we get all of the Magna Mopa companies, and not just 8 of them, to IPO. I could maybe quibble as to whether Anthropic really is the first frontier lab to go public, or whether xAI is technically a frontier lab. Certainly, I think Elon fashions xAI as a frontier lab at this point.

But I think it is a public good that they're IPOing, and a public bad that it has taken this long for retail investors to have access to equity in the Magna Mopa companies. I certainly hope, and will do what I can, to ensure that this is the last time in at least foreseeable history when we see so much private wealth accumulation happening outside of the public markets.

I will do what I can to help make public markets the place that private startups think of first and not last as an aspiration to be.

Emad, you're building an incredible company with Intelligent Internet. You've been raising capital, and you've been building breakthroughs that you can't discuss here, that I know of. How does this kind of price hyperinflation hit you in your efforts to raise capital?

I mean, it's positive: the amount of money in AI is insane, which is why you're seeing these crazy raises. But at the same time, they're not crazy because you've never seen revenue growth like this. What is Anthropic's revenue right now? It's probably $6 billion, $5 billion, $6 billion, so it's coming at 20 times revenue, which is high, but it isn't Palantir.

It used to be 50 times. These things are kind of making it up, and you wouldn't be surprised if you saw $100 billion or $150 billion in revenue from Anthropic next year because it's just that useful. This is the reality. When you build useful things, the market gives you the value.

I think that the trillion dollars, as you said, is just going to come back because it's a wealth-creation event. You feel oddly bad for Sam Bankman-Fried, actually—one of the greatest investors of all time. If you look back at the Anthropic Series B, where he put $500 million—

All those people that bought those shares will be reinvesting straight into the giants of AI.

Yeah, we're making so many billionaires in Silicon Valley, for sure.

Yeah. And outside of that, everyone else is coming in as well. So it's not like, to your first point, is there going to be a lack of demand? No, this will be oversubscribed. SpaceX will be oversubscribed.

I think there's probably a trillion dollars of money that wants general access, and Elon is going to do his best to make that possible. But we're not going to run out.

This gives support to when Elon said triple-digit GDP growth by 2030. Setting aside GDP as a metric, our next story is: How fast are companies achieving a trillion-dollar valuation, to the point we just made? Apple took 42 years to reach $1 trillion. Google took 21 years, SpaceX 24 years, OpenAI about 10, and Anthropic now roughly 5.

Alex Wissner-Gross

But the real story here isn’t how fast they’re reaching a trillion dollars. It’s how many employees they’re using to reach that level of valuation. Anthropic generates about $9.4 million in revenue per employee—$9.4 million, almost $10 million per employee—on about 5,000 people. Apple is about a quarter of that, $2.5 million per employee, and Google is at $2.1 million.

These companies aren’t just growing faster; they’re extracting more value per person, 4 times the rate of the best tech companies. This is Salim’s Exponential Organizations thesis playing out in real time. Impressive numbers. Something will beat this—something soon will beat this. Peter, dare me to estimate when, since we had this discussion.

I dare you.

Alex Wissner-Gross

6 to 12 months ago, we said we’d see the first one-person unicorn, and we achieved that. Dare me to estimate when we’re going to see the first one-person centacorn or teracorn.

Okay, I dare you, Alex. What is it?

Alex Wissner-Gross

Sometime in the next 10 years.

Okay. I think that’s an easy estimate. I would join you in that bet.

Alex Wissner-Gross

Yeah.

All right. I just think the revenue per person is extraordinary, and the question, of course, is: Do we start measuring this as revenue per agent?

Alex Wissner-Gross

Soon enough.

Interesting.

Alex Wissner-Gross

I think the issue with revenue per agent is that agents right now feel like really discrete entities, just as we’ve discussed on the pod in the past when talking about taxing per token. But ultimately, there are scenarios where the boundaries between quote-unquote agents start to vanish, where we see, for example, end-to-end differentiable teams of agents, and we start to ask the question: Is it really multiple agents, or is it really a single team-level agent? It’s not obvious to me that the agent paradigm survives long enough into the future that we’ll be asking revenue per agent.

I think the other really important aspect of this is that the AI economy is now funded to live within itself. Historically, if I talked to bankers a year ago or car dealers a year ago, they would say, “Look, Anthropic and OpenAI can’t become really big companies until they’ve delivered something to me as a bank, and I’m paying them,” because the iPhone didn’t become big until a billion people used it, and all this legacy stuff wasn’t big until it interacted with me in the old-world economy. It doesn’t actually have to be that way with AI.

If Dario wakes up after his IPO and says, “I want to spend $100 billion with you, Peter, to solve these 10 diseases,” he can just spend it. And you can spend it right into the agent economy. Now your agent economy has $100 billion of cash flow. It had nothing to do with JPMorgan. It had nothing to do with Main Street. It had nothing to do with any legacy business.

So this conduit of capital through these IPOs back into the pure AI world could entirely build an economy of its own, and you might not even notice it in Europe. You’d be like, “What happened?”

So, Dave, we’re seeing the ceiling get higher, right, as these numbers are again going into multiple trillions. How does it feel at the bottom rung, when you’re coming in as pre-seed investors into these companies and they’re escalating? Are the valuations climbing faster than ever before?

David Friedberg

Yeah, it’s crazy. The seed-stage valuations of the companies are about where they were because the founders don’t care about maximizing valuation on that day. They care about being in the perfect ecosystem to get on this wagon.

Mm-hmm.

David Friedberg

But the escalation—the timelines are so short and the escalations are so big—I’ve never seen anything like it. Multiple companies here in our incubator went from an idea to a $1 billion valuation in a couple of years, but one of them is on a trajectory to get there in less than a year. I’ve never seen that.

The reason I’ve never seen it is because it never existed in the world prior to AI. I think it’s here to stay. I don’t think this is a bubble or a flash in the pan. I think you can get so much more done so much more quickly than ever before that it’s here to stay, which is why we continually show these numbers to people.

You’re crazy to do anything other than be an entrepreneur. Then you get all the pushback in the podcast notes, but the numbers are just overwhelming.

All right. Here’s some interesting news. We haven’t heard about Microsoft in some time. We haven’t discussed Microsoft on this pod in a good 6 weeks, but they’ve just come back into the scene in a big way. Microsoft just dropped 7 in-house AI models at Build 2026, spanning reasoning, coding, image generation, video, and transcription.

Here’s the thing that should make every AI lab nervous: They’ve built all of these in-house from scratch. No distillation of OpenAI. No reliance on anyone else’s weights. Friend of the pod Mustafa Suleyman, who runs Microsoft AI, put it bluntly: AI training compute has increased 1 trillionfold, with another 1,000× coming in the next 3 years. Pretty extraordinary.

These models are already being tuned for Microsoft 365, with their tuned model for Excel matching GPT-5.4 while being 10 times more efficient. They’ve also announced a collaboration with Mayo Clinic to co-create a frontier AI model for healthcare.

My interpretation here is Microsoft is saying, “We’re not just a distribution layer for OpenAI anymore. We’re building the whole stack.” Alex, over to you on this. Your thoughts? Are they here? Are they in the game again?

Alexandr Wang

No, they’re not in the game. That’s the bottom line up front. And not only that, Mustafa previewed this position for us when we interviewed him at Microsoft HQ. He foreshadowed it for those who want to go back and rewatch that. I thought it was a really fun interview. He foreshadowed the Microsoft–OpenAI divorce.

Before OpenAI, Microsoft was working on all of its own foundation models, and that wasn’t moving very quickly. Microsoft, I thought, very strategically entered into this relationship with OpenAI and got a boost, became the distribution partner, and got the channel.

Granted, this relationship was formed over a period of time, so it wasn’t as if OpenAI launched its foundation models and then Microsoft invested. It was exactly the opposite chronology. But Microsoft really started piling on the capital once it became obvious that OpenAI was onto a solution for general intelligence.

Microsoft, for better or for worse, was in a position where it didn’t have enough compute—maybe talent as well—outside of the OpenAI relationship to build its own in-house, first-party models. That’s pretty perverse if you think about the history of Microsoft and what Microsoft did to IBM back in the day.

That was the same history rhyming now, with OpenAI playing the role of Microsoft to Microsoft’s IBM, basically running away with the next wave. It’s sort of amazing that Microsoft allowed that to happen to it.

That’s a great analogy, buddy.

Alexandr Wang

So now Microsoft is trying to recover from the OpenAI quasi-divorce and develop its own in-house models, but it’s still relatively compute-starved and relatively talent-starved. I’ve looked at all of the models that it launched, and I really want more hyperscalers and more frontier labs. We need more competition, but Microsoft isn’t there yet. These are mid-tier models that are, at best, competitive with models that Anthropic and OpenAI were launching months ago, not current models.

Emad, do you agree?

Yeah, I think you don’t need AGI to make a PowerPoint, right? And Microsoft—

What if it’s a really good PowerPoint?

There’s no such thing. No such thing. I hate PowerPoint. Again, it’s a very different intention. I think history would have been very different if Sam and Co. had joined Microsoft during the coup. That was a big kind of difference.

Almost, maybe. Yeah.

Yeah. What a different world.

Because it’s a very different environment when you’re heading toward AGI versus not. Microsoft had a lot of issues. The WizardLM team was fantastic, and then they went to Tencent because they couldn’t actually build toward AGI. Then Tencent built one of the best open-source models.

Where Microsoft is right now is that they’re at the level of a Chinese lab—a good Chinese lab—and I don’t think they’re going to go that much further, because you get to a certain point now with AGI where it’s a generalist model of different types. But now they’re going to hyper-specialize because they need to serve 400 million Teams users or whatever. Poor guys.

Whereas you’re seeing the closed models in Anthropic and things like that, they’re never going to see the light of day from the big labs, I think. The big labs are going straight to AGI, whereas Microsoft has gone up like that. But they’re not going to 1,000× the compute that they spend on training a model, because that will then require 1,000× the inference costs.

And if you look at their Maia chip, which these models were probably partially trained on and which they can also run inference on, those aren’t designed for large-scale, massive MoE-type models either.

I think that it's, again, humanist, office-based intelligence as opposed to general superintelligence. So, are they supporting the older generation of entrepreneurs and CEOs? You have to remember Microsoft, a 50-year-old company, was the most valuable company in the world for a pretty long run. They did extraordinary things. Nobody lasts that long at the top. Dave, any thoughts here?

David Friedberg

Well, to me, it's a really cool historical case study, and it's a battle of people, not companies. When Alex described it by saying, “I can't believe this is happening to Microsoft, given what Microsoft used to be.” But if you said Bill Gates instead of Microsoft, Bill Gates would not lose. He would do whatever it takes to win. And here Mark Zuckerberg decided what it takes to win is to offer Mark Chen a $1 billion comp package to come over from OpenAI, and he turned it down.

Yeah.

David Friedberg

Can Microsoft offer Mark Chen a $1 billion comp package? Are you kidding me? How? Now, if Bill Gates were there, he would find a way. But it's the difference: it's a battle of people. It's not a battle of corporate brands. Nobody has positioning—even Apple. Nobody has positioning that's a moat that's purely defensible. You just have to have the will to win and do what's necessary to win. It just feels like, for whatever reason, Microsoft is not doing what it takes to get the 5 or 10 cannot-miss great AI researchers to work with Mustafa to build a true frontier model.

Companies have momentum, and Microsoft does not. Anthropic most definitely does.

Don Mucalo

Don, let's talk about cancer. I know that members come in who think they're healthy. It turns out 3.3% of them have a cancer in their body they don't know about.

That's right. The majority of cancers that we screen for aren't necessarily the ones that are taking lives when found at a late stage. We know that when cancer is found early, the chances for a cure are much higher. We know it's much easier to treat a cancer when found early versus when found late. What we're finding in our members is that over 3.3% were found to have these cancers that otherwise wouldn't have been found or detected.

People, you don't feel a cancer until stage 3 or stage 4. If you don't know what's going on inside your body, it's like driving your car with your eyes closed. So, how do they detect cancers?

Don Mucalo

We're doing full-body MRI, and we also do early cancer detection screening. This is very important. These are not typical tools used in the conventional care setting when it comes to prevention. This is a hard thing because currently these are not studies that insurance would yet be covering. But the goal is to collect these numbers, do the research, and work hard to democratize wellness.

Yeah. So at the end of the day, you can know what's going on inside your body. It's your obligation to know. So check out Fountain Life. You can go to fountainlife.com/peter to get access to the latest technology to help you detect cancer at the very beginning, at stage one, when it is curable, before it gets to stage three or stage four in your world of hurt.

All right. This next story pisses me off. This is out of The New York Times. The New York Times ran a piece analyzing 602 goals that Elon stated publicly over 15 years. Their headline: “He only met 19% of them.” I would bleep my own words out toward you, New York Times. Who in The New York Times is setting any audacious goals and doing anything worthwhile?

The framing misses the point entirely. In 2015, he hit 75% of his goals on time. The ones he missed—well, he's still building Tesla, the most valuable car company on the planet. SpaceX is the dominant launch provider by a huge margin. Neuralink is working on BCI and is in humans. xAI is a frontier lab and a hyperscaler. The man is worth trillions of dollars. Who on this planet has set any audacious goals like this and actually met—okay, 19%—but he will hit all of them. I'm clear: he's always directionally right. His timing may be off a little bit. Anybody want to disagree with me?

David Friedberg

Not by a long shot.

Shocked. Shocked that the paper of record is launching, apparently, an ideologically motivated attack against the greatest technologist on Earth. Shocked.

Alexandr Wang

Well, I don't read The New York Times anyway. You could not pay me enough money to read what the editor wants to put into my neocortex. I linked to this in my daily newsletter, and I excerpted the only statistic that I thought was interesting, which was hitting 75% of goals for 2015. I think anything else is really underselling everything that Elon has accomplished and is accomplishing.

For those who will preemptively say, “Oh no, I'm fawning. I'm kissing the ring,” or whatever the cliché line is, that's not what I'm saying. He has accomplished nothing short of miracles in multiple sectors. A hit piece that focuses on which promises aren't on time completely misses the point. I love it when Starship launches and does miraculous things, and when its earliest flights—it's always a test flight—and the headlines are, “Starship explodes and fails.” It's like, you got the point wrong, for God's sakes.

Yeah. Yeah. Look, I like to think of it in terms of VC—one for you, Dave. About 10% of companies return 90% of all returns. You can't do what Elon did unless you actually have this distribution. And it's actually shockingly high.

It actually goes to what we were just talking about with Microsoft. The culture in Microsoft is not to take chances. So, you're never going to have breakthroughs, are you? You'll just repeat what's happening. Whereas the biggest, best breakthrough is when people actually take a chance on model training or trying different things.

You know, the most misleading thing about this article—the whole thing is misleading—but the most misleading thing is it implies that somehow investors aren't happy with their investments in Elon. You go and talk to Antonio Gracias, and he's like, “Elon could invent a new urinal, and I would put $1 billion on it,” because everything he does works. It's just cringey to have them peel out a couple of edge cases, which even the edge cases are not bad. So, it's—yeah, anyway.

Crazy.

Alex Wissner-Gross

That's the media today. You've said it a million times, Peter, that—

The media absolutely has no budget, so they have to create controversy to drive any readership at all. They're just trying to drive your eyeballs to their advertisers. Don't give them freely.

All right, this is a story for my 2 math polyglots on this pod here today, both Alex and Emad. Let me make sense of this. More than 130 mathematicians signed something called the Leiden Declaration, backed by the International Mathematical Union. Yes, there is an International Mathematical Union. They're warning AI-generated mathematical proofs can look completely convincing but contain subtle, hard-to-detect errors. They're also worried that AI companies could end up influencing which math problems get studied and funded—basically steering the direction of pure mathematics based on commercial priorities. Alex, to you first, pal.

Alex Wissner-Gross

This is a bad look for mathematicians. It's a bad day for mathematicians. I read the declaration. I talked about it in my newsletter. I view this as rear-guard action in the wake of the Erdős conjecture being solved by OpenAI regarding planar unit distances.

Are they trying to maintain their relevance?

Alex Wissner-Gross

Yes. And it's not going to work. This is a terrible idea. It's on the wrong side of history. AI is going to cook math. AI is cooking math. AI has cooked math. No number of whiny declarations by mathematicians regarding the cooking of math by AI is going to change that.

I guess you're not a union member.

Alex Wissner-Gross

I am not a card-carrying member of a mathematical union. No. I think we're going to make such tremendous progress in math and in the physical sciences, as you and I talk about in Solve Everything. I think declarations with fear, uncertainty, and doubt regarding AI solving or bulk-solving math or other fields are just on the wrong side of history. I'd rather see mathematical researchers focusing on using AI to increase the body of our mathematical knowledge, not whining about AI undermining trust in math.

These mathematicians who signed this are obviously fighting for their lives and for relevance. Emad, what do you make of it?

Yeah. I talked to a number of very sad mathematicians after the Erdős conjecture, and they're like, “Do I have a future?”

It was so close.

The thing about these types of things is nobody feels close to them because they're like, “It can't be done,” because they don't dare to do the impossible. Truly original, daring math is not what you get tenure for. It's not what you get your PhDs for. You don't look at things in different ways. Like when Perelman figured out the Poincaré conjecture by saying, “This isn't topology; this is a PDE.”

That wouldn't even be—that would be frowned upon, taking that type of approach. So you had to go to a shack somewhere and figure it out. I think it's understandable that you see this, because you're going to see this in industry after industry. The mechanical side, we know AI can do, but now it can actually pull from multiple different areas, like we saw with the conjecture.

We had overspecialization in mathematics in particular. I think the classical thinkers, like Erdős himself, had so much breadth, but now we force people to look at just one thing, and so they can't see out of their hole. Now, I think, using these tools, everyone should be saying, "Well, I can explore other areas."

I think the one thing this declaration does get right is that the models can get a bit weird, particularly when you're pushing the edges of math, and be very convincing and completely wrong. So we do need to have some help with that, but other than that, I think, again, to Alex's point, we should look and take advantage of—

AI psychosis, is it?

It's not quite AI psychosis, because it's more like AI confabulation and confidence. It's what you'd expect a grad student who's really super talented and convincing to do.

And you sometimes will gloss over that because it's so convincing, but it's not classical psychosis, where you're like, "I figured this and this and this, and it's all—"

Super quasicrystals.

All right, here's another story that pisses me off. This is the American Federation of Teachers, 1.8 million educators strong. It just dropped a 10-point plan for AI in schools. Here's the highlight: no screens at all for pre-K through 2nd grade. I can kind of agree with that. AI safety and privacy protections for K–12, limits on AI use in order to keep teachers responsible for instruction and assessment, and my personal favorite: they propose a tech tax on big tech companies.

The AFT president, Randy Wearten, is leading the charge here. This, again, I think we should have an effort to have the teacher unions require the use of AI by all teachers. All teachers should be AI-literate and understand what it means and where it's going. Dave, you want to weigh in?

Dave

Well, I think anyone who proposes a new tax should go straight to jail. I mean, is there any great person you can name that created a great new tax and they're famous for it? Taxing is the least of our worries. In every one of these topics, the very first thing the person proposes to solve a problem is, well, let's generate a tax and then that'll create a pool of money and then some magical person will figure out how to use the money to solve the problem. Just work on the problem.

Yeah. Don't work. We have plenty of ways to tax people; that's not the issue. Work on the actual underlying problem. But I think it's similar to the last story, where people feel threatened and their reaction is to ban, stop, propose stopping, or eliminate—eliminate all data centers, eliminate all use of AI, eliminate, eliminate, eliminate. As soon as you use the word "stop" or "eliminate," you're on the wrong path. So I agree with you, Peter. The story is frustrating.

Yeah, for sure.

Alex Wissner-Gross

If I might, Peter, I'd like to quote Mahatma Gandhi, who said, "First they ignore you, then they laugh at you, then they fight you, then you win." I think we're at the "then they fight you" stage here, both with this story, with the teachers' union, and the previous story with the mathematicians' union.

Alex Wissner-Gross

They're fighting progress, and this is, again, on the wrong side of history. The solution isn't to tax new technologies to pay as a subsidy to old technologies or old ways of doing things. It's exactly as Dave says: focus on the ultimate objective here. If the goal is education, or the goal is discovering new math, focus and shape the charge of superintelligence on that ultimate target. Don't focus on cross-subsidizing old ways of doing things that are less efficient.

I think I'll give you something very practical, again extending from math. If you look at Math Academy, it's one of the most effective AI-enhanced tutors, but it just uses AI to coordinate the different tasks and things. They have an amazing book, The Math Academy Way, that brings all the science of adaptive learning together. I've seen 8-year-olds finish their entire high school just by doing Math Academy.

There's no science to what they're putting forward here. That's the really frustrating thing. So, again, I'd encourage everyone to look at The Math Academy Way and look at that in relation to education, because it's a very interesting, scientifically backed treatise on this, as well as being a great platform.

All right, let's move on to our next story here. Senator Bernie Sanders introduced the American AI Sovereign Wealth Fund Act. The idea requires OpenAI, Anthropic, and other major AI companies to contribute 50% of their stock to a public wealth fund—not cash, 50% of their stock.

His goal is to have every American effectively own a piece of the AI revolution. It's the boldest proposal yet for distributing AI-generated wealth. A 50% equity stake is pretty much a nonstarter politically, but the Overton window is shifting. Even if this specific bill fails, it's going to start to normalize the idea of the public deserving a part of the AI value.

I can imagine a 5% or 10% stake being put forward, structured as a condition for operating on U.S. soil, using U.S. data, U.S. energy, and so forth. Let's listen to Bernie Sanders, and then we can discuss it.

Bernie Sanders

And that is why, in the coming weeks, I will introduce the American AI Sovereign Wealth Fund Act. This legislation would give the public a direct ownership stake in the largest AI companies in America through a one-time 50% tax, not on profits but on stock.

It would do two extremely critical things. First, it would give the American people a direct role in determining the future of this technology. No longer would the future of AI be dictated by a handful of big tech oligarchs while the rest of the world sits back and watches them do what they want.

Secondly, it would guarantee that the trillions potentially created by AI are used to improve the lives of all of us, not simply to make the richest people on Earth even richer. And I have to tell you, this is not an original idea. OpenAI has proposed creating a public wealth fund. Anthropic has proposed national sovereign wealth funds with stakes in AI. Elon Musk has said direct federal payments are the best response to AI-driven unemployment.

The principle is quite simple: when a public resource generates wealth, the public should share in that wealth. Artificial intelligence is being built on a public resource far more valuable than oil.

So, guys, pretty extreme, but I think he's going to get a lot of play on this. What are your thoughts? Dave, you want to jump in first?

Dave

Well, the idea is obviously stupider than stupid. Sorry, Bernie—I don't mean to throw it at you. Do you not understand that you can never trust an administration? If you have equity, you have to sell it to generate cash to give to the people. When you tax something, you generate the money. You can use it to do anything you want with it. That's income tax. That works fine.

If you have equity, you, in theory, have $4 or $5 trillion through this one-time tax. But you have to sell the stock to turn that into any public good. You can't dump $5 trillion of stock on the market. There's nowhere near that much liquidity. You'll tank the whole economy, but you've promised all the people all of this value, and you have nothing to give them unless you sell the stock.

Which president is going to sell it? Well, it's going to be the very first president who's missing their budget and doesn't want to raise taxes because they're trying to get reelected. "Oh, let's just dump all of our stock, and then I don't have to deal with it." So you'll own this stock for exactly 1 election cycle. It's the stupidest idea ever. I don't want to go off—

Alex, it's such a strange future that we live in. Recall that either in the last episode or the one before that, we were discussing how OpenAI was now using its foundation as an instrument for exploring what a UBI might look like and allocating a few hundred million via the OpenAI Foundation.

I predicted on this pod that OpenAI was opening the door to a very slippery slope that would encourage some allocation of either the OpenAI Foundation or the frontier labs overall to be mined for either UBI or UBE. A few days later, that's exactly what's happened.

So, on the one hand, I think Senator Sanders is correct that it wasn't his idea. The frontier labs are basically inviting this, seemingly. I agree with that point. I also agree with the notion of a sovereign wealth fund benefiting, in part—or maybe in large part—from radical advances in technology, including in AI.

I think if universal basic equity—which is to say, we have, say, a sovereign wealth fund that distributes dividends to the population—if that is one of the directions that the U.S. goes in, then I think it's only natural that the sovereign wealth fund would have major stakes either in a broad market index or in particular companies. If it's a broad market index, then, thanks to the Magnificent Seven phenomenon, naturally some of those companies are going to be OpenAI, Anthropic, or SpaceX.

We also see, at the same time, this administration taking 10% stakes in Intel and golden-share equivalents in other companies as the basis, potentially, for a sovereign wealth fund, which was one of this president's first executive actions: to order the exploration by the Secretary of the Treasury of creating a sovereign wealth fund. So I don't think a sovereign wealth fund is intrinsically a bad idea for the U.S. I think it could actually be a wonderful idea for the U.S.

I can quibble, as is probably obvious, with particular execution, like the government forcing itself into these frontier labs and basically forcing them to divest half of their equity to that wealth fund. But I do think some sort of sovereign wealth fund is very likely to happen. I just don't think this is the right way to go about building one.

Alex

Yeah, I think this changes the dynamic. If you're starting an AI company, all of a sudden you're deciding where outside the U.S. you're going to base yourself. You're deciding if you're going to stay private a lot longer and whether you're going to distribute profits differently. It basically changes the way companies think. Iman, what are your thoughts?

Iman

Yeah, I have an alternative proposal that'll be out soon. I think, obviously, it's a lot better, but this is something I've thought about a lot. If it was 1 trillion dollars—half philanthropic, half of OpenAI—it'd be about 2,000 bucks per American citizen. That's what we're talking about here: not very much at all. I think he doesn't realize as well that what that would lead to is the AI companies being too big to fail.

A.W.G.

America could not let them go. And what we've seen with OpenAI is some slight governance issues from that as well. Because it's not like the individual citizens of America will be in control of OpenAI, or definitely not in control of Anthropic with its PBC structure.

Iman

Exactly.

A.W.G.

So it basically entrenches them as the biggest political power in the world.

Iman

And then the controller of the sovereign wealth fund is even bigger.

A.W.G.

So I think it's a massive power gain. I think what the AI companies should do, because, as AWG said, this is a slippery slope, is put Anthropic and OpenAI shares into the Invest America funds of every single child in America.

I love that idea. The quote I pulled out at the bottom of the slide—I'm going to read it again—is, “The principle is quite clear: When a public resource generates wealth, the public should share in that wealth.” AI is being built on a public resource.

So it's interesting where the government is going to start to claim American intelligence, American power, American data: “You've built it on our backs, and you owe us some of that.” The right time to pursue an action like this, I think, would have been before the privatization of the NSFNET or before ARPANET was converted to NSFNET. I think that train has left the station.

The internet is filled with tokens that are contributed by non-Americans and Americans and AIs at this point. I think trying to go through the exercise of rationing or allocating which pre-training tokens are attributable to which persons in which countries is a hopeless problem to solve. Never mind the fact that the frontier capabilities at this point are largely being driven by synthetic advances and not just pre-training tokens from humans.

A.W.G.

Actually, you can use nothing but tokens from people who died and then use those to generate synthetic data from there.

Then who gets the money?

Iman

But mark my words, guys, this is not the last we're going to hear of this. I think what's going to happen is that every single government is going to introduce a token license, and probably the most sensible thing is to tax them on GDP value.

GDP value, you mean?

Iman

Yeah, GDP value.

Our next story: yet another incredible news source, The Washington Post. [Laughter.] The Washington Post has laid out 5 policy approaches for dealing with AI's impact on employment.

Number 1: tax the robots. Okay, original thought. Number 2: cushion the blow with stronger unemployment insurance. Number 3: make workers AI-proof through retraining and upskilling. Number 4: spread the AI wealth through dividends and public ownership. Okay, a page from Bernie. And number 5: do nothing and wait. Full spectrum.

What's interesting here is that every option except do nothing assumes AI will be displacing significant numbers of workers. I want to tie that with the next story, which comes out of Forbes. Here we go. This is the counterpoint to all the doom.

Torsten Sløk, chief economist at Apollo Global Management, put it bluntly: “AI is a net job creator.” Companies are citing AI just to justify cuts that they're making. Anyway, the data and the narrative are diverging. He's looking at actual employment data, not headlines. The data says jobs are not being displaced—not yet, anyway.

We talked about this a lot in a recent story in Fortune as well. Cognizant CEO Ravi Kumar said he's hiring 20,000 graduates this year alone. We talked about this being murky waters. We're hearing, on one side, people saying it's a job apocalypse. We saw Sam and Dario last time saying, “No, we've reversed our position.” And, of course, folks like Senator Sanders are depending upon that doom and gloom to put their action into motion.

Thoughts, gentlemen? Dave, what are you seeing?

David Friedberg

What I'm seeing is incredibly optimistic. It's all coming down to HealthCare.com. Sean Taylor, the CEO there, has discovered that people who understand the industry—which, for him, is health insurance—can write code and build products without an engineer. And that's the linchpin.

Employment across our network of companies is up 2x, not down. If you asked me a year ago, I would have said it was going to go down because AI was going to automate everybody. In reality, the ability for people who previously were not builders to now become builders is having a far bigger impact than any job loss through automation. And so, everything—jobs are going up.

The expansion of the people contributing is not just core technologists, not just geniuses like Alex and Ahmad, but actually anyone who understands any business can now be a builder and a creator within that business. And so it's a broader pool of talent than ever before that's participating. It's all looking very, very good right now.

Iman, how's it look on the other side of the pond?

Immad Wak

I think we've only just achieved actually competent intelligence, right? Us on this pod, we're at the cutting edge, and everyone's receiving the harnesses and other things. We're not really seeing job losses yet, but we're not seeing job hiring. I think we're starting to see the first aspect of that in the data.

However, the AIs will be incredibly competent next year, as will the robots, and that's the real danger. And that's why you've got the policy things on the other side. I think we've got to prepare for that future because it's inevitable, and we've got to articulate the future which we want, which is: the robots do everything, and we have really fun lives exploring the universe and doing arts and culture, and there's no hunger and there is no disease or anything like that. And we explore the universe in that.

What does the flow of money look like? What does the ownership of infrastructure look like? So let's look to that future.

Alex.

Alex Danco

I would say that this is almost, maybe inevitably, turning into the moral-panic episode. I want to make sure that we lift our sights and not get bogged down with all of these, I think, morally panicked narratives of AI destroying jobs. That's not the long-term outcome.

We are so well positioned with superintelligence to solve truly hard problems for the first time, and everyone who's hand-wringing—if you might permit me to say that—they're just focusing on jobs that are going away, if they're going away at all, that probably humans shouldn't have been doing in the first place. We really want a humanity and a civilization where people are able to solve the most interesting, the most hard and valuable challenges.

Yes. Do what they love, right? Did you want to be a call-center person? Do you really want to be working at Amazon, shipping and packing and shipping? No. Extraordinary.

We talked about this a little bit on the last pod: it doesn't look like there's job loss as much as a pause on hiring. So I still believe early-entry jobs—and this is the big push toward, “Please consider becoming an entrepreneur.” If you can't get a job, build a job for yourself.

David Friedberg

Well, look, in our companies, the long-tenured employees at the more mature companies are builders, and they're thriving. The people coming out of college can't find a job to save their lives, so they're becoming entrepreneurs. That's sort of 90% of the story right there.

Moving on to NVIDIA. NVIDIA is about to drop its first Arm-based PC processors, the N1 and the N1X. The N1X is the beast: 20 CPU cores, 6,144 CUDA cores, which puts its GPU performance on par with the RTX 5070 in a laptop chip.

This is the direct shot across the bow to Apple, Intel, and AMD. NVIDIA has owned the discrete GPU market for years. Now they want the whole processor. Alex, what do you make of it? Are you excited?

Alex Danco

I'm surprised, yes and no. I probably won't end up using it because it seems highly unlikely that Apple would ever adopt this, and I'm mainly on the Apple ecosystem for laptops. But it's not forever. It's sort of a headscratcher for me that NVIDIA has taken this long. NVIDIA is 8 to 10 times Intel's value at this point. NVIDIA tried unsuccessfully to acquire Arm. Why is it taking NVIDIA this long to launch a serious effort to take over the laptop CPU space?

Maybe one could argue, in connection with the Vera CPU portion of Vera Rubin—CPU plus GPU—that it's timely for NVIDIA to finally reattack the laptop space. But it's a bit of a headscratcher to my mind. Maybe you all have a better head canon for why it's taken this long.

Yeah, NVIDIA. Wait, I have a theory to run by you.

Alex Danco

Yeah.

Sorry. No, no. Here's my theory: the laptop industry is tiny. The smartphone industry is so much bigger, and then the data center industry is so much bigger. So why—who cares about the laptop industry?

Well, you care if you think it's a leading indicator to the whole OS becoming AI, and you want to get your toe in the water with laptops because Microsoft is moving to an AI-oriented Windows, and the laptop is going to become the conduit, kind of the test bed, of consumers who purely interact through AI. If NVIDIA gets through this channel, they'll have their first direct-to-consumer contact. Right now, they have no consumer contact at all. They don't get any consumer data. They don't get any automated profiles here. They'll start to gather that information for the first time, and they'll be well positioned when AI can suddenly disrupt Apple. What do you think?

Alex Danco

I think that's plausible. I'll pose maybe a complementary conspiracy theory, since this is Conspiracy Theory Kremlin Corner. I think we read stories every day about how cheap smartphones in Africa are no longer continuing to exist because of the semiconductor shortage and global memory shortages. We also are tracking that NVIDIA is now the majority of transistors coming out of TSMC at their bleeding-edge node. It's no longer Apple.

So maybe a complementary theory would be that NVIDIA basically has this pipe into frontier-node semiconductor production, and many laptop vendors that want access to frontier nodes may be expecting to get it from NVIDIA, thanks to NVIDIA's new distribution muscle. Maybe Intel or other Arm licensees—remember, Intel used to be an Arm licensee but gave it up—maybe NVIDIA, insofar as it has this amazing pipe into TSMC production, is now the best way to ensure bleeding-edge nodes and other compute for laptops that otherwise would be pushed out of the frontier.

Yeah. You know what's interesting about that particular conspiracy theory is that NVIDIA and Apple collide like crazy at TSMC. They're both fighting to get capacity from TSMC because they both can sell as much as TSMC is willing to make for them: Apple through Mac minis or whatever, and Jensen through data centers.

But years ago, according to the lore, Apple really pissed off Jensen by not using NVIDIA GPUs in the Macs, and it really, really practically destroyed NVIDIA. So maybe there's some legacy bad blood there, and that could be a factor in this decision. You mind any thoughts here?

David Friedberg

Yeah, I think this is a blocker to an AMD Strix Halo play with the integrated 128 GB chips that we're seeing coming out of them. The play here is Jarvis: intelligence throughout the home. It's upgrading your home to have that intelligence at the edge.

If you look at something interesting here, it's an RTX 5070. It's good enough, fast enough, and relatively cheap enough to run a 3-billion-active-parameter model. Or if you look at Liquid AI's latest LFM-1B model with 1 billion active parameters, that one's super-fast. NVIDIA has gone aggressively now with Cosmos and NeMo alliances into fully open source. They're going to provide the intelligent substrate and sell chips to the very edge here to block AMD, because AMD is coming up as a threat, and to try and own Jarvis at home, I think. I think that's what this play is.

All right. Let's stay in the innermost loop. But this move from chips up to data centers—we've talked about this topic a number of times in the past. People's concerns about water use. We've debated it, we've heard about it, and people are protesting in the streets: “Not in my backyard.” All right. Here is a poignant presentation by Satya Nadella responding to this issue.

Satya Nadella

It changes with the cooling system, right, and water. So, in fact, the cooling loop is filled once, and the data center can operate effectively with zero water consumption. In fact, the daily water usage over the course of an entire year is roughly equivalent to what a single restaurant would use, right? I mean, that's—

Dave, that's one of the coolest Satya clips of all time. I gained so much respect for him. I had to take it on. I had to show this. Let me throw in one other data point here because I'd seen this reported a number of times. I went and looked it up.

How does data center water use compare to California almond farming? Almond farming uses 1.3 trillion gallons per year, compared to all the U.S. data centers that use 150 billion gallons per year, right? 1.3 trillion versus 150 billion.

David Friedberg

Just almonds.

So, just almonds. Everybody out there, please protest. Stop eating almonds. Crash the market. We need the water. Stop going to restaurants. I mean, it's—Dave, you want to continue?

David Friedberg

Well, look, Satya took it on. I love the fact that he took it on. But the point isn't that. The point is the haters are going to hate. They'll find something else to hate. That's just what they do. Taylor, ask Taylor Swift.

It's cool that Satya took this particular one on. They're just going to move on to something else. And look, at the end of the day, you want to know the actual scientific truth underneath. That's why this podcast exists. Water was never a problem. Post anything you want, but water and data centers are not polluting things. They're not nuclear reactors, which are also safe now, by the way. But I don't want to touch that. Real data centers don't pollute.

Data centers don't destroy your local economy. Data centers don't cause traffic in your neighborhood. Data centers are really a wonderful thing to have in your state or in your community for employment, for your tax base, all of those things. In the future, we're going to start to see all of the hyperscalers actually delivering lower-cost energy to your neighborhood, right? Make that deal. You want to build a data center, give us free electricity for our 100,000 citizens here. Emad, what's your take?

Yeah, I think this all came about because someone did bad math. I think it was Karen Hao in Empire of AI, where she accidentally put 1,000 times the water usage. She said it's the same as almonds. It's the same as U.S. golf courses. It isn't a big deal, but there's fear, and people will attach to anything here because they don't feel part of that control. So, again, we need to figure out ways to enable that to happen, for them to be part of the story.

Yeah. Alex, final word here.

Alex Fink

I think the default outcome is the data centers are going to sun-synchronous orbit anyway, where no one can credibly complain about their water usage. That said, if they stay on Earth and the SSO-based Dyson swarm doesn't happen, I want to recommend to Microsoft and Google that they take these complaints head-on and, to the extent that data centers are already collocated with electricity production facilities, maybe consider also collocating them with water production facilities, with distillation facilities and other facilities that produce clean water. Just own the narrative.

Yeah. Yeah, own the narrative. That's great advice, by the way. If you take a page out of, well, all of recent history—Elon Musk or the Trump administration—just talk more and just tell the truth as you see it, as aggressively as you can, and don't pander to nonsensical arguments. Even though you'll generate haters, you'll also win in the long run.

Alex Fink

Yeah, the truth needs to win out.

Which brings us to this next story. Trust in media has hit an all-time low. It's down to 19%. 1 in 5 people trust what they watch on the news. Again, you could not pay me enough money to watch the Crisis News Network and have some producer decide what's going into my neural net. It's insane. This is only going to get more harrowing as we have more AI-driven fake news and narratives out there. Any comments, gents? Alex?

Alex Fink

I was at The Washington Post when they bought CourseAdvisor and they were going through this crisis. At The Washington Post, nobody's at fault here. The writers got into writing because they want to tell the truth. They want to write the best possible articles. They want to do investigative journalism.

And then the managers are struggling with the loss of all the revenue to the internet. Without the revenue, they have no choice but to generate more and more garbage, fluff, and arguments—anything to generate an audience. So it becomes more like a talk show, a late-night talk show, and less like news every day. That creates this death spiral: the writers get frustrated, and then they quit or get fired because they're spending too much money developing a real story instead of just reading one off the wire. It starts this spiral.

So there's no one at fault here. It's just an unstoppable trend toward zero. If you're watching this on YouTube, peak trust in media was about 80% in the mid-1970s, and it's been an absolutely straight-line decline to 19% today. It sort of hits zero by 2030 or thereabouts, where media disappears. So I think you need to pick the people you trust, people you align with and understand, and hopefully, from the feedback we get on this podcast, you know that we're here to share with you how we feel—absolutely openly and deeply.

Alex Fink

And we don't pull our punches. I'll speak for myself, at least. I'm all in doing this podcast. I'm not pulling any punches. This is from the heart. I'm trying to do my best to call balls and strikes as I see them. Message to the audience: Don't trust all those other media filter bubbles. Trust this one.

Look, I think trust is built in one way: by helping people. This basically tells you the media is not helping people.

Alex Fink

Yeah.

And so the thing that I'm very saddened by is, of all the entrepreneurs—maybe Dave put out a call for this—why don't we use AI to build an actual transparent, trustworthy news source? Use open models, have open reasoning traces, and allow people to contribute. There's a massive space there, and I've seen nothing of people actually using this wonderful technology to build a trusted site for news, opinions, and other things like that.

Alex Fink

And trusted reporters as well.

Yeah. Enable the best reporters to actually report. Again, I think there should be a new lab for trusted news.

Alex Fink

That would be a massive thing.

This episode is brought to you by Blitzy, autonomous software development with infinite code context. Blitzy uses thousands of specialized AI agents that think for hours to understand enterprise-scale code bases with millions of lines of code. Engineers start every development sprint with the Blitzy platform, bringing in their development requirements. The Blitzy platform provides a plan, then generates and pre-compiles code for each task. Blitzy delivers 80% or more of the development work autonomously while providing a guide for the final 20% of human development work required to complete the sprint. Enterprises are achieving a 5x engineering velocity increase when incorporating Blitzy as their pre-IDE development tool, pairing it with their coding copilot of choice to bring an AI-native SDLC into their organization. Ready to 5x your engineering velocity? Visit blitzy.com to schedule a demo and start building with Blitzy today.

All right, let's jump into one of my favorite subjects: longevity and health news. This one is fascinating. Out of Russia, it's been reported that Russia has committed $26 billion to anti-aging research after making longevity a national priority. The targets are ambitious: 3D-printed human tissues, transplantable organs, and epigenetic reprogramming, all by 2030, by the way. That aligns immediately with our $101 million Healthspan XPRIZE, and with what Ray Kurzweil calls longevity escape velocity by 2033.

Their goal here is to save 175,000 lives by the end of this decade. I've had this conversation with Elon when he was in the White House. I said, “You want to help save the budget? Make people healthier.” Right now, in the United States, the average lifespan is 79, thereabouts. The average healthspan is 63. You're spending your last 16 years in pain and suffering and spending all your money.

If we could reverse that and give people health, they'll be more productive and they'll continue to work. The goal here is that you have the aesthetics, cognition, and mobility that you had at 40 when you're 80. If you've got that, you've got GDP—or whatever we're going to call it in the future—going through the roof. Alex, any thoughts on this?

Alex Fink

I'm having flashbacks to For All Mankind, where, without spoiling it too much, I think in the 2nd or 3rd season, Russia prematurely pulls back from Afghanistan and instead invests its entire GDP in its rocket program and in continuing the space race. Imagine an alternative history where, instead of invading Ukraine, Russia just decided it was going to pivot its entire GDP to longevity.

Oh, my. Everybody would come to you.

Alex Fink

Yes.

Go on, please.

Alex Fink

I was going to say that, to the extent that the agenda of Russia's leadership is either rebuilding some sort of pre-end-of-Cold-War empire or increasing its international prominence, I think that would have been a far better and more effective—and, by the way, far less deadly—way to do it. Establish Russia as the world's leading longevity outlier rather than just killing and wasting money.

A couple of points here. One, if you're the absolute ruler of a country and you're in your 60s or 70s, where else would you spend your money? Especially if you continue being the president forever.

The second thing is, the country that's able to really scale longevity the fastest is going to have massive GDP growth. When I'm on stages around the world talking to wealthy audiences, whether YPO groups or hedge fund managers and such, I say, “How much of your wealth, honestly, would you spend for an extra 30 healthy years of your life?” If they're honest about it, it would be nearly everything, right? It's hundreds of trillions of dollars of long-term potential.

Crazy.

It's worth noting that there are numerous billionaires funding this work besides national leaders. Here's one of them: Brian Armstrong, the CEO of Coinbase and a friend of the podcast. We're going to have Brian on this show next week. NewLimit, co-founded by Brian and Blake Briars, just raised $435 million at a valuation of $3.1 billion. They're developing therapies for epigenetic reprogramming to reverse cellular aging, with human clinical studies targeted for next year. Their first indication is alcohol-related liver disease.

Just to make the point here, if you're doing research with the FDA, you can't use aging as your target. You have to pick an existing disease that's identified and reverse that. If it happens to reverse your aging, so much the better.

So, besides this, obviously, we've got Sam Altman backing Retro, and we've got Jeff Bezos and Yuri Milner with Altos Labs—a lot of capital—and we have the $101 million Healthspan XPRIZE to reverse functional aging by 20 years. Who wants to talk about epigenetic partial reprogramming for everyone?

Alex Fink

It seems that everyone has an epigenetic reprogramming startup, usually AI-based or AI-guided. I think this is wonderful. It's borderline miraculous that there are so many different startups all tackling epigenetic reprogramming from different angles. There are many different tissues that would benefit from it.

I would query whether focusing on particular organs or tissues is going to end up being the optimal strategy, or whether maybe some sort of long shot involving more systemic exposure to, say, GLP-1s or 3rd-, 4th-, or 5th-generation GLP-1 derivatives ends up being what actually helps us achieve longevity escape velocity. But one can quibble over the organ-versus-organism divide. I think this is wonderful, and more power to Brian. I'm glad he's spending money on this.

I run a longevity and abundance longevity trip every year in October. Folks can learn about it at abundance360.com/longevity, and we're going to have some of the top epigenetic-reprogramming companies there. We'll have the frontier labs there, and we'll have the XPRIZE there.

Yes. I raised $157 million for this prize. Dave remembers—he's on our board there. The goal is: Can you give someone a therapy in under a year that reverses their functional losses in cognition, immune function, and muscle? In other words, can you give them the immune system, the ability to build muscle, and the cognitive capabilities they had 20 years younger?

830 teams have entered this competition. We expect a winner by 2030. We're in the midst of a healthspan revolution. If you're out there listening to this, you've heard Demis Hassabis say he's going to cure all disease within the next decade. I think that timeline is actually within the next 9 years at this point. You heard Dario say that, at the current rate of growth of AI systems and biology, we could double the human lifespan. His numbers were 5 to 10 years. Let's call it 10 years.

So, a question for you: Are you saving enough money if you're going to be living an extra 20 or 30 years? I just want to put this on people's radar. You want to take the best care of yourself possible so that you can, in fact, intercept these therapies coming our way at light speed. Immad, you've been thinking about this for a while—the use of AI and health.

Immad Akhund

Yeah, I think that it's tractable for the first time.

Bryan Johnson

Like if I told you 10 years ago, “$26 billion into longevity,” you’d be like, “Into what?” Now, I think all of us around this table can see I can definitely spend $26 billion, and it almost certainly will increase lifespan. And I mean, this will save more lives than anything.

So this is why, as you said, the billionaires want to put their money in, but they didn’t know what to put their money in. Now teams are coalescing. You have tractability, and we’ve got to make it so people can thrive. I think it’s incredibly exciting, as long as you hit that in the next few years. You said, “Take good care of yourself.”

Yeah, I think I’m the oldest among this group, so I’m leading the charge here.

Dave

It’s not a competition, Peter.

We all win. Well, here’s an important point I want to make. If, in fact, Vladimir Putin succeeds in his $26 billion journey, we all win. We all have the same biology, everyone around the world. Something that works in Beijing will work in Boston. I think it’s a beautiful thing. We all share the same biology.

Bryan Johnson

Yeah, I think if I just add one thing, Dave, earlier you said, “Everyone get into robotics.” I think if you want to make lots of money and do incredibly well, longevity will have huge amounts of money, even faster than robotics, and is more accessible. So I encourage young people to go full-on into longevity. So, everyone—

David Friedberg

Not my area of expertise, but I 100% agree, and I’m actually—

Bryan Johnson

That’s where I cover you, buddy. That’s where I cover you.

David Friedberg

Yes. Thank you.

Everyone get into longevity and robotics. You heard it here. [laughter]

Dave

There you go.

Daniel Oliver

Daniel Oliver.

All right. Our last story in the pod might be one of the most important longevity stories out there. VERVE-102 is a gene-editing therapy. It’s a single infusion, a single shot that permanently switches off the PCSK9 gene in your liver.

I take a shot every 2 weeks of a monoclonal antibody called Repatha to deal with this. It sort of blocks the PCSK9 proteins there. VERVE-102 just shuts it down. The PCSK9 gene destroys LDL receptors, which are the things that clear bad cholesterol from your blood. Turn off PCSK9, and your liver keeps its LDL receptors and your cholesterol drops.

In a phase 1 trial published in the New England Journal of Medicine, the highest dose reduced LDL cholesterol by 62%. This is the bad cholesterol, sustained for up to 18 months so far. PCSK9 protein levels dropped 88%. Here’s what makes this different: this is a one-and-done infusion. Alex, you’ve been tracking the story.

Alex Wissner-Gross

Oh my goodness, I love this story to pieces. Do you remember the scene in Star Trek IV where Leonard McCoy goes back in time and is at a hospital in San Francisco and walks by a woman who complains that she’s on dialysis? He asks, “What is this, the Dark Ages?” And he gives her a pill, and then by the end of the act, she’s regrown a new kidney and is telling the doctors, “Doctor, some random doctor gave her a pill. She’s regrown a new kidney. Medical miracle.”

This is at the level of Leonard McCoy giving a woman a pill to regrow a new kidney. It combines so many technologies. I love it. It combines CRISPR base editing with mRNA-based drug delivery. It’s a one-and-done shot that basically, to the extent that LDL cholesterol is the lion’s share of the cause of heart disease, it’s basically—squinting at it—this is a one-time shot to cure heart disease.

That is right out of Star Trek IV. This is Star Trek-level medicine that we’re starting to see, so I’m very excited by this.

Bryan Johnson

Editing human software. That’s what we’re doing.

Alex Zhavoronkov

I think that’s what most disease is. It’s just that we haven’t quite figured out the code. This is, I think, the first of many therapies that will be very similar, because most of it is just that our prompts have gone a bit wonky, and this is one of the things that adjusts them.

And it gets better. This drug, if memory serves, was discovered as a result of a small minority of humans who have a mutation that causes them to have naturally low LDL. So you have to ask the question: How many other diseases—how many other human variants in human natural biodiversity—are there for people who never get Alzheimer’s, never get cancer, never get fill-in-the-blank? And are there, out there in base-editing space, comparable therapies that could be delivered via single-injection mRNA LNPs? It’s very exciting.

Amazing. All right, we have a few questions to speed-run with the mates. Here we go. Alex, first choice is yours.

Alex Zhavoronkov

All right, I’ll take question 1, which is: If we get the perfect algorithm or AI, do we even need this insatiable compute-energy budget anymore? And this is from Jackie Lampert, 6KN.

I think this question was directed to me because I talk from time to time on the pod about this idea that eventually we may get to a perfect or optimal AI algorithm at the end of this scaling race. Short answer is yes. I do think we’re going to need quite a bit of compute even if we develop a perfect algorithm.

It’s entirely possible we develop the perfect algorithm, it gives us a dopamine rush of maybe a few more orders of magnitude in terms of effective capability, and maybe we see a DeepSeek demand crash on steroids for about a year until we figure out, Jevons’s paradox-style, how to saturate all of that new capability and capacity that’s come online thanks to algorithmic advances.

But then, yes, I do still expect the horizontal scaling to resume. If anything, if we hit a perfect asymptote, if we hit the ceiling in terms of algorithmic improvement, that puts major new pressure on hardware- and infrastructure-level improvement. Right now, algorithmic improvement, depending on which estimate of the scenario you believe, is probably absorbing about half of all the hardware improvements that we need otherwise.

There’s the classic anecdote: Would you rather take a chess algorithm from 2000 and run it on 1980-level hardware, or take a chess algorithm from 1980 and run it on 2000-era hardware? The answer ends up being you’d rather take a modern algorithm and run it on older hardware.

But if the algorithmic advances stop, if they saturate—now we’re out of further algorithmic improvements because we have the perfect algorithm—that puts even more pressure on hardware improvements.

Dave, number 3 is meant for you, buddy.

David Friedberg

I guess I have to take it, then. “With 170 agents, what is Dave doing with all of them?” [laughter] From New Rave World 9733.

I think mostly just incinerating my bank account. Well, so I’m coming at it—I’m going to flip the question around on you, New Rave. The reason I’m running so many agents is because ultimately we’re all going to want that many, and I’m trying to work backward to how do you make them do something productive.

Over the weekend, I had them all build a particle simulator where the particles have gravity and electric charge, and they’re all interacting with each other. And I asked each agent—

Well, it was just something. It could have been anything, whatever I was thinking of.

David Friedberg

But I asked each agent, “Try and make it as cool a demo as you possibly can,” and then compete with each other. That actually worked pretty well. But I’m trying to figure out how you synthesize work in parallel and get it to come back and be something productive.

Usually I have them working on neural network research ideas, and you can generate thousands of ideas a day, very few of which work, but many of them working in parallel can figure out the good ideas. So I do a lot of that as well.

But the meta-idea here is, look, ultimately we’ll have access to billions of these, and we’ll want to advance humanity with billions of them. So getting a head start on how you wrangle them into a productive workforce is a really good meta-idea by itself, and it’s a lot of fun.

Immad, may I suggest number 2 for you?

Immad Akhund

Can someone now catch up with OpenAI and Anthropic?

This is from @1.156.

Immad Akhund

Yeah. Can someone come out of nowhere—a lab nobody’s heard of—and catch up with OpenAI and Anthropic? I think the answer is yes, but it’s going to be very difficult because distribution effects count for so much.

I don’t think it’s necessarily an algorithmic thing. We’re already seeing potential algorithms that kind of match, but they have a data advantage and a distribution advantage that they’re going to now spend hundreds of billions of dollars to lock down. And this is typically how we see markets in terms of winner-takes-all, unless the lab has a very different distribution mechanism.

I think there are some there, but then it’s not a technological race. It’s more of an “I’m better at go-to-market than you are” race.

I’ll take number 4 from friend of the pod @CJTruheart. How do you measure non-material human abundance?

I love this question, CJ. It’s obviously easy to measure material abundance: lots of goods, lots of Teslas, lots of Optimus robots, and the like. But non-material abundance isn’t impossible to measure; it’s just harder.

I would look at metrics like happiness, access to education, creative output—more music, art, and writing—and connection. The other thing is—and Alex, you and I have discussed this before—it’s optionality and agency. It’s the ability to choose. Are you unconstrained in all the things that you might want to do? For me, those are great non-material measures of abundance.

All right, another round here. Alex, why don’t you go first?

Alex McCaw

Yeah. Well, questions 7 and 9 are pretty similar. I wonder if I could answer a linear combination of those.

Sure, go for it.

Alex McCaw

Question 7 asks: Can anyone offer an empirical, quantifiable, objective—those are the 3 adjectives—definition of AGI? No podcast, interestingly, has ever had a single source of truth. What exactly are we close to? In scare quotes.

Question 9 is: Is there a common AGI benchmark everyone agrees on, or is the missing piece just agreement itself? I think these are both facets of the same question. Well, this podcast has a definition. Actually, this podcast has multiple definitions.

We’ve talked about all of the benchmarks. Almost every time there’s a major new model released from one of the frontier labs, we talk about a variety of evals and how they perform. Those evals are, by and large, correlated with each other.

So I want to answer this question at a meta level. To question 9, the missing piece really is just agreement. We have lots of benchmarks at this point that all seem to correlate with each other, and one can squint and just say, regress a line through all of them and call that AGI, since they’re all pretty correlated with each other at this point.

To question 7, can anyone offer an empirical, quantifiable, objective definition of AGI? If you’re dissatisfied with just pointing at all of these very practical evals, I would say we need to go back to Jürgen Schmidhuber and AIXI, his theory, with a number of other collaborators. Jürgen Schmidhuber was always, you know, sort of the joke in the community because he claims that he invented everything first. Jürgen, this one is for you. I’m giving you credit for having defined AGI.

Take a look at the AIXI theory, which is a mathematical, information-theoretic formalization of what, theoretically, a perfect intelligence would look like. It’s, in some sense, a Bayesian superintelligence that takes the perfect action at any time step in order to perform the optimal actions toward a given objective. If you’re dissatisfied with all of these practical, generalist definitions for AGI that we talk about here on the pod, take a look at AIXI.

All right, Dave, I think number 8 is yours. Why do solopreneurs only seem to get traction at incubators? Isn’t there room to broaden the reach so opportunity democratizes? And that is from Philip T., 8514.

You’re dead right. When we started incubating companies, what, 15 or 20 years ago, less than 10% of all unicorns came through an incubator. Now it’s like 70% and rising, so incubators have completely taken over success.

The reason for that is because time to market is so short. A company like Mercor went from an idea to a multibillion-dollar valuation and then to a $10 billion valuation in 2 years. It’s now in its third year. When companies are growing that quickly, they don’t have time to get office space, figure out payroll, accounting, and food.

If we move into robotics or biotech as the next great frontier for entrepreneurs, just the process of getting a CNC milling machine and starting to grind out parts would take you 2 years. If an incubator already has all that infrastructure ready for you, you can grow much faster in that environment.

The way to democratize it is actually to create many, many more incubators all over the world, not to fight the trend toward faster growth and higher valuations. The fundamental flaw, though, is in the financial structure. Venture funds right now usually charge a 2% management fee, which pays the salaries, but that’s nowhere near enough of a fee to actually build out all the infrastructure needed for a really good incubator.

The investors generally vomit if you go to a 3%, 4%, 5%, or 10% load, but it’s the right thing to do. If we can solve that problem so that investors are comfortable with the incubator structure, then it’ll democratize very quickly.

I am. I’m working on my Abundance Studios to parallel what Linc Studios is doing. Emad, do you want to take number 6?

Number 6 is the real backlash, actually: anti-corporate AI sentiment, not anti-AI sentiment itself. As someone who received hundreds of very nasty messages doing open-source AI when we were doing image generation, I think it’s genuinely anti-AI sentiment that’s stirring up.

I think it’s because it’s come from Anthropic, OpenAI, and others. It isn’t so much corporate as it is fear of this technology that has suddenly gone from being weird to suddenly being good. People can see it looming and coming for them, taking away their agency—to kind of put words in Bernie Sanders’s mouth. It feels like taxation without representation, as it were.

People are being told that AI is being trained on all of their data, and it’s taxing their future, and they have no representation in this. So I think it’s generalized anti-AI sentiment, and it almost doesn’t matter where it’s coming from.

But it’s easy to go anti-Elon or anti-Sam Altman, or some of these bigger-than-life characters, because they are coalescing so much around them. How are you going to fight the shoggoth? It’s very difficult.

I’ll take number 10: How does the average Jane and Joe get a piece of the action, a piece of the pie? This is from Dave Galloway 56.

I would say there are immediately 3 ways. You can buy stock. You can buy Nvidia, Microsoft, and Google, and soon SpaceX, xAI, Anthropic, and OpenAI. Buy it in the public markets and own a piece of that. Even putting away $100 a month into an AI-focused ETF would put you on the right path.

Second, you can use AI to increase your earning power. Learn to use the tools. A real estate agent using AI for market analysis, listing copy, or better communication can do 3 times the number of deals that he or she was doing before.

A third option—we’ve been talking about this pod forever—is to build. The cost of starting an AI-powered business has dropped orders of magnitude. You used to have to hire an engineer, a marketing person, a sales team, lawyers, and accountants. That’s cooked. That’s gone.

I guess there’s a fourth way: You could back Bernie Sanders and have the government take half of it. But everybody can participate. This is demonetized and democratized.

Emad Mostaque, thank you, buddy, for making it past midnight with us. I love your brilliance. Having both you and Alex on this show, I’m really and truly surrounded by brilliance. And Dave, as always, I love you, pal.

That robot has the hots for you, Alex.

Alex McCaw

I love that. I’m also wondering, have we been uploaded to that data center?

Oh, we must have. Amazing. Gentlemen, have an awesome day. See you next week with Brian Armstrong. I’m excited for the SpaceX IPO. We’ll see how it’s going to go. Anyway, living in the singularity, with no better time ever to be alive.