AI业内人士揭秘Elon Musk赢下AI的主计划：AWG与Dave Blundin对谈 | 第192期

xAI launches Grok Code Fast 1. I had to double-check the numbers on this because they were pretty epic. It's insane. Elon went from zero to building Colossus 1 in 122 days. Everyone said it couldn't be done. This guy will not slow down. He wants to be number one.

We're seeing these data centers leapfrogging each other.

Elon, the entrepreneur of all entrepreneurs, knows that it's all or nothing. You don't build the second-biggest data center. You either win the race, or you don't win the race.

This is the bitter lesson as applied to hardware scaling. It's a case study in brute-force hardware scaling, where we're seeing the power, the chips, and the data centers all being brute-force scaled. These incredible tools are demonetizing and democratizing at an extraordinary rate. All the students struggling with their largely irrelevant curriculum: this is what you should be doing.

Now that's a moonshot, ladies and gentlemen. Welcome everybody to Moonshots and our weekly episode of WTF just happened in Tech. I'm here with my Moonshot mates, Alexander Wissner-Gross and Dave Blundin. This is the news we hope that you watch that makes you optimistic about the future, that raises your IQ points 20 and gives you a chance to see the future. Salim is in India right now; we'll talk about that in a minute. So, Dave and Alex, good morning, guys. Good morning.

So, how was Labor Day for you?

I got a lot of grief, actually, for being—

Distracted and antisocial.

I was talking to my AI agents. They were pestering me with progress all week. They have an IQ of 148 now, so it's pretty hard to deny them. They're getting needy for the first time. I've been waiting for this moment since I was 14 years old, so it's pretty hard for me to tune it out.

I missed you. But they are getting needy.

Alex, did you labor over Labor Day?

Absolutely. Never a dull moment.

Yeah, me too. I'm in Port Townsend, Washington, near the San Juans. I've been here for a week with the family, which means I'm getting up super early, like at 5:00 a.m., just to get my email done, work, and do my writing, and then spend time with my 14-year-olds on the beach, in the woods, and fishing.

I went fishing and caught a massive fish, which was about six inches high compared to Alaska. Anyway—

Did you eat it? Just eat it raw, right out of—

Catch and release. Catch and release.

Okay.

Well, so listen, I just want to take a second to appreciate the fans we've had on this podcast. It's been pretty amazing to get the feedback, and you should all know we love doing this. We spend a lot of time working on this to deliver the news that we think is worth learning about and that gives you a positive view of the future.

I want to take a second to read you guys some of the comments and say thank you to everybody for the awesome feedback. So, why not? Jack says, “Moonshots is the best thing I've ever found.” High praise. The Crypto Canvas says, “Best podcast in technology right now. Thank you guys for doing this consistently.” And we do love doing it consistently.

Steve Darus1234 says, “An exciting future. Thank you for providing an optimistic long view amidst the constant doom and gloom of the news cycle.” One of the principles here is that if you are constantly watching all the negative news, it's going to shape your mindset in a really dystopian fashion.

Carl Rankin 5385 says, “Quite simply, the very best and most relevant AI and digital technology podcast available today. Thank you, Peter, for allowing us to hear Salim, Dave, and Alex and their collective brilliance.” You're welcome, Carl.

Renise IB6532 says, “I absolutely love this podcast. You guys are doing a great job keeping up with everything. And yes, Alex is brilliant.”

Okay, let's follow up on that note. From Polymerper, who says, “Polymerper, thank you. I was a little iffy on Alex at first. Then I realized I was just jealous of his intelligence. Now he's my favorite to see in the world.”

That is exactly my experience with Alex when I first met him years ago.

Oh, he's great. A sweetheart of a guy.

We'll wrap it up with Bill Jacobs 30386, who says, “The Fab Four are back.”

Fab three.

Well, today it's the Fab 3. Yeah. Before we jump into where Salim is, I just want to say thank you to our production team, who've been amazing: Nix Dacanay and Gianluca Mignano. Thank you guys for all the hard work you do making this easy and fun.

Right? I mean, it's pretty amazing.

So, Salim is in India, about to get on stage with his Singularity University mates at an SU summit there. Before we hung up with him a few minutes ago, our edict to him was: bring back a box of iPhone 17s and please fix the U.S.-India trade issues. So, he's taking that on. We'll see how he reports out.

And this is, Peter, back-to-school week. I don't know if you're in phase, but everyone is back on campuses now, grinding away on these soon-to-be-irrelevant curriculums that are falling by the wayside. I got lots of questions from my kids, nieces, and nephews over the weekend about what they should be doing and what they should be studying. It's so great that we have Alex here to help add to that guidance because it's changing so quickly, and it's very, very hard to keep up.

I know that what they're learning is irrelevant and becoming more irrelevant by the minute. We've got that much figured out, but then what is relevant, and how are we going to keep up with it? Class starts on Thursday morning.

Well, MIT Foundations of AI Ventures—

My boys as well, but not quite MIT curriculum. They're in eighth grade, but hey, that's good.

Well, I don't know if you know, but MIT added a new thing this semester: 6E. Remember how Course 6, which is computer science and EE—

Had always had 6A, where you go to companies for a semester or two and learn how the real world works.

They added 6E now, which is incredible. E is for entrepreneurship.

So, you basically take a couple of semesters and go either work at a startup or a venture fund—

And see how the startup world works. It's incredibly popular.

This semester, I'll be teaching advanced algorithms in that curriculum a couple of times this semester, and then full-time the following semester.

Yeah, I do believe the career of the future is entrepreneurship, period. We should have that conversation, and we should talk about education on the next pod. We go a little bit deep on that, and there's some news developing there.

Alex, I don't want to miss this point. Right now, if you're an incoming freshman to college, what's your recommendation? Skip it? Skip college? Or what do you do?

It's a tricky time. I think it depends entirely on the freshman's goals. If your goal is to build a startup, I think there's a strong macroeconomic incentive to just do it now. Consider dropping out and moving to Silicon Valley, or doing it in Boston or elsewhere.

But I think AI, AGI, and ASI timelines are so short that almost any conventional career plan—if we'd had this conversation 20 or 30 years ago, I think it would have been far easier to project out a conventional life plan or career plan. I think now the singular bit of advice—no pun intended—I'd have for any college freshman is to assume that AI timelines are incredibly short.

Assume that we're going to have superintelligence, to the extent it doesn't exist somewhere already and just isn't evenly distributed. Assume that we're going to have superintelligence in the next 2 to 3 years, and guide your career plans accordingly.

Yeah, I'll add my opinion there, which is, as I've said over and over again: find a problem you're passionate about, right? The technology is going to constantly change, but the problems are going to be fundamental for some time. Then apply intelligence to that problem. Apply AI to problems that you care deeply about.

If you don't know your massive transformative purpose, wherever you are—high school, college, or graduate school—pause what you're doing and really focus on what's your driver. My favorite Mark Twain quote is, “Two important days in your life: the day you were born and the day you found out why.” So, why are you here? Then apply AI and digital superintelligence to that why.

The other one, Peter, of course—the other Mark Twain quote that's apropos to the college experience—is, “The classics are books that everyone wants to have read but no one wants to read.” So maybe those 2 end up colliding in this case.

We have Google NotebookLM summaries now, so we can listen to those books in brief in a podcast.

Every week, my team and I study the top 10 technology meta trends that will transform industries over the decade ahead. I cover trends ranging from humanoid robotics, AGI, and quantum computing to transport, energy, longevity, and more. There's no fluff, only the most important stuff that matters, that impacts our lives, our companies, and our careers. If you want me to share these meta trends with you, I write a newsletter twice a week, sending it out as a short two-minute read via email. And if you want to discover the most important meta trends 10 years before anyone else, this report's for you. Readers include founders and CEOs from the world's most disruptive companies and entrepreneurs building the world's most disruptive tech. It's not for you if you don't want to be informed about what's coming, why it matters, and how you can benefit from it. To subscribe for free, go to dmandis.com/metats to gain access to the trends 10 years before anyone else. All right, now back to this episode. Should we dive into the AI wars?

Let's do it. Ready?

All right. Absolutely. Let's see where we are first. First up is our friend Elon and his rollout of Colossus 2, coming online in a couple of weeks—a 1-gigawatt data center in Memphis. I love the fact that we're beginning to talk about these in terms of energy, not the number of GPUs.

Again, this mythical 1-gigawatt center is finally coming online, fitted for 500,000 NVIDIA Blackwell GPUs and then doubling it again next year, in 2026. This is where Grok 5 will be born. Let's talk about it. I mean, this guy will not slow down. He wants to be number 1, and we're seeing these data centers leapfrogging each other. Alexander, what are your thoughts here?

This is the bitter lesson as applied to hardware scaling. It's brute force. It's sort of a case study in brute-force hardware scaling, where we're seeing the power—as you mentioned—the chips, the data centers, all being, in the style of the bitter lesson, brute-force scaled.

What we haven't seen yet, and will, I think, be very interesting to see, is that same level of brute-force effort being applied to the software side of the stack. I wouldn't be surprised to see that kick in as well. But for now, in the style of the bitter lesson, it's just absolutely incredible to watch, with such vigor and such brute force, the hardware-scaling side of the equation taking place.

The bitter lesson is from a guy named Richard Sutton, who made this, I think, incredibly important observation that has permeated the AI community. Alexander, it'd be great if you gave us a quick summary of it.

Sure. I'll caricature it, but the core thesis behind Sutton's bitter lesson is that all of the decades that AI researchers spent developing artisanal solutions to problems—to speech recognition, to language understanding, to computer vision—were all basically wasted. In the end, all that really mattered was taking large data sets and lots of compute and off-the-shelf algorithms and just scaling them up to very large sizes.

So this is the bitter lesson. This is why it's bitter: all of this effort by humans, by human researchers getting their PhDs over the decades, coming up with artisanal new algorithms that they could publish. You remember, several years ago, it was breaking news when researchers were able to build a cat detector. That was news.

That was just 5 years ago, right? Not even a long time ago.

And that's just been steamrolled by completely general algorithms with very little human-injected prior knowledge, combined with huge amounts of data and huge amounts of compute. So the bitter lesson has been, I think, one of the core themes of the AI revolution we've seen over the past few years.

Going back to Colossus 2, the bitter lesson applies not just at the software layer but also at the hardware layer.

So, brute force.

Yes.

Yeah. What I want to talk about here is, again, just this leapfrogging of data centers and the hyperscalers. Elon went from 0 to building Colossus 1 in 122 days. Everyone said it couldn't be done. Right now, he's up against others, and he's built the largest data center and is going to maintain that lead.

We're going to see OpenAI coming out with the Stargate centers. So how will Stargate compare to Colossus?

Toe to toe, actually. Isn't it? They're exactly in line. This foot race is very much winner-take-all. I know we had that debate the last time we podcasted. It looks like there'll be 5–7 big AI labs, but I think Elon, the entrepreneur of all entrepreneurs, knows that it's all or nothing.

You don't build the second-biggest data center. You either win the race or you don't win the race. These things, once they're trained, compile down to something crazy fast and very easy to tailor into specific use cases. But nobody wants to start with the number 2 model or the number 3 model.

And so you either win the race or you don't. Elon's all in, Stargate is all in, and Sam Altman's all in. But, yeah, Alexander, do you know the exact numbers? I think they're pretty much right in line.

Also, the numbers are a little bit tricky because I would distinguish between data centers for training new models and data centers that are going to be used primarily for inference. The world appears to be moving toward just tiling the Earth's surface with inference-time compute models.

It's important, just for a second, for our viewers and listeners, Alexander, to talk about the difference between training and inference. If you haven't heard the terminology before, it's fundamental.

Sure. Conventionally, the way one would think about this is that an AI model like ChatGPT is trained. It's created basically from large data sets at one time—a fixed cost upfront—and then later on there's so-called inference time, when the model is actually used.

If we were to analogize this to software engineering, there's compile time, or development time, when a computer program is created, and then execution time, when a computer program is actually run. Same idea here. Inference time is when an AI model, like the GPT series, for example, is actually deployed and being run in practice.

Those inference-time use cases, based on the headlines that we're seeing, are going to be run in data centers around the world. We're standing up data centers as a human civilization all around the world—in the Middle East, in India, in Norway, elsewhere. But we'll get to this later.

The training time, when the frontier models—the strongest models that we have—are being created, appears to be more geographically localized in the US at the moment. So, going back to Dave's question, it's a tricky distinction between data centers that are intended primarily for training and creating new strong AI models versus data centers that are intended primarily for running models that already exist.

I think another analogy here is if you go to school and learn a language. You'll spend a few years learning a language, but once you've learned it, actually speaking it is a lot quicker. Inference is just that: having the language uploaded into your neocortex and being able to speak it.

I think one other thing worth mentioning here is that Elon has basically got access to infinite capital. Every time he goes to raise capital, it's oversubscribed. There's a massive amount of family-office money and sovereign-wealth money that's just prepared to fund his continued growth. It's never been that way.

I think what we're going to see here is that the differentiator for the United States in building these companies and building these frontier models is access to this risk capital, which doesn't exist at this level anywhere else. I think that's pretty awesome.

Yeah. Well, just to put some numbers on that, this is 1 million Blackwells. They're $30,000 each. Remember, a GB200 has 2 Blackwells on it. So I'm pretty sure he's talking about 1 million Blackwells, not 1 million GB200s.

Alexander, you probably know the answer. Anyway, 1 million Blackwells at $30,000 each—that's a $30 billion investment in the chips alone. Then whatever on top of that for the racks, the power supply, and all that.

Amazing. It's also probably worth briefly mentioning the electricity supply side of this. Colossus 1 famously has a self-contained electricity source. It's using natural-gas cogeneration facilities on-premises. To my understanding, it's not drawing electric power materially from the grid. It's generating its own electric power on-site.

One can extrapolate all sorts of interesting questions. Does this mean that there's going to be a pocket economy of data centers that are being forced to co-locate with nuclear power plants or natural-gas cogeneration simply because the rest of the grid and the outside economy are too slow to catch up?

Yeah, I think that's going to be the case, right? Where do you have cheap electricity? Just move your data centers there. All right, let's go on to the next one.

Yeah, go ahead. Oh, so Daniela Rus and I took a tour of the Markley data center here. It's the first quantum deployment, but Jeff Markley—a great guy who built the data center—bought all the 3-megawatt generators in the country.

I said, “What do you mean, all of them?” He said, “Well, all the 5-megawatt ones were already sold out, and I panicked. I'm like, you know, because we need to generate a gigawatt. There are only so many of these generators.”

So now that's the issue. Even if you get access to a power supply, you need the generators to turn it into electricity, and those things are completely sold out.

One of the things I want to talk about on the podcast here is where you would make your next investments. We missed the Intel option call we talked about last time with Leopold. I'm thinking—and this is not investment advice; this is me advice—that investing in the cutting edge of energy production is worth considering.

The drawdown right now, as Eric Schmidt said to us, Dave, when we were having our podcast with him, is that AI is energy-limited—not chip-limited, not intelligence-limited. It's energy-limited.

Well, I’ll tell you what, guys. Why don’t we go down Leopold’s holdings from his 13F filing on a podcast maybe next time and look at everything? His biggest position isn’t just Intel; there’s a whole list of things there that are all direct implications of what you just said.

And so, let’s just analyze them one at a time.

You want to turn this into a financial investment podcast? Okay.

We could label it. So, if people want to skip that one, they can.

All right. The next topic for discussion is X, keeping on the Elon theme here. xAI launches Grok Code Fast 1. I had to double-check the numbers on this because they were pretty epic. This is optimized for agent coding, right? You see here on this graph of model performance per 1 million tokens, Grok Code Fast 1 just trounces everything.

I put the numbers down at the bottom here. Input tokens are $0.20 per million tokens, and output is $1.50 per million. Compared to GPT-5, input is $1.25, compared to $0.20, and Claude Sonnet 4 input is $3. So, we’re talking about 15 times cheaper on input tokens and 10 times cheaper on output tokens. How do you compete against that? I mean, this is a race to the bottom. Thoughts, Dave?

It’s definitely not a race to the bottom, even though it appears to be. This is a “get the market share, don’t lose no matter what” strategy. People get addicted to this stuff so quickly, and then they want an infinite supply of it. So, this is much more like a crack dealer giving out the first hit for free than it is like a race to the bottom.

I think people are completely misinterpreting whether AI is a race to the bottom and also whether the chips will commoditize. Neither is going to happen because the demand is infinite.

Yeah. Alex, your thoughts here, buddy?

What’s worth noting is that if you try to interact with Grok Code Fast 1 via a browser, you will not find it. We’ve talked on the pod in the past about the browser wars, about browsers—web browsers—as distribution channels for AI.

I think it’s quite notable, but sort of under the category of burying the lead, that as a consumer, you can only access Grok Code Fast 1 via one of several different coding environments, like Cursor or Windsurf. To the extent that we talk about distribution channels for superintelligence, I think it’s quite notable that those coding environments are becoming almost competitors for the browser for accessing superintelligence.

Fascinating. Yeah, entry points. I’m looking at Cursor right now on my screen. I don’t see it there. Is there something I need to do to get it?

With Windsurf, it’s a bit of a hassle to get to. You have to search specifically for Grok Code Fast 1 in order to get it.

It’s not even one of the recommended ones.

No, it’s here. Sorry, it’s way down at the bottom. I don’t know why.

Well, it’s just out, and it’ll gain popularity. But I think the point here for everybody is that these incredible tools are demonetizing and democratizing at an extraordinary rate, right? We’re going to see literally billions of coders. Everybody will learn to code. The language of coding is going to be basically your mind—your English or your Hindi or whatever it might be.

All right. Before the next podcast, I’ll test it and see how it fares just on quality. Obviously, on price it’s incredible, but I want to see how it fares on quality. I don’t understand how this price point is possible, just counting up the FLOPs and working it back to the chip costs. Alex, I don’t know if you have any insights on that.

There was an article that said a lot of this is being basically carried by Silicon Valley investment, at least.

Yeah, but you can’t do a little bit of that and then burn money by the billion. I mean, you can for a little while, but you can’t do it sustainably. Do you think that’s what’s going on here, Alex? I don’t know.

It’s not clear. I see different accounting schemes. Without having direct access to the chart of accounts for the frontier labs, it’s difficult to know whether inference time, as we were discussing, is profitable or not. I’ve seen claims either way.

But I do think, more broadly, Jevons’ paradox, broadly speaking, is that the cheaper a given commodity gets, there’s a sort of paradoxical net increase in demand, resulting in potentially greater expenditures in this case. I think we’re going to see that with code generation as well. As the cost of intelligent code generation trends toward zero, as it has been for the past few years, I think we will see, to Peter’s point, just-in-time code demand for everything, and we’ll be awash in new code that otherwise never would have existed.

One thing that came up when we were talking to Kevin Weil two weeks ago at OpenAI—the chief product officer at OpenAI—is that there’s a huge amount of routing optimization going on. When I use these things, I’ll bounce back and forth between a trivial question and then, “Can you solve cold fusion?” back to back. The models are very intelligent now about routing each question to the minimal model that will actually answer it correctly.

There are many different layers of those types of optimizations. So, I think you’re going to see innovation at this ridiculous pace, and then the price point continues to come down. But again, the demand is infinite.

Let us know when you solve zero-point energy. I’d like one.

Okay. This is another fun article: Elon poaches 14 Meta engineers with a different offer. We’ll talk about what’s going on in Meta in a second, but I think this is fundamentally what Elon does extraordinarily well, right? He has a massive transformative purpose: open up Mars for humanity, make humanity a multiplanetary species. He has that as a pure signal.

Entrepreneurs who are willing to work hard and who are builders want that. They want to work on something epic. It isn’t about the money, since eventually we’re heading toward a post-capitalist society anyway. So, Elon’s offering purpose and equity over cash, and his equity has done incredibly well. I don’t think he’s ever started a company that’s lost money. It’s increased. All the companies, from Starlink to xAI to X, have basically just skyrocketed in value.

We’re going to see a start-up intensity here. Alex, what are your thoughts on this?

I remember back to the early days of, call it, the post-2012 ImageNet revolution in AI, when there was a lot of concern, including, as I recall, from Elon, that AI would end up being a monoculture. One lab, maybe called Google DeepMind, would completely capture the future light cone with AI.

I view this and other related headlines as a very helpful sign that we’re not ending up in that future of a monoculture, of a singleton, where just a single culture from a single frontier lab has complete dominance and complete hegemony over AI. We’re seeing multiple competing AI frontier labs with different cultures.

This should be one of those cultures: the manic focus and intensity of just delivering state-of-the-art results. We’ll see lots of other cultures as well, and we’ll have them compete. That’s the world that we humans, I would argue, want to live in.

Yeah. I think purpose—a purpose-driven life—is going to be far more important in the future than anything else. Being clear about why you’re doing what you do, waking up in the morning with an epic mission, and being excited about building is a future that I want. It’s always been part of my life, and I want that for my kids.

That’s what’s going to win over just cash, especially if you’re working someplace where you don’t like the culture. As you said, I want to play a short video. This comes as part of Elon’s Master Plan Part 4 on sustainable abundance, and let’s chat about it afterward.

Cool.

I love that term, “sustainable abundance.” It’s amazing to see abundance becoming an underlying theme for the hyperscalers and for the tech world. It’s part of the optimistic vision of the future, right? You’re not going to see this on the 6:00 or 7:00 news. You’re not going to see it in The Washington Post or The New York Times.

It’s important to realize that the technologies we’re talking about on this pod are going to shape every aspect of our lives, and there is so much positive news. It’s so easy to focus on the negative. But understanding this is important, and Elon’s been on this mission since I met him back in 2000, when he had sold PayPal to eBay.

He’s been on the mission of autonomous vehicles, electric cars, solar, making humanity multiplanetary, and people love those epic grand challenges, right? These moonshots, which this podcast is all about, are what they gravitate toward.

So, Peter, you invented XPRIZE and used X about 10 years before Elon stole it from you, and then Abundance was the name of one of your books.

My first book. Yes.

So, what’s next in the Elon-takes-all-of-Peter’s-ideas category?

Oh, God, it was very funny because the XPRIZE logo looked identical to SpaceX’s and then to X’s logo.

I didn't have the heart to call him out on it, but hey, it's fine. He's been generous supporting XPRIZE and supporting our work over the years. He's given probably $150 million in capital to support some of the XPRIZEs we've done.

But sustainable abundance is a real thing, right? It's about digitizing, dematerializing, demonetizing, and democratizing everything. The way I like to describe this is: when Google came out, Google was a for-profit company that had the biggest nonprofit impact. In other words, it uplifted all of humanity. The poorest child on the planet using Google and the wealthiest child—Larry Page's kids, Sergey Brin's kids—Google was identical for all of them. It was a leveling and democratizing capability.

We're going to see that here for food, water, energy, healthcare, and education. That's extraordinary. Alex, you've been on this journey with me.

Oh, yes. Yeah, no, I love this term. I think it's sort of an implicit recognition that another important part of the equation—autonomy, a.k.a. superintelligence or post-scarcity intelligence—is the missing factor here.

When I look at the video and hear the term and see its usage in practice, immediately this screams to me the intersection of abundant energy and materials on the one hand, and abundant intelligence and autonomy on the other hand. I think it's a rare and precious time in human history when we can have near-total clarity as to what the technology tree looks like.

Every time we talk about having scarcity of something like lithium—“Oh my God, we're running out of lithium”—it's like we discover these massive supplies. Off the coast of California, there is nothing that's truly scarce, period. I think the sooner people get that technology is a scarcity-destroying force, the better. Dave, you were going to say something.

Yeah. The other thing on that slide that's kind of sandwiched in the middle in bullet 2 is the 7-day workweek. We have a bunch of companies at Link Studio working 9-9-7, and they just declare, “We're doing 9-9-7.” What's 9-9-7? It's this China thing where you work 9:00 a.m. to 9:00 p.m., 7 days a week, while you're sprinting toward some massively transformative purpose.

Those two things go hand in hand. No one's going to work 9-9-7 on something irrelevant, right? It has to be something—

—world-changing and imminent. You know, 9-9-7 is no way to live your life. You do it in a sprint to get to a very specific destination, but that is by far the winning strategy. If you do this massive million-GPU data center and you do it a year late, it's worth zero.

It's absolutely worthless. There's no point in doing it on a 4-day workweek. You either do it or don't do it. If you do it, you can sprint. It's like an Olympic gold medal—you've got to sprint.

Yeah. I think it's also interesting. I think it's 996. China was 996, but I think Dave is saying, with the reaction to 996—popularly known as lying flat—people are opting out of the 996 culture.

It's interesting to think about whether sustainable abundance actually obviates that entire discussion altogether, rather than lying flat under the implicit assumption that 996 and the need for enormous amounts of human labor are going to continue in perpetuity. What if we actually, a few years from now, find ourselves in a sustainably abundant future where the need for 996 human labor is actually only a short-term need? A few years from now, we hand over that workload to autonomous systems.

Then we go to the stars. Then we go to the stars. All right. On the flip side of this conversation, people are beginning to bolt from Meta's new superintelligence lab.

Two months after its launch, at least 3 top researchers have resigned to return to OpenAI, while Meta's long-term product director also joined Altman's ranks. The departures are raising questions, despite recruits being offered 9-figure pay packages. To be clear, we don't actually know what's going on inside. This is just reporting what came out in Wired.

But I do think that when you're capturing an employee by offering them a lot of money, that's not going to capture their time and attention and their heart. It's mission and purpose that captures them. Dave, you agree?

Well, absolutely. I think also Elon has a reputation for everything always working. Having an MTP has to be married with a mission that will succeed to develop that reputation—not only is it massive in implications, but it's going to happen. It's real.

It's very easy to go out there and say, “Hey, I'm going to build electric cars.” Like, come on. Then, if it doesn't work out, no one will join you the second time. You have to have a track record of succeeding.

Mark Zuckerberg is probably the most successful of the young CEOs in the country, with the most cash flow and every opportunity to win. Elon has done it repeatedly, though. I think a lot of people are flocking to him: he's always been right before. Why would he be wrong this time?

Yeah. That really helps.

Never doubt Elon. Right. Yeah.

Yeah.

All right. I don't want to spend too much time on this, but it's been a big week for Google. We're on the verge of the release of Gemini 3, which will be coming out any day now. The top news, though, is the extraordinary conversation around Nano Banana, powered by Gemini 2.5 Flash Image. Let's take a look at this video.

A few days ago, image editing changed forever. Google released Gemini 2.5 Flash Image, Nano Banana, that has everyone buying puts on Adobe because Photoshop is officially dead.

Instead of learning how to use all these antique tools, you can now just prompt Nano Banana for changes, and it's able to deliver any photo alterations you can imagine—most importantly, while maintaining the consistency of the original image.

Not only is Nano Banana an exceptional image model that's already at the top of the LMArena leaderboard, but it's also extremely fast and affordable, costing only 3.9 cents per image via the API.

The upgrade that most people are talking about, though, is character consistency. If you start with an image of a person or pet, for example, the model can blend it with a different image or make minor changes to it without noticeably altering the original character. Or multiple characters and objects, like this guy did by blending 13 different images together.

What's kind of crazy about this model, though, is that it also has an understanding of the real world. If you point to a spot on Google Maps and ask what a person would see there, it can generate a realistic photo.

Just epic.

Just epic is right. Very few people—basically nobody under the age of 40—remember life before the GUI, the graphical user interface, on your computer. But we do.

When the Apple II first came out, you would boot it up—or a TRS-80, or whatever—and this little flashing prompt would be there. What can I do with this prompt?

Yeah. Hello, what do I do? All you can do is just start writing code. It's basically all you've got. Just start writing code.

Then, in 1984 or '85, Steve Jobs comes out with the Mac, and now everybody's lived in this kind of world of stasis of the GUI for 30 or 40 years. Everybody's like, “Yeah, nothing ever changes.”

This is all going to change imminently. It'll be the biggest step function. It'll be much bigger than going from no computer to computer, or from command line to GUI. It's been so long that nothing has changed that people are completely underestimating how different the world will be a year from today, when everything has, “I just asked the computer to do this for me,” just like on Star Trek—

And it just did it. But that's happening literally right now.

The implications for startups are incredible. If Adobe gets destroyed by this—

I think our friend Greg Bellis is still over there, so it'd be kind of sad if that happens. But it'll be very important as a wake-up call that if you've been camping on your software installed base for the last 20-plus years, milking it for money, your days are numbered.

Yeah, I mean, because everything's going to change. If you were an Adobe Photoshop or Canva specialist making your living that way, you understood how to do layers and masks and manual adjustments. But Nano Banana is just editing through language, not layers. It's literally, “Do this.” How do you describe what you want in a way that AI will understand it? That's going to be the skill base.

Yeah, exactly. If you're a graphic artist or a writer or whatever, you get so used to these tools and all their proprietary interface components. Then you're afraid to shift to something else because you get so invested in knowing where the menus are, knowing where the buttons are, and knowing how it responds.

Then you're locked in, and you end up paying for that product for 10 or 20 years. Now everything's wide open again. The interface is trivial. My mom, who's in her 80s, has no problem creating images. She could never use Adobe Photoshop—never figured it out. Now she can just talk to it.

I would argue, actually, that I've been using Nano Banana quite a bit. It's a much bigger deal than just some of the headlines that would say this is a Photoshop killer.

So, in using Nano Banana, some of the most striking new capabilities that I've seen are that you can feed it an image and then ask to view the same scene from a different perspective. That's way more than just pixel-level, Photoshop-style editing. It smells to me like this is just a sliver or a distillation of a larger world model, sort of like we've spoken about Genie 3 in the past.

It feels to me like this is some sort of tendril from a much more monstrous model. If that is indeed the case, and to the extent that Nano Banana has basically become merged into mainline Gemini model releases by Google DeepMind, I think this portends a future where world models, like those in the class of Genie 3—the videos that we've discussed previously—just merge into GPT- or Gemini-type models as the ultimate modality of interactive, simulated, maybe even streaming realities.

I tell you what else, Alex: all media competes with all other media. There are no swim lanes. Everything competes for users' time.

Time is scarce.

At least while we're stuck with finite attention, maybe we can make attention post-scarce as well.

Okay, well, for the next couple of years. But this time of year, normally by now, I would have done a fantasy football league, signed up for my players, and known when the first football games would start. I haven't even paid attention. I don't even know what's going on because the stuff that you post in our Link chat is so much more entertaining and engaging than mainstream media.

The stuff that you, as a single-handed person, can create is so much more interesting and relevant. It's a capability that never would have existed a year ago.

Well, look, a singularity probably only comes about approximately once per planet, so it's a special time.

It is a special time. We're going to celebrate that. I'm going to have a singularity party when it happens. I hope you guys will join me.

We made this point last time: AI won't take your job. It'll let you do any job. I think this is a perfect example. Literally, designers who have made their careers based on understanding how to use a specific tool really well—now anybody can do that.

But here's a question for you: If I go and say to Nano Banana, "Hey, place me on the Moon in a spacesuit getting into a Starship for a return flight to Earth," and it generates that, who's the author of it? Is it the software? Is it the human who prompted it?

We're going to start to have some interesting conversations around ownership, blurring the lines of authorship and human-in-the-loop creativity. That's going to be an important conversation. But let's talk about the real issue here, which is the ability for this to drive accelerated misinformation and the erosion of visual trust. The old saying, "Seeing is believing," is out the window—period. Thoughts?

Yeah, my comment on that would be that we're in a post-Veo 3, post-Sora, post-natural-language-generation era. To the extent that one wants to have faith in the accuracy of any visual inputs—images or videos—I can see a future where there's some sort of cryptographic chain of trust between cameras, video and still cameras, and browsers.

It's sort of like the way there's a cryptographic guarantee that when you put in your credit card information to pay for something on a website, that credit card information is handled in a cryptographically safe way between you and the ultimate counterparty. One can imagine some sort of cryptographic guarantee that the image that you see in social media was actually unaltered, in some sense, from the original capture without any AI involved.

That said, my baseline expectation is that that's not going to be very popular. Blockchain to the rescue. But this is deepfakes on an industrial scale.

Right. I mean, just to put it where it is, this is deepfakes on an industrial scale.

Now, this is what people also said right before GPT-2 and GPT-3: This will empower all sorts of misinformation and disinformation. Yes, there's probably a lot of false information being generated by these models, but I have to look at it as a risk-reward trade-off. There's so much new scientific information being unlocked by these models that it's very difficult to get too bothered by the potential downsides.

I'm not worried about that. I'm just saying, for the majority of the 8 billion people on the planet, if they keep on seeing this—when I see a video, my first reaction is, "Is it real?" And my second reaction is going to be, "No, it's not real."

Yeah, no doubt. You've read all the Neal Stephenson books, I'm sure, Peter.

Of course.

Of course.

It's not an "of course," but of course. The Diamond Age, to me, was everything he's ever predicted in those books has happened. He invented the word avatar in his first book, Snow Crash, and foresaw cyberspace; that word was invented there.

In The Diamond Age, everybody moves into communities that have different rules around technology and how you manifest it because it gets too weird too fast. The big step-function change for society is cameras everywhere, right? That started years ago, and now we're living in a camera-everywhere world. There's no concept of privacy. Anytime you're outside, you're being filmed, at a minimum by a satellite.

You can know anything you want, anytime you want, anywhere you want. The data is there: layers of drones, layers of satellites, and layers of autonomous-car cameras everywhere. Everything's being imaged. There is no privacy. That's another conversation we can have sometime.

Well, those 10-megapixel cameras now are 50 cents each.

Yeah.

So they're going to be everywhere. That already happened to society, and social media happened to society. It totally changed the whole election process, and everything's been disrupted tremendously in the last 10 years.

Now you layer the deepfakes on top of that. It's just the third act in this massive, turbulent societal change that's only going to accelerate. And, of course, governments do nothing. They just sit there and assume that people are talking about the next election as if it's going to be anything like the last election. It's going to be a different world by the time we get to the next election.

Yeah, I would just maybe add that I would speculate this moral panic that we're engaging in right now is going to look very quaint in a few years. It's difficult to imagine people with smart glasses doing real-time augmented-reality overlays of everything that they're seeing. That basically means everything becomes photo-edited. Everything that you see becomes doctored by default.

On the other hand, you can clutch whatever it is—the moral panic that you're worrying about. I think it's far more likely that this will look quaint and nonsensical in a few years. Whenever panic becomes quaint, it's worth noting that there's a digital invisible watermark that Google is putting on these images: SynthID.

We're going to start to have an arms race between generative AI and detection tools as well. That's going to be part of it. It's always the virus-antivirus war.

Hey everybody, there's not a week that goes by when I don't get the strangest of compliments. Someone will stop me and say, “Peter, you've got such nice skin.” Honestly, I never thought, especially at age 64, I'd be hearing anyone say that I have great skin. And honestly, I can't take any credit. I use an amazing product called One Skin OS01 twice a day, every day. The company was built by four brilliant PhD women who have identified a 10 amino acid peptide that effectively reverses the age of your skin. I love it and, like I say, I use it every day, twice a day. There you have it. That's my secret. You go to onskin.co and write peter at checkout for a discount on the same product I use. Okay, now back to the episode. Let's go to Google's next big announcement of this past week: Google Translate, another incredible product coming out right now. Google's AI-powered live translation—historically, Google has translated about a trillion words per month for 600 million users, supporting 243 languages, which, by the way, is 58,806 language pairs. Amazing.

But now we're being driven by Gemini 2.5 with live translate. Let's take a look at this video.

Hi there. My friend told me there's a sandwich here that's really good, but I'm not sure which one it is. It's spicy, has really tasty cheese on it, and avocados, I think.

I think I know what it is. It's seasonal. In fact, we've already taken it off the menu, but let me see if we can still prepare it for you.

Pretty amazing. The question is, what's this going to do to the language-translation industry? What's it going to do to people learning languages? I used to want my kids to learn multiple languages. Now the question is, do they invest their time in doing that?

What was that company that all the kids used to cheat on their homework with? It was a public company and got obliterated.

Alex, which one is it?

Not sure. But, as a historic reminder, remember the Transformer architecture that helped kick-start a lot of the generative AI revolution.

It was originally developed for language translation, for machine translation. It was an encoder-plus-decoder architecture. Right now, we mostly use the decoder part. Nonetheless, it's sort of ironic that the original targeted application for transformers was statistical machine translation, or machine translation.

It's only now that we're starting to see pervasive machine translation finally tackling the real-world use cases of real-time conversational embeddings. That's my first thought. My second thought is, it's interesting to speculate: What does this do to language diversity in general? Does this promote languages?

There's been a lot of hand-wringing over the past 20 years about low-resource languages dying out in favor of usually English, but sometimes other languages. Or is this a net promoter of diversity, where, once all languages, thanks to AI, become fully interoperable, as it were, there's suddenly no reason to collapse down to one monolingual English language? Do you know the joke here? What do you call someone who speaks 3 languages? They're trilingual. Someone who speaks 2 languages? Bilingual. If they speak 1 language, they're American.

I love that.

Well, that's going to turn out to be the winning strategy. What do you know? The company was Chegg. Check out its stock ticker, or we'll splice it into the podcast here, but it went down from $90 to $1.40.

Wow.

Just because ChatGPT is a better way to cheat on your homework or whatever—that's not really what they do, but it's that—

But the point here is that Google is also providing a language-practice mode that allows you to personalize speaking and listening exercises, right? And so, the impact on Duolingo—

That was a 10% stock drop. Duolingo is a $13 billion company. It's done incredibly well, with 130 million active users. Only 10% of users pay, but nonetheless, it's generating real revenues. You can see this drop that occurred on August 29, when Google's live AI translation capability was announced.

This is another example where these large frontier models, sort of in their wake, whether or not they know it, are going to be disrupting existing companies, which are going to have to constantly be pivoting. Yeah, Netflix. Duolingo is absolutely doomed unless it becomes an AI company. And if it becomes an AI company, it can go through the roof. But a lot of these companies don't have the AI talent to get started. So you have to turn the battleship somehow, but if you do succeed in turning the battleship, your valuation can go through the roof. You've seen that a single incredibly talented AI researcher can be worth $1 billion.

So, the value is there. In general, in the short term, I want to generalize from just this one instance. In some sense, I think the cliché here is every software-as-a-service company is under existential threat from generative AI models that will simply cannibalize them from below.

Whether you're doing software for some enterprise purpose or you're just offering software subscriptions to help people learn a new language, that frontier model, that chatbot, is going to devour you because you've become just one special case among countless cases that a generalist model can handle. I think that's critically important, right?

Every CEO out there, every board of directors, needs to understand that if they're not building on an AI base that's accelerating alongside everybody else—if they're depending on their old business model, software as a service—they will be marginalized. But here's—

Remember what we said before, too: If you're in a regulated industry, you have a little bit of time. You can actually get ahead of it. You have to get the AI talent now, but you can get ahead of it. If you're not in a regulated industry, like Chegg or Duolingo, and you're just a user-installed base, then you're really vulnerable.

I mean, you have an advantage of a user-installed base and a brand. Use that to your advantage to actually leapfrog forward. Don't—

Like, idle.

Yeah.

Yeah. I just want to dwell for a minute on what that leapfrogging looks like. I think I agree that, in the short to medium term, differentiated user experiences are a bit of a moat, if you will. But in the medium to long term, what I'd like to see from every single SaaS that feels existential risk from being devoured by a generalist model is: step up your ambition by 100x, 1,000x.

If you're Duolingo and you happen to feel existentially threatened by generalist models, maybe consider becoming a brain-computer interface company. Wouldn't it be wonderful if we could sideload new languages, in the style of the movie The Matrix, into the human brain? Rather than spending days, weeks, or years learning a new language, why can't you enable your clients or users to learn it in a minute?

Yes. Pick a moonshot. I mean, that's the whole purpose of this podcast: Get people to go 10x, 100x bigger. Pick your moonshot, or—

Hire Alex as a consultant for 2 weeks, and you'll have a moonshot at the end of that.

No, he has no time for that.

Okay. All right. So let's take a look here. We've got OpenAI's real-time API bringing smarter voice AI. Let's look at this quick video here. I love this one.

Are there any homes in my budget near water, with a view of the skyline and Mount Rainier?

Sure. Let me look. With your buying ability of $824,000, Wallingford would be a great fit. I think you'll love 404 North 33rd Street. It has those skyline and Rainier views you're after. With this week's market, I'd book a tour with an agent soon. Want me to set that up?

This capability is an example of using this on Zillow to find your home, describing exactly what you want, and having it actually scrape and generate an efficient answer. But just the ability to do all of this and actually get you to the point—what you want next is: Find the house, buy it for me, arrange the mortgage, arrange the moving trucks, and let me know when to show up in my new place.

Yeah.

Yeah. No, I think AI as a management tool—a general-purpose management tool—is hugely underrated because everybody loves the graphical stuff: the image creation, the self-driving car, the stuff you can feel. But just as a general way to manage large-scale projects with hundreds of people, moving parts, and logistics, it's unbelievably good at doing that.

So I think we can expect far more efficient construction, management, manufacturing, and supply chains than we've ever seen before, because the sensor data with all the cameras everywhere has been available for a few years now. But it's all kind of dumped into big databases. You throw it into Snowflake or something like that, and then it's very hard to make sense of it.

The missing ingredient was this AI overlay that can just take the unstructured, free-form data and turn it into conclusions, actions, schedules, buying things, scheduling things, and managing things. We had our condo in Vermont built many years ago—20 years ago—and they put the Tyvek on upside down, so it's overlapping the wrong way. It grabs rainwater and funnels it into the wood. So, years later, everything's rotting; the whole thing's falling apart.

Why would you put the Tyvek on upside down? Now, it's very easy for the AI to say, “Hey, dude, stop. It's just as easy to put it on right. You're putting it on overlapping the wrong way.” Just a trivially easy AI problem, all of a sudden. Thousands of things like that can suddenly be converted—

Sorry, Peter.

I was going to say the point here, though, is the real-time API. Alex, let's chat about that.

Yeah, so I've played with this. The underlying model is called GPT Realtime. If you've played with AVM—the advanced voice mode of OpenAI—that's the mode where you can chat in real time with very low-latency responses with ChatGPT. It's a lot like that, but in API form, so that it can serve as a backend for third-party applications.

I really do think this is transformative, in part because imagine taking low-latency voice-to-voice, but generally capable, intelligence and now embedding it everywhere. I think it probably ends up being transformative for customer-service-type applications, probably many other sectors as well.

But even bigger picture, I think this is a preview, albeit a tiny preview, of a future where every single audio segment, every single pixel on screen, is generated in real time, streamed interactively on demand. Our user experiences, our user interfaces, are just completely, just-in-time generated. It's going to be a very, very interesting future.

And there's a single interface. There's a single interface to the world, right? Your Jarvis will go and interface with everything out there, whether you know it exists or not, and give you the answer you finally want. We don't have a good catchphrase for this one.

Our voice customer-service company, Vocara, doubled in ARR during the past week and is planning to 10x between here and the end of the year, just using this exact capability for complex customer-service and sales conversations. So far, consumers dramatically prefer it to a human call-center agent.

It's so knowledgeable in the link studio.

Mhm.

Yeah. Yeah. IT team.

To my knowledge, we're missing a term for this. I've definitely come around to the view that it's important to coin new terms whenever there's this important new concept. I think we're missing a term for this. It's not conversational user interface because it isn't always conversational.

The best term, if I had to coin a term for what I think we're seeing the beginnings of, would be something like streaming interactive models. It's not necessarily just voice. It could be like Genie 3, where, if there's a visual component, it could ultimately be like a brain-computer-interface-type component. So try it on for size: streaming interactive models, or SIMs.

And because everything becomes a TLA—streaming interactive models, or SIMs, right? Okay, Alex, you...

The Zillow example is really important for people to look at. Rewind the pod and watch it again, because customer service is typically a phone call today. It's very hard to explain complicated things on a phone call.

So this is very quickly going to move to multimodal, where it's talking to you while creating images in real time. People haven't experienced that before, because no human call center operator can create an image or pull up thousands of pictures in real time, but the AI can do it very easily.

Whole new experience. Let's move on if we can. A lot to cover still. We're still in AI. We're going to be covering a lot more in energy, health, and Starship.

NVIDIA beats revenue predictions, defying fears of an AI bubble. I think that's great news. Up 56% year over year from 2024. The company's at $4 trillion. The stock is up 700% since ChatGPT's 2022 release. How awesome is that?

Yet we still have a bunch of U.S.-China turbulence. NVIDIA gave 15% of China sales to the U.S. to keep exporting. Just reporting this news, NVIDIA continues to be leading the pack. There's another piece of news I want to hit on regarding this, which is out today. Alex, I'd like you to chat about it: investors bet on Cambricon as China's next AI chip champion. Would you chat about this?

Yeah, maybe looking at these 2 stories together through the lens of where value is accumulating in the stack. I think there are 2 competing worldviews. One is, call it, the pyramid model, where the broadest part of the pyramid is at the base. In this case, under this worldview, most of the profits in the AI revolution that we're living through will accumulate at the lower infrastructure levels, like the chip designers, the fabs, or data centers—the lower levels.

There's also a competing worldview that we ultimately move to, or maybe are living in but just don't realize it yet: an inverted pyramid model, where most of the profits and most of the value accrue at the upper layers, the application layers—the startups that are being built on top of these frontier models—and the frontier models themselves just become profitless or profit-sucking commodities.

I think if you look at these 2 stories through a common lens, at the moment these would seem to bias me, at least, in the direction of thinking that, for the moment, most of the profit is accumulating at the bottom of the stack, at the chip design level and at the data center level, regardless of geography.

But let's talk a little bit about this new company, about Cambricon, if you would.

Yeah. No, it's difficult to know what precisely is going on inside any given company, regardless of whether it's U.S.-based or China-based. I do think, again, just generalizing over Huawei, Cambricon, and then obviously a whole cohort of American AI chip designers, I think we're seeing the beginnings of a non-monoculture, where there are diverse chip architectures—a diversity of chip architectures for AI acceleration from the U.S., and seemingly the beginnings of a diverse set of non-NVIDIA-based AI accelerator or accelerated-compute architectures coming out of China.

Where all of this goes, I think, Peter, your bet is as good, if not better, than mine, but I think that the headline here is that there may be the beginnings of a post-NVIDIA, post-CUDA monoculture kind of—

That's the point I want to make. Whenever you restrict China's ability to sell them products, they will develop products there, right? We have a lead, but that lead is getting shorter and shorter on chips and AI. We saw this as well in the satellite world, right? When the U.S. State Department started limiting the ability to export satellites from the U.S. to different parts of the world, the industry finally materialized and competed back against the United States.

And so this strategy of scarcity doesn't work in a global culture of innovation. I'll also tell you, you know, David Sacks is talking to you right now, but if you look at that 7-nanometer capability and we're operating at 2 nanometers, you're like, “Oh, we're miles ahead of China.”

But the algorithmic improvements can be massive, like 1,000,000×-kind of improvements.

Yes, they are way more important than the 7-versus-2-nanometer gap. We're not used to that in government, because we're used to the nuclear arms race or the space race, where you're not going to get a 10× advantage by magic. There's no rocket fuel that you can throw in there that's one-tenth the weight of the competing fuel. It just doesn't exist. But in software, it does exist. In fact, it's common. It's everywhere.

And so, it's very easy to get complacent where you stand.

You, in fact, when you restrict, you cause innovation in different areas here. And you're right, algorithmically, we're going to see 100×, 200×, and 1,000× improvements there over the next few years.

And just to remember, China has won the Math Olympiad year after year after year. They have incredible talent, and 50% of Meta's AI staff is Chinese. Let's not fool ourselves. The intelligence is there to innovate as well as here.

Though it's different here, of course. We mentioned it earlier: it's the risk capital, the entrepreneurial drive that has people working around the clock. I would just make this note again as we talk about NVIDIA: this is not going to be solely an NVIDIA world. We're going to see China step up and compete.

I love this article. Again, we've talked about the idea that AI is no longer U.S.-centric. We're seeing the world step up and get involved. This is “Billionaire Mukesh Ambani Taps Google and Meta to Build India's AI Backbone.” Mukesh Ambani launches Reliance Intelligence to build India's AI infrastructure.

I know Mukesh. I've been to his home a number of times in India. I was at his epic wedding a year and a half ago or so, and the guy is an incredible entrepreneur. Just for people to understand this: he enters India's telecom market in 2016. He's the 10th mobile provider, right? You've got Vodafone, Airtel, and all the players there. But he comes in with a completely different business model, and that's his brilliance.

Reliance Jio launched a radically different model: free voice calls for life, ultra-cheap data, and months of free service trials. The other thing he did was, it used to take you 2 days to get a mobile phone. He basically said, “Show up in the store, sign a few papers, and it's instantly up and operating.” Then he used his capital to leapfrog over 2G and 3G and build out a nationwide 4G network.

So he literally destroyed the competition, and they are the major mobile phone provider. Cell service in India right now, when I was there, was 5-bar service, 5G across the nation everywhere. It's pretty extraordinary, and I expect he's going to do the same thing here in AI.

Well, something big is going to happen in India because you saw Kevin Weil saying they're making a huge push at OpenAI into India. You're like, “Oh, that's kind of odd. Why are you doing that?” Well, if you look at the demographics of the country, it has by far the most untapped talent in the world. I mean, by far.

Yeah. China's in a terrible spot because of the aging demographic problem. The one-child-per-family thing caught up to them in a big way, and now they've got a massive aging demographic problem. The U.S. is in great shape because immigration is strong. It always has been, and hopefully always will be.

But India has the best latent talent pool in the world—right in the sweet spot, the 20-to-40-year-old talent pool. The reason per-capita GDP has been so bad in India for so long is that it's incredibly corrupt; all the structures are terrible. But I think AI might have a way to cut through that and just go directly to the people.

And also poor transportation, with the roads being flooded out. We're going to aerial-deliver in India as well.

All right, let's keep moving on. TIME100 AI 2025. This was their issue. I sent Marc Benioff, “Congratulations.” I said, “Marc, you're not listed here, but you need to be on this list as well.”

But check this out: if these are in order, Matthew Prince is number 1, Elon is number 2, and Sam Altman is number 3. It's fascinating that Matthew Prince is number 1. Any idea why?

I would be remiss if I didn't note that sometime in the media cycle over the past week, there's been a lot of interest in the future of Cloudflare, which Matt leads, and agentic AI. There's a lot of interest in what a web where AI agents are independently surfing the web on the same level, with the same rights, as human web surfers looks like—or whether there should be a separate entrance to the web and to the economy for AI agents.

So, if I had to speculate, I would say the intersection of Cloudflare and special handling of AI agents could be one possible reason.

I did a little digging. Let me tell you what I found out. Matthew Prince stands at the top of this list for one reason: he’s been focused on safeguarding the value of internet content. He’s all about making sure there’s proper attribution and that you’re basically not stealing from the publishers. And, of course, TIME magazine is a publisher.

I think they’re flexing their muscle here to say attribution is critically important. I think it’s going to be maybe even worthy of much more dedicated time, actually doing a deep dive on the issue: Should AI surfing the web on its own be treated the same as a human web surfer, or should it be treated differently? I think there are so many nuances there.

Yeah.

The other thing we see on this list is a huge amount of global diversity, and it picks up leaders in different countries. This is no longer just a Silicon Valley play. This is a global play, where countries are beginning to invest heavily and really double down on this.

Next topic here—I love this one. Airbnb’s co-founder, Joe Gebbia, who’s been on my stage at Abundance—he’s amazing—has been named the U.S. chief design officer, appointed by Trump. His goal is to redesign government sites and services to be simple, modern, and friendly. I love his quote: “I want to make government services as satisfying to use as the Apple Store.” That would be awesome.

I think it’s perhaps not obvious, but there is actually an open-source library hosted on GitHub. I think it offers Joe enormous amounts of leverage for the task that he’s taking on. It’s called the U.S. Web Design System, USWDS. In principle, it’s a common set of user-interface components underlying most—not all, perhaps, but most—U.S. government websites. That’s sort of a seminal place, I think, for Joe to start, with such high leverage, if the goal is to radically improve the user experience directly, at least.

The key point of this story here is the Trump administration tapping entrepreneurs to come in and help move the government forward. Despite whether you’re a Trump lover or hater, it doesn’t matter. This is about bringing in the smartest people, because historically, going to work for the government was not where an intelligent entrepreneur would go. There’s been an incredible shift in that regard.

Yeah. No, you phrased it exactly right. I think when the U.S. government said, “We’re going to have a chief technology officer,” back under Obama originally, the first 2 CTOs of the United States had law degrees, and they were just buddies of the president. Then we created that HealthCare.gov site. It was $1 billion to build a website, and then it never launched. It failed. So, okay, why don’t we get some real technologists into D.C.?

I can’t believe it’s actually happening, though. It’s amazing.

Well, these people are post-capital, post-abundance themselves, right? They’ve made their money. They could be working on their next moonshot, or they could be building a moonshot that will hopefully right the ocean liner of the United States.

I think during COVID, a lot of people who normally didn’t care about government suddenly started caring a lot. They realized how much government can change your day-to-day life. Forcing you to stay inside—that’s pretty extreme in terms of government intervention in day-to-day life. So, whether it was right or wrong, they felt like, “Wow, this really matters. I need to get involved.”

Yeah. Well, good luck to Joe. I’m sure this will have a huge impact. Making something actually usable—this is like when we had ARPANET, usable by a few individuals at MIT, Harvard, Stanford, and the defense industry, and then Marc Andreessen comes and builds a layer on top of that with Mosaic. If Joe can do that—make it easy to use and functional—that would be amazing.

All right, here’s our debate and discussion for today. I’m going to read this out, and I want to hear your thoughts here. Here it is: Will people vote AI into power? This is a tweet from V Razer X:

“Because they’re tired of corruption and broken promises, AI will provide laws without loopholes and policies based on measurable outcomes. Election by election, trust will shift. Eventually, the ballot will include a new option: governance by AI. Citizens will choose it, not out of fear, but hope for fairness. Power won’t be inherited or bought. It will be optimized and accountable. Democracy’s paradox: people will freely vote to be governed by something beyond human flaws.”

Here’s the question: Do you believe this? Do you believe that we will be voting AI into power? Dave, what’s your position here?

Well, I think there’s a long history of laws having people’s names on them, like, you know, Obamacare or Glass Deagle Act or Graham Dodd or, you know, PE The fact that AI is coming up with the idea and writing the law won’t change the fact that someone will put their name on it and say, “This is my act.” But it’ll still be AI creating the law under the covers.

I think it’s inevitable. It’s going to happen very, very quickly because the number of things that need some kind of framework is explosively growing, exponentially growing. So the traditional process of passing it through Congress, passing it through your local legislature, is way too slow to keep up with the rate of change.

This is definitely going to happen, but not quite the way you’re suggesting. You’re not going to vote in AI to be your politician. It’ll still look and sound like a person.

So, let me be clear. I wish this would happen. I’d love to see this happen. I don’t think there’s any way in the world, short of a revolution or starting a new country off-world, that we’re going to see this happen.

There are lots of reasons. For me, the most important thing is the entrenched bureaucracies, right? Politicians, bureaucracies, and entrenched interests will fiercely resist bringing this on. Courts will strike this down. And talk about getting rid of corruption: corruption doesn’t vanish; it just shifts, right? Corruption will shift from the politicians to the engineers, corporations, or states that are manipulating the AI.

As much as I’d love to see this happen, I don’t think it will. Alex, how about you?

I’ll take a third position in this debate. The question is nonsensical. This is a very old trope in fiction.

Just 2 examples. If you remember the original version of The Day the Earth Stood Still, based on the sci-fi novella Farewell to the Master, the entire premise was that alien civilizations had decided that they themselves—the biologicals—couldn’t be trusted to maintain peace. So they ceded all authority to a race of robots that policed them.

One can look back even further. Remember, famously, Henry V 6th: “Let’s kill all the lawyers.” This is a very old trope in fiction. I think it’s completely nonsensical. What I expect to happen is humans will merge with the AIs, and so the question then degenerates to: Will people vote people to power? The answer is yes, but it’s sort of vacuous in my mind to ask whether people will, separately from that, vote AI into power.

We will couple, and we will—

Yeah, we’ll merge. We’ll speciate.

All right, so next article on our economy: Jensen Huang announces that he expects $600 billion a year on AI alone. We’re seeing a massive continuation of investment. This is a good thing. We’re also seeing the AI spending frenzy propping up the U.S. real economy, and we’ve seen the impact, surging our GDP by 1%.

I think it’s interesting that AI infrastructure will reach $375 billion by the end of this year and is expected to be at half a trillion in 2026. The money is flowing out of Silicon Valley, out of sovereigns, out of family offices, and into the U.S. economy through the piping of AI.

I want to pause on this conversation here. This is “Nasdaq bubble soaring past dot-com records.” So I’ll read this. Here’s a chart looking at the Nasdaq market over the last 25 years. The Nasdaq’s market value has surged to unprecedented levels, now equal to 176% of the entire U.S. money supply and 129% of GDP. Both ratios are far above the dot-com bubble peak, signaling stock prices are racing far ahead of the real economy.

Let’s talk about this for a second. I think it’s important. Is this different from the dot-com bubble? Alex, any thoughts?

I want to pose a thought experiment. If we were on the verge of artificial superintelligence, what would you, Peter and Dave, expect the ratio of the Nasdaq market cap to M2 to look like?

Approaching infinity. Ripping upward.

Yeah.

Exactly.

So, this has all the hallmarks of the signature that one might expect. At least in the short term, one might reasonably expect there to be a concentration of rents around key publicly traded providers of AI infrastructure on the Nasdaq.

And then maybe at some point—again, this is not investment advice; this is an idle thought experiment—one might expect perhaps all of these rents and profits to become more evenly diffused throughout the economy, and then maybe we see a plateau at that point.

But this is exactly the signature that I would expect to see: this ratio of market cap to M2 ripping upward in the context of the eve of superintelligence.

Yeah, I completely agree, Alex. It is exactly what you would expect to see.

I also don't think the dot-com bubble was really a bubble. At the peak, Amazon was probably a bargain, and then it went down 90%-plus in the trough. Then we had 9/11 right after that, which turned out to be a great buying opportunity in the market. The internet was real; it was always real. The valuations got very high, but some great companies were in there. Google got started right at the bottom and then went public in 2004, as shown on this chart.

I think there's a possibility of the market coming down through panic, but it's not rational, because what's going on should drive this to the moon.

We also have companies that are real, profitable, and have real products and real services. That's very different from the dot-com world, from the Pets.com days and eras.

A lot of these charts are meant to scare you, too, because here you're looking at the NASDAQ. The NASDAQ is a bigger fraction of the market now because tech has become so big.

The P/E is a little high for the market as a whole right now, but it's not nearly as outlandish as this chart makes it look.

I like Alex's explanation the most. This is the signal that digital superintelligence is arriving.

Okay, let's move on to a conversation around health, one of my favorite subjects. We just saw an announcement out of the UK about an AI stethoscope detecting major heart disease in 15 seconds. This is a perfect use of technology, right? You put the AI layer right there at the stethoscope because, in medical school, you're listening carefully to all of these heart sounds, trying to hear a murmur and trying to hear “lub-dub” and variations thereof. AI can pick it up far better than this can.

We had a $10 million Qualcomm Tricorder XPRIZE. Of course, everything comes back to Star Trek, Alex, doesn't it? Reinventing or making the Star Trek universe real.

Such a strange universe, Peter. Again, biotech without longevity. Very strange universe.

Yeah, we'll get to that in a minute. We had this $10 million competition that Paul Jacobs, who was CEO of Qualcomm, funded at the time. To win this competition, you basically had to diagnose 13 different conditions, including anemia, diabetes, pneumonia, and sleep apnea.

The device had to weigh under 5 pounds, which is huge. Eventually, these things will become embedded. You also had to record 5 vital signs. This is a step in the right direction. It's still the beginning, but it portends what's coming next.

One of the companies I venture-backed through BOLD is called Echo EXO, and they build an ultrasound platform. Think about the kind of device: it's a handheld ultrasound that you can use to look at your baby or look at your carotid artery and so forth. The key was that it had an AI layer that would direct you on where to move the probe. It would say, “Can you move it upwards? Can you rotate it inwards? Can you hold it there longer?”

If you had this ultrasound probe, you became the physician. The AI guided you to do what you needed to do, then analyzed the imagery and gave you a diagnosis. I think that's pretty amazing stuff.

Totally. I would also invoke the statistical folklore that everything is correlated. I think this is just scratching the tip of what's possible in principle.

Going back to the Star Trek Tricorder, the key scenario I would like to see, and would hope to see unlocked, is using the power of AI and relatively diminutive hardware to simply infer the physiological state of an entire person at a distance from a few key, at-a-distance biomarkers with AI.

Well, I mean, AI is going to drive health care out of the doctor's office and out of the hospital into the home, where you're being sensed all the time and your AI agents are just watching and listening. It's going to transform health.

All of this will be cheap and free. It's going to be free because your company or your insurance company is going to pay for you to have those sensors in your home, on your body, and in your toilet because it just saves all the cost, right? Health care insurance is about keeping you healthy, not paying you after you've been sick.

Anyway, here's another fun article. I first saw this from David Sinclair, who posted it: “Psilocybin Shows Striking Anti-Aging Effects in Old Mice.” I added this because we have a community of folks out there who are interested in psychoactive molecules.

Check this out. This is a 2025 study that extends cell lifespan by up to 57%. In this study, they took aging mice, and you can see on this chart that about 20 weeks in—sort of late middle age—they started dosing them with psilocybin. At the end of the experiment, which was 28 weeks long—and mice really just live 2 to 2½ years, typically—the survival rate of those on psilocybin was 80%, versus 50% for those who were not on psilocybin.

I'm trying to find out what the dose equivalent for humans would be, but this is just another example of the continuous discovery of different molecules' impact on longevity.

Just to comment, Peter: I read the paper. It's a very interesting paper, and I think it's potentially promising. The authors do a great job of extracting downstream impacts. They were dosing with psilocin, which is a metabolite that normally emerges in humans and other large mammals from metabolizing psilocybin.

They looked at the downstream impacts. There was certain gene expression, and there were changes to the way telomeres in chromosomes are managed and regulated. That's a template where, ideally, one wants all of the anti-aging effects without all of the central nervous system psychedelic effects.

In an ideal world, we would find that we were able to extract that template. Notwithstanding the CNS effects, we could subtract those out and distill a template for a non-CNS version of this. I think that would be enormously impactful.

I wanted to share something I did on my summer vacation with our viewers and readers. I'm always experimenting and hopefully doing intelligent experimentation. When I see technology come along that I believe has a pro-longevity, high-reward, lower-risk approach, I'm open to trying it, researching it, and sharing the results.

A couple of weeks ago, I posted this on X. I went and did something called stem cell re-education. This was the work under Dr. Zhao. I want to share it because, for a certain group of people, this will be transformational.

For stem cell re-education, I was there for about 6 hours. My blood supply, which is typically 5 liters, was put through a machine 2½ times, for a total of 12 liters of blood cycled. My immune cells—my T cells, macrophages, lymphocytes, and so forth—were extracted from that, and I filled up a bag with about 300 cc, or a third of a liter, of my white cells.

Those cells were then co-incubated overnight with cord-blood stem cells. These are stem cells from a newborn, and my immune cells effectively went to school. They were put through a factory reset. That happened over about a 24-hour period, and the next day I had about 1.27 billion re-educated immune cells flowed back into my body.

My goal is to bring my immune system levels back to a much more youthful state, reduce inflammation, rebalance my immune system, pump up my stem-cell functionality, and increase my immune function. That's my goal for myself.

We're going to be flowing this technology in through Fountain Life as well. Our goal is to set this up at our Florida centers. It was amazing, and kudos to Dr. Zhao, who pioneered this work at Throne Bio.

I want to show 2 remarkable videos. If someone in your life is dealing with type 1 diabetes, alopecia, Parkinson's, or ALS, this technology is lifesaving. A lot of these diseases turn out to involve your immune system attacking your own body. Alopecia—the loss of hair—is your immune system attacking the hair follicles throughout your body, and you lose your hair. This process basically cures that. You regrow all of your hair.

Let's take a look at 2 videos. I'm going to show you first a 17-year-old teenager who has type 1 diabetes. This is where the immune system is attacking your islet cells in your pancreas, and you're no longer producing insulin.

This young man has developed neuropathy, and you can see him here prior to treatment. He cannot get out of his bed to get into a wheelchair. That's his normal state of function. Now let's take a look at 2 months later, after the treatment. I mean, it's a resurrection. He has been able to regain his function, and his type 1 diabetes has been eliminated. It's extraordinary.

Equally extraordinary is this Fountain Life patient who has ALS. ALS is a death sentence for most individuals, and I want to show you the pre- and post-treatment videos. This is the pre-treatment video: his inability to raise his hands above his shoulders. This is massively difficult.

Now let's look at a couple of days later and his ability to basically regain his function for someone with ALS.

I wish we had this for Stephen Hawking while he was still alive. I don't know if you want to comment on this, but I just wanted to share because I think this is the kind of regenerative medicine that our health-span revolution is undergoing right now.

Well, the only comment I'll make is that when you—

Talk about allergies and autoimmune disorders, there are so many interactions going on. It's immensely complex.

Yeah.

And it's a perfect fit for an AI that is specific to your body and your results. It's so promising that you can actually do something with this immense amount of data we can gather.

Yeah, I think, Peter, you're courageous, and now we know what you did last summer.

I'll report on the results. A huge list of markers were collected prior to my treatment, and then I'll report it out at 1, 3, 6, and 12 months. I'm excited for it, and we'll see where this goes.

All right, let's dive into robots, energy, and transport. I love this. We've seen NVIDIA with—I love the name of this—the Jetson AGX Thor generation of robot brains, right? Enabling real-time intelligent interaction at the edge. The Jetsons are here.

This delivers 10× more than their previous chipset, Orin. It runs generative and reasoning AI models at the edge. I love this. I double-checked these numbers: 2 million developers are using the Jetson Thor development kit right now. Alex, 2 petaflops of FP4 compute. For reference, that's approximately a tenth of a Blackwell, or maybe about 30 iPhone 16 Pros.

So when we talk about all the capex that's flowing into AI data centers, I don't think that's going to be bottled up in data centers for very long. We're going to see these AI chips—the AI compute—literally start to walk out of the data centers and onto the streets of the rest of the economy, the so-called real economy.

I think Thor is such an interesting case study in how this AI compute is going to be embodied in humanoid robots in an ergonomic form factor—both energetically and physically—and literally walk out onto the streets of the real economy.

Amazing.

Yeah. I really think this is important for all the students struggling with their largely irrelevant curriculum. This is what you should be doing. Your phone or your laptop will do about 30 teraflops now. So it's about 1/100 of what you can get from a probably $4,000–$5,000 NVIDIA chip.

Get yourself an accelerator on the side, or if you go to Andrej Karpathy's libraries on GitHub, you can get a really fast start. A lot of things that were daunting 6 months ago, you can just voice-code and vibe-code into existence on your laptop in under an hour.

You really can get your hands dirty with these toolkits, and then all of a sudden you're one of these people getting the $100 million signing offer. I got—how did I get from here to there? Well, I just jumped in and got my hands dirty.

I played.

Fun. It really is fun. We have an image here of Jensen with Brett Adcock, CEO of Figure. We had Brett on stage at the Abundance Summit last year. We're going to have at least four, maybe five, of the robot CEOs on stage with us at the Abundance Summit this coming March. And by the way, you're going to have the Moonshot mates, all four of us. We'll have Salim, Dave, Alex, and myself on stage at the Abundance Summit as well. We'll be doing a Moonshot recap, what WTF just happened in tech in the last three days of the Abundance Summit in March. If you want more information, you can go to abundance360.com to learn more about the summit. It is, for me, an epic part of my year and my life getting ready for that.

China's humanoid robot sales are expected to exceed 10,000 units in 2025, with year-over-year growth of 125%. I love this image of robots marching down the street. What could possibly go wrong?

But this is just the beginning, right? We heard when we interviewed the CEO of 1X Technologies, Bernt Øivind Børnich, that he expects to see hundreds of thousands flowing out of his factories. Of course, we'll see that from Figure, and we'll see that from Tesla. The prediction of 10 billion humanoid robots by 2040—it's coming.

Here's another article. I think this is pretty much Elon classic: Tesla is shifting Optimus' training strategy to vision-only.

We saw this with self-driving. He said, “No LiDAR.” I introduced the CEO of Luminar to Elon at a party, and Luminar makes LiDAR, and Elon just went, “Nope, no LiDAR. We're vision-only.” If a human driver can drive with one eye, we should be able to have AI do the same.

The switch here is no longer motion-capture suits. It's just going to be training robots based on video recordings of workers doing the work. Alex, you buy this?

Yeah. Obviously, this rhymes with the LiDAR versus non-LiDAR episode with autonomous vehicles, but I think the real story here is that, to the extent that you believe we're about to all be wearing smart glasses—that that's the next major form factor after smartphones—I can only imagine what fleet learning is going to look like when you have billions of people basically doing visual-based motion capture for humanoid robots.

Going back to the beginning of this pod, where we're talking about the bitter lesson, the bitterest lesson of all, arguably, for humanoid robotics is going to be when we have billions of people wearing smart glasses, doing fleet learning to power every single trade, every single manual trade, just by passively watching through the smart glasses—

—and recording all of human history in detail at the micro level. Right. No, that's right. That did come up when we were at 1X Technologies. Bernt Øivind Børnich, in his first 10,000-odd units, you have to use fleet learning; there's no option to turn it off.

So all the data, telemetry, and everything from your household is getting transmitted to the central learning engine. There's no human analog for that. Every job is training the centralized version of the robot. If it knocks over a coffee cup in somebody's house, then in the other 9,999 houses, the robot doesn't knock over the coffee cup.

Can you feel the acceleration? Can you feel the singularity coming?

Oh my God, I can for sure.

This is just in today. Apple is mandating all of its manufacturers—all of its tier-one suppliers—to automate, automate, automate; use robotics instead of humans wherever possible to increase reliability and reduce costs. Alex, a quick comment on this one.

Yeah. I think if the world finds its way toward completely redomesticated supply chains, robotics is probably the missing X factor for how, just as we were discussing earlier, tiling the world's surface—tiling the Earth's surface—with inference compute.

One can imagine a not-too-distant future where robotics enables essentially every sovereign country to, in some sense, redomesticate its entire supply chain if it has inference-time robotic capabilities to onshore every last bit of manufacturing.

Amazing. We'll see this at Amazon, FedEx, and all the companies that survive. The companies that don't do this aren't going to survive. I think it's going to be pretty cut-and-dried.

I'll hit one more robot story here, which is the competition between Waymo and Uber. It isn't really competition right now because we see Uber delivering 30 million trips per day, while Waymo is at 700,000 trips per month, right?

But here's the point: one Waymo robotaxi outperforms 99% of Uber drivers on a daily basis in terms of daily trips. These Waymos are efficient, and they're running 24/7, except for their charge time, of course.

Imagine, as these roll out more and more, that they will displace Uber. Uber is trying its own autonomous play. It's doubling down in San Francisco. It'll be increasing the fleet by 50% there. It's trying to get into New York. It'll have a lot of resistance there, but we'll see these technologies. And, of course, cyber taxi and cyber cattle are coming. Alex—

And remember, Peter, the post–World War II consumer automobile arguably created suburbia, created the suburb. What happens to urban planning when the cost of mobility is driven to zero? Do we see suburbs expand? What happens to roads? What happens to parking lots? What are we going to do with all the parking lots?

But again, the elephant in the room, on top of all of the hand-wringing over urban planning, is that there are so many other changes that are going to happen, probably on a much faster timescale than we can replan cities and suburbs, that maybe it's all meaningless anyway.

My plan for the parking lots: you're turning them into vertical farms, right? Each layer of the parking lot is growing a different—

Which is great if we need to be densely clustered together. But if we don't need to be densely clustered together, maybe it's something else entirely.

And by the way, I talk about the demonetization of everything, right? So driving will be 4 times cheaper than owning a car. The poorest people will be chauffeured around first and foremost.

And then how do you change the cost of living? Well, if you can live an hour from downtown LA, where the real estate is cheap, and you fly an Archer Midnight eVTOL back and forth to work—or you don't go to work.

You’re using Starlink to telecommute.

Or we move to other planets, or we upload to the cloud. There are so many different options.

I’ll take all of the above, please.

All right. I’m going to close on this particular piece, which is the U.S. electricity spike begins. This is a chart of the U.S. Consumer Price Index for electricity, and we’re beginning to see a spike that started in 2021 and is continuing. Thoughts on this, Alex?

Superficially, price signals convey demand. That’s why we have a price-based system. But I think the elephant in the room here is what happens if and when we get to recursive self-improvement.

We’ve already seen at least one demand shock, if you will. That was the DeepSeek Sputnik moment, if you will. What happens if and when there is some new algorithmic breakthrough that suddenly radically reduces the compute intensity of frontier models? Could the law of straight lines be violated, and could this burning upward, tearing upward of electricity costs reverse?

Yeah, remember—

The electricity is so easy to predict because we know recursive self-improvement is here right now. Well, we—I know it, anyway. I tweeted, “The world will know it very soon.” I tweeted this last week, right?

The human brain operates on 20 watts of energy, and I was playing with GPT-5 and asking what the equivalent compute cost, in terms of energy, is for one of the frontier models. It’s somewhere between 100,000 and 1,000,000 times more energy than the human brain. So this is a massive potential for improvement here.

We’ll get new chip designs, new strategies, and new approaches to make it more efficient. So we’ll build out all of this energy and data centers, and then if we get a 10^5 to 10^6 improvement in energy efficiency, that means we get that level of improvement in our total AI capabilities.

We can do better than that, Peter. The Landauer limit—we can blow past it with reversible computing. The human brain is by no means an optimal computer. There are lots of other, better ways we could build computronium.

Amazing, guys. Listen, as always, I love spending my time with you. I feel smarter afterward. I hope everybody listening enjoyed this episode. A lot more coming. We’re going to see the release of Gemini 3. We’ll be back on to discuss that.

We’re going to be coming on with a lot more of WTF Just Happened in Tech. Hopefully, this is your dose of optimism to counter all of the moaning pessimism coming off the media channels that people normally consume. I’ve stopped watching the news. For me, this is the news—the news that really matters, that transforms our planet, that is giving us increased longevity. It’s going to increase sustainable abundance—that’s the word.

Alex, some closing thoughts from you, then we’ll go to Dave. Black hole supercomputers.

Okay.

Okay. That’s a closing thought—

—which assumes we’re not living in the black hole right now.

Agnostic. We’ll follow up on the short-term implications of Leopold’s list and a bunch of other things. I think we’ll be back online again very quickly, within a week, with Salim back.

There’s so much happening now. I have a bunch of things we couldn’t even get to today, and then more will happen within the week. So, yeah, just a lot to keep up with, but this is the place to do it.

Yeah, I was wearing my Occupy Mars shirt in expectation that we’d discuss the Starship 10 launch, which was a huge success. So congratulations to Elon and his team and the team at SpaceX for that launch. It was awesome.

If you haven’t seen the video, please see it. It is proof that we’re living in the year 2025. Humanity is building fusion, going to Mars, and heading toward longevity escape velocity. You know, the only time more exciting than today to be alive is tomorrow.

On that note, gentlemen, have a beautiful week. Talk to you all soon.

Every week, my team and I study the top 10 technology meta trends that will transform industries over the decade ahead. I cover trends ranging from humanoid robotics, AGI, and quantum computing to transport, energy, longevity, and more. There's no fluff, only the most important stuff that matters, that impacts our lives, our companies, and our careers. If you want me to share these meta trends with you, I write a newsletter twice a week, sending it out as a short two-minute read via email. And if you want to discover the most important meta trends 10 years before anyone else, this report's for you. Readers include founders and CEOs from the world's most disruptive companies and entrepreneurs building the world's most disruptive tech. It's not for you if you don't want to be informed about what's coming, why it matters, and how you can benefit from it. To subscribe for free, go to dmmandis.com/metatrends to gain access to the trends 10 years before anyone else. All right, now back to this episode. [Music]