AI最新动态：失业、Elon 与 Sam Altman 的芯片竞赛，以及“AI泡沫”——与 Brian (Blitzy) 和 Emad 对谈

AI is useful. People pay for it because it has economic value. AI is not a bubble.

Nothing’s going to change the world more than what’s going on right now. It is definitely not a bubble. It’s not even vaguely like a bomb.

Until a few months ago, AI was like having a very smart goldfish-memory buddy next to you that you had to oversee all the time. Now it’s “set it and forget it,” and it can use millions of tokens and millions of lines of code.

Gemini overtakes ChatGPT in the United States. This is based on iOS sales.

Grok 5 could reach AGI first. It’s going to be a crazy couple of years with this, and there’s just not enough energy, compute, infrastructure—anything.

As a CEO and entrepreneur, do you worry about getting access to the compute you need?

You really want to be a player where everybody else wins when you win.

There will be some kind of a breakthrough in compute. Is it going to be from quantum?

Now, that’s a moonshot, ladies and gentlemen.

Hey everybody, welcome to Moonshots. Another episode of WTF Just Happened in Tech. I’m here with my Moonshot mates: Dave Blundin, the CEO and head of Link Exponential Ventures; Emad Mostaque, the head of Intelligent Internet and a dear friend; and Brian Elliott, whom you guys met on a previous podcast. Brian is the CEO of Blitzy. Emad, you’re in London today?

Yep, in London. Nice.

There’s a lot to discuss, as always. If you are ready to plug in guys, start taking notes, start listening. This is the world that is transforming how we live our lives. Before we begin with anything else, I want to talk about the wake-up call for colleges and universities. It’s pretty extraordinary.

This is a chart that just came out about how Americans perceive the value of college. People who say it’s very important have dropped from 75% in 2010 to 35%. That’s the wrong direction if you’re a college or university. On the other end, “not too important” has gone from 5% up to 24%.

For me, universities have a problem. Dave, we’ve been talking about this for a while. Your thoughts?

I couldn’t believe it. I knew this was happening, but these numbers blew my mind. I immediately sent it off to David Siegel, the founder of Two Sigma, because we’re going to go meet with Sally Kornbluth, the president of MIT, in a few weeks. I thought, “Holy crap, this is a really, really big deal.”

Peter, you’ve been saying it for a long time: the cost of tuition goes through the roof, and the perceived value of the education has been plummeting—not because it’s worth less in any fundamental way, but because what you can learn has grown so quickly and it hasn’t made it into the curriculum.

Remember, Peter, we had that meeting with one of the top people at MIT. We can build a nuclear reactor on campus faster than we will ever change this curriculum.

Oh my God. If you’re an accredited university, you’re not iterating your curriculum fast enough, and it just becomes irrelevant before you graduate. Tuition is up 180% since 2005. Room and board at a private university today is a quarter of a million dollars, and you’re saddled with debt. You don’t make it back because you’re not getting the jobs.

Brian, you’re closer to college than I am right now. How do you think about this?

College has been a credentialing program for a long time, right? The act of getting into MIT was actually impressive. It had less to do with what MIT could specifically teach you, because its curriculum is taught all over the world.

One indicator of this has been MIT dropouts. Yeah, totally—for free. Dropouts get funded incredibly fast. So what’s the point of staying for those extra few years?

There’s this unbundling happening right now between the credentialing that you can get just from getting into a specific school, going to Y Combinator, or having a really good portfolio site. There are other ways to get a credential that just weren’t possible before.

I’d love to see the graph of dropout rates in years 1, 2, and 3 increasing over time, especially in the last few years. You sort of stopped and started and finally went and collected the piece of paper. Tell me how you think about this.

It took me 20 years to get my pieces of paper from Oxford. That was a hell of a thing to do.

I think there were probably 2 things here. The first boom was the tuition-expense boom that we saw. I think that’s captured in the first part. The second part was probably COVID. COVID was a terrible experience for a lot of people in college, and I think it shook things up. Now we’re heading toward the AI drop, as it were.

We saw that paper by Erik Brynjolfsson and others that showed early-stage graduates starting to lose the ability to get jobs. That’s just going to grow. Again, it puts into question what that is, and that’s before we even get into the foreign students and what’s going to happen there with the visa changes.

This is the second blow. This is the kill shot. This graph is titled “College educated are unemployed longer.”

It used to be that you’d go to college to get your job, and we’ve talked about this on the podcast a number of times. The only career of the future that really matters, in my opinion—in all of our opinions—is being an entrepreneur. It’s not marching up the career path.

This graph runs from 2000 to 2025. What we see in terms of unemployment is unemployment among college graduates increasing, while everybody else with some college or just high school—in fact, high-school graduates—are becoming more employed. If they didn’t go to college, they went to trade school.

I love that one pixel there just last summer, where the most unemployable people in the world are college graduates. It’s hilarious, but it’s bad PR for colleges.

If you look back to 2000, that’s what I’m used to hearing: “If you go to college and graduate, you’re over twice as likely to get a great job.” That’s exactly what you see in the data just 20–25 years ago.

This is a pretty rapid shift in society’s perception of the value of a degree. Keep in mind that, within this entire chart, the unemployment rate is extremely low. It’s around 4.5%, so most people are finding jobs. Actually, it’s very hard for this year’s graduating class to find jobs. It’s shockingly hard.

What does this look like if we stratify the top 10 versus everybody else? I think there’s been a blowing up of people getting college degrees at, I would say, sub-tier institutions. It’s incredibly profitable for these institutions even though they maintain a nonprofit status. They’re growing the size of their employee base and their student base as a way to fund poor education.

I do agree with what you said earlier, Brian: what really matters out of a college education is the fact that you got accepted by a specific university. When someone says, “So, Emad, you went to Oxford,” or, “Peter, Dave, or Brian, you went to MIT,” they don’t ask, “Did you graduate?” They don’t ask what your GPA was. They don’t ask what you studied.

It’s just, “Yeah, I went to MIT.” That’s all that matters. That’s the highest-order bit. It’s crazy. There should be a brand-new program that MIT offers where it accepts you but doesn’t expect you to go.

All right. If you go to the next slide, it really makes Brian’s point. Here’s your tuition getting completely out of control, but those top 10 or top 15 schools are hugely endowment-driven. In fact, the endowment returns contributing to the budget are over twice as much as all tuition combined.

I want to read this for those who are listening. It says, “College tuition versus other expenses: cumulative percentage price change since 1983,” which is when I was at MIT. Tuition is up almost 900% over that time—a 5.6% average annual increase.

You’ve got a handful of schools that don’t even care about the tuition. They’ll be fine because their endowments are so big. Then you’ve got this really slippery slope of schools that need the tuition desperately to stay open at a time when people are not really perceiving the value of the degree.

That’s where it gets really ugly—right around schools numbered 40 to 400. If I’m a board member at MIT or Harvard, I’m probably not as worried. But if I’m at a second-tier school, I’m thinking, “Holy, what do we do?” We need to reinvent how we educate.

Emad, you and I have talked about the value of education and the fact that the best educator in the world will be AI. What’s your thought here?

I think there’s the credentialing part, but university became something that you just passed through by default, versus programs like Gauntlet and others where you actually have to work really hard to succeed and get through, with high dropout rates.

I think the world we’re going into is a very competitive one, where people who use AI—I mean, there’s nothing you can’t master with AI now, faster.

We've seen Alpha School and others show that, with just 2 hours a day of tuition, they're in the top 0.5% in the world. Even with academic papers, I think it's going to be similar. I think there'll be a huge amount of arbitrage because it just got too expensive. In the UK, Oxford costs $13,000 a year for tuition, or about $60,000 a year if you're foreign.

I think people will go to the networks, and they'll go to the places that embrace the technology to actually do what universities are meant to do: networks, knowledge, learning, and more. But we haven't seen the first AI university yet, which I think is going to be really interesting.

And really important. We’re going to have Mackenzie Price, the CEO, and Joe, the co-founder, who’s funded it, on a podcast coming up. For those of you who are moms, dads, or educators, we’re going to get ready for a fun episode on how to reinvent secondary education in high school.

All right.

Actually, I think just one final thing. Yeah, please.

Maybe the endowment should be putting big supercomputer clusters down because the universities in the US don't have them. That'll probably be the biggest determinant of research quality in universities: how many GPUs you have.

I totally agree. I love that.

I think it's a no-brainer.

All right, MIT, listen up here, and put the endowment to use.

They want to. There are forces in the school that desperately want to do exactly what Emad just said. I don't know what the friction is, but we'll work on it.

Get JPMorgan to fund it. There you go.

All right, let's jump into the AI wars, our favorite subject every week. We're going to kick it off with the fact that Gemini overtook ChatGPT in the US. This is based on iOS sales and 150 million users. We've seen Gemini go through this viral element. I love Nano Banana V3 and others, and they've jumped into the number-one position. Any particular thoughts here?

I did not believe it because ChatGPT had such a huge lead, so I checked the App Store data directly, and it is absolutely true. Now, this is the US, so ChatGPT is still miles ahead globally. Google can use its massive distribution power to push that. That's how Chrome bypassed Firefox: you just push it out.

It didn't bypass it. It blew it away.

It blew it away, actually.

You know, I checked Polymarket on this. Interestingly enough, I checked Polymarket, which is really fun to do if you guys—if our subscribers—haven't done that, look at it. I first asked when Gemini 3 was coming out. We've been waiting for Gemini 3 to do an episode on Gemini 3. The current top prediction is 40% by October 31, so maybe by the end of next month.

But here's the other prediction: which AI model will be in first place—the best—by the end of September? There was a 99% prediction for Google. But what was fascinating was the prediction for the second-best AI model by the end of September: 91% for Alibaba, for Qwen. I find that amazing. How do you think about that, Emad?

Well, I think we've seen the gap close dramatically between those models. Qwen is releasing almost daily now. Today, they had 6 model releases.

Wow.

They're just accelerating. I think it would be difficult for them to have the best model, but they have such reach with the billions of users that Alibaba has, the amount of data they have, and they've got a really kick-ass team there. Distribution matters so much. I don't know anyone that uses Threads at number 3 there, right? It has 400 million monthly active users and 115 million daily active users. I feel that this Gemini-ChatGPT thing is the same, which is why people are going to be doubling down on distribution.

Now that we've got reinforcement learning really coming through the models, that's actually how they'll get really good. I think that's going to be a real differentiator as we go forward. Again, we'll probably see the Qwen models keeping up because they're used so widely now, everywhere. They are closing that gap.

All right. Not to be left out of the conversation, Grok 5: this chart says Grok 5 could reach AGI first. We've seen it beating all the AGI benchmarks, and in particular, Grok 4 has reached the top mark on the ARC-AGI benchmark, which is the Abstraction and Reasoning Corpus. Right now, on ARC-AGI v2, Grok 4 has hit 15.9%, which is the highest known. Are you tracking these, Emad?

Yeah, I think we're continuing to see scaling coming through here, and this is going to be the first big mega-run that we'll know about. Again, OpenAI might release their verifier runs, but all these benchmarks are saturating so fast. I think Epoch AI predicted that every benchmark in the market today will be saturated within 3 or 4 years.

We need new benchmarks.

Just simple extrapolation—

Dave.

We need new benchmarks. But the crazy thing—

Crazy thing is—

This is V2, though. We already saturated V1, but V2 is crazy hard. If this one saturates, then you're beyond superhuman intelligence. If this one saturates in 3 years, we're in another universe, which it probably will.

I think more important is the cost per task, right? On the x-axis, it's very clear that if you're willing to throw more dollars at this, you're able to increase performance. I could care less, on the last slide, who the consumer user is. It's about when we throw more dollars at which models, whether we increase the quality of performance. That's going to determine who ends up winning.

Yeah. It's still only been 1 year since o1 was announced.

That's crazy. Ancient history.

That's funny. And so, here we go: Grok 4 Fast Reasoning. I love these names, right? Just appending things on the end of them. Grok 4 Fast Reasoning ranks number 1 on the extended New York Times Connections benchmark. What is that? This is based on New York Times puzzles where players must group 16 words into 4 groups, each belonging to a common semantic category.

The original version has 436 puzzles, and the extended version has 759 puzzles. These are just vanity benchmarks to brag—to get bragging rights. I talked to Alex about this one; he's off in Europe today. By the way, I should say Sem is MIA. Sem, where'd you go, buddy? Alex is on a top-secret mission in Europe. I'll leave it at that.

This is a really fun benchmark, though, because if you go to The New York Times and do the Connections test—it's a daily puzzle—it's really fun. My wife does it every single day with her friends. They made it harder by adding more categories to it, 4 more categories, and you have to get it right the first time.

When you do it on The New York Times website, it gives you 3 wrong answers before it says, "No, you're wrong." But the AI has to get it right the first time. It really is a good test of general intelligence. Shockingly good.

The theory here is that the big foundation-model companies are going to benchmark it. They'll train on a bunch of data specific to this puzzle type to try to max it out. You saw when Brian did the Blitzy announcement on our podcast, he was very careful to say, "We topped SWE-bench, but we did not tune or benchmark to that test. It just happened this way."

We're almost sure that they're trying to get the PR by benchmarking and optimizing toward the problem.

But you can't prove it. It is a crazy-high score, though, to get into the 90% range on this.

Amazing. All right, here we go. This is going to be the data-center wars: xAI's Colossus 2, a gigawatt-scale data center with 110,000 GB200 GPUs, 119 air-cooled chillers, and Tesla Megapacks. I love this. This is the beginning of Colossus 2. Emad, you're tracking this, I'm sure.

Yeah, I think Elon said he's going to be the first to a gigawatt, the first to 10 gigawatts, and the first to 1 terawatt.

Yeah, this is his tweet from today. OpenAI is bragging about its NVIDIA partnership, and here he is saying, "Just as we were first to bring a gigawatt of coherent training compute online, we'll be the first to 10 gigawatts, 100 gigawatts, and 1 terawatt." I love that.

It's basically as much as a state now. These things will be drawing down. The whole of Bitcoin's energy use is about 20 gigawatts, if you look at it as well. That's about as much as all of Argentina. Argentina is a 10-gigawatt power center.

I think what's going to happen now, though, because you don't have the infrastructure, is that we're going to see massive solar and battery buildouts, and that's going to be super interesting as you scale there. I don't know how else you're going to do it unless you have these small-scale, literally nuclear reactors. In fact, Microsoft has co-opted nuclear power everywhere. So it's going to be power wars across the US.

Amazing. Here's our article on NVIDIA investing $100 billion into OpenAI. There was a great CNBC piece that had Sam Altman, Greg Brockman, and Jensen Huang speaking together, and let me just quote what they said.

Sam said, “$100 billion is a small dent in the scale of our plans for 10 gigawatts of compute. This data center will be a multi-square-mile level of infrastructure. The stuff that will come out of this superbrain will be remarkable.”

I love that: multi-square-mile superbrain. I mean, tiling the world. Greg Brockman then comes on and says, “We really want everyone to have their own GPU, so agents can do work for you while you're sleeping,” which means that we're talking about on the order of 10 billion GPUs. The deal we're talking about with NVIDIA is for millions of GPUs. We're still orders of magnitude off. We're heading toward a future where the entire economy is powered by compute, and it's a future where compute is scarce.

And then Jensen comes on finally and says, “This project is 10 gigawatts, or roughly 4 to 5 million GPUs. That's approximately what we're putting into one project—what we sold all of last year, and double what we sold the year before, and double what we sold the year before.”

So, just a massive increase. Dave, how do you think about this?

Tie together those last few slides and really open your mind to the compute scarcity that's coming up. So you've got $100 billion. The US venture industry is about $200 billion a year. Here you've got a single investment by a single company that's half of all US venture in a year. Where's that going to go? It's going to go into buying chips and building data centers to support the users. Well, how many chips is Jensen going to be able to make this year? It's about 5 million.

Yeah.

Okay, 5 million chips. This deal buys a lot of them—20% to 30% of those, by itself. When you look at the other slide that Brian commented on, on the x-axis, wow, this stuff gets more and more intelligent and useful as you throw more hardware at it. How much more hardware? A lot more than we actually have on the planet. All these demos you're seeing, all these things, these benchmarks—there aren't anywhere near enough chips to deliver that to 7 billion people around the world.

So we talked about where do you invest? Chip manufacturers, the construction to build out these data centers, the power plants to power these. I mean, we're converting electrons into intelligence and into crypto. Emad, how do you think about this?

Yeah, I think whoever controls this is the marginal producer in the economy, right? If you look at OpenAI's projections to get to $200 billion, $80 billion comes from this brand-new AI agents line, and another $20 to $30 billion comes from elsewhere. They're going to be rolling out AI workers that work around the clock, and then the investment, as you said, is a supply chain.

But then it's also the companies that can have the expansion in margins because they have pricing power and they'll be replacing humans with AI. And then downstream, the impact, I think, is going to be probably actually in the attention economy, because it's about the only thing that isn't scarce: human attention.

So we're going to look more and more toward media, which might be a bit counterintuitive.

Interesting. Brian, as a builder, as a CEO, and entrepreneur, do you worry about getting access to the compute you need?

You really want to be a player where everybody else wins when you win, right? And so you really want to be sort of model-agnostic. You want to be provider-agnostic. You want to lift all ships. So I don't think, if you are a healthy player in the ecosystem, that it is a huge concern, but you can't be 5th or 6th, right? You have to be the most important to these folks. And so economies of scale are going to matter a lot here.

Here's a chart reinforcing this. Lab compute has 3x'd in just 1 year. We see a graph showing OpenAI, xAI, Meta, and Anthropic. OpenAI is at the top, with xAI coming on strong. We don't see Google on this, or Alphabet, which is interesting. I don't know if anyone has any comments on this.

I'm going to couple it with the next slide here, which is that data center capacity is expected to go up 4-fold by 2030, going from 44 gigawatts to 156 gigawatts. Forty-four gigawatts today, 156 gigawatts by 2030. And from what I'm hearing, that seems like a lowball estimate as well. McKinsey always underestimates.

Demand, I think, somewhere else in here, is going up 10x year over year. Supply is going up very quickly, but nowhere near as fast as demand.

So what does that mean?

What you see with all the model providers is that they're trying to offer fast or smart or whatever. But what it's really doing is rerouting your query to the smallest model that can answer the question to try and save some compute. Meanwhile, they're all working on internal self-improvement, so that's eating up a lot of compute at the same time.

So you're starting to see the cracks in the supply-demand curve here. Your question for Brian was a really good one: Do you worry at all about getting access? I think that a lot of the use cases that'll be deprived of access are the virtual girlfriend and doing your English homework, because Brian can overpay 100x or 1,000x over those use cases, so he won't get cut off. But there is going to be a huge supply shortage, for sure.

Oh, yeah. It's about the marginal dollar, right? Costs are going to go up dramatically, I believe. At the same time, costs are depreciating from the actual cost basis of the chip. But the model providers are going to be able to increase costs if they're number 1, and we're willing to pay that. We're willing to really pay anything because it's much more valuable, what we're able to provide, than these consumer-type services.

Yeah, I think economic value per FLOP is just going up dramatically because you're at this inflection point. AI, until a few months ago, most people were using GPT-4o. It was like having a very smart goldfish-memory buddy next to you that you had to oversee all the time. Now it's set-it-and-forget-it, and it can use millions of tokens, millions of lines of code, and then be proactive.

And as Brian and Dave said, it'll be the marginal dollar going up. But if we look at the previous one—just put it in context—we had our launch party at Stability AI, I think 3 years ago, at the Exploratorium, and we had a slide go up saying, “We have the 10th-fastest cluster in the world at 4,000 chips.” Now people are talking about 114,000-chip deployments.

The reason for that is literally just because of this economic thing. The amount of economic labor that AI can do has gone from maybe 1% or 2% now to, in the next few months, it'll probably be 50%—in the next year, actually. And so this is all complete. This isn't a bubble. This is all very reasonable because your TAM—your total addressable market—has gone up so much. And so, yeah, it's going to be a crazy couple of years with this, and there's just not enough energy, compute, infrastructure—anything.

Amazing. This is Greg Brockman on that very subject.

I think part of the 2030 outlook is we will be in a world of material abundance, right? I think that AI is going to make it much easier than you could almost imagine to create anything you want, right? And that will probably be true in the physical world, in addition to the digital world, in ways that are hard to predict. But I think it'll be a world of absolute compute scarcity.

And we've seen a little bit of what this is like within OpenAI, right? The way that different research projects fight over compute, or that the success of the research program is determined by the compute allocation. And so one thing we think about a lot is: How do we increase the supply of compute in the world, right? We want to increase the intelligence, but also the availability of that intelligence. Fundamentally, it is a physical infrastructure problem, not just a software problem.

Could you imagine the ongoing conversations inside of OpenAI and the arguments about, “No, I need the compute to do this project”?

Yeah. Crazy.

I don't know if you remember, Peter, but when we were at OpenAI headquarters a few weeks ago talking to Kevin Weil—

Yeah.

We asked—or I asked him, anyway—about the division of labor between him and Mark Chen, Sam, and Greg Brockman. And he said, “Well, Brockman's out there just getting compute. We need compute, like you, so he's just out there finding it.”

So, yeah, I kind of miss the days when Greg and Sam used to do these things together. Sam is on the road constantly now, so Greg has got to be in the house finding the compute, but he used to do a lot more podcasting. It was really nice when they were a 2-person team.

But I don't know—everything he said is exactly what you were just saying.

The theme of today: abundance everywhere except compute scarcity.

There will be some kind of a breakthrough on compute, right? Emad, what’s your bet on where we might get some sort of new breakthrough at 10x efficiency or power use? Is it going to be from quantum? Is it going to be from thermodynamic computing? What do you think?

I think it’s probably a data story right now. If you look at the Qwen model from Alibaba’s Tongyi team, they managed to score, I think, 22% on Humanity’s Last Exam with 3 billion active parameters, with a self-reinforcing continual-learning model that runs on a smartphone.

They did it through improved data. Again, this thing Dave said about better reinforcement learning and better approaches—I think there’s a data hybrid, reasoning, and other things coming together to optimize for specific tasks in the economy. Again, that 50% of tasks—that’s how you can route this down to be highly efficient, and we don’t know where the lower bound is because we could have more breakthroughs.

We could have improved chip performance. Again, we’re going up 5 to 10 times a year on chip performance, and it’s just very hard to extrapolate this. The only reason you can say it’s going to reach this crunch point is simply because the amount of work that can be done in the global context is so large. At this inflection point, there’s no way that we’ll be able to get them efficient enough. That’s the only way we can look at it.

And if I could just riff on that for any of the entrepreneurs out there, what Emad just said is a really good barrier to entry if you work on it within your domain. If you said, “Hey, there’s all this technology and research related to transfer learning and distillation that allows me to get the exact same quality of result with 1% of the parameters and, therefore, 1% of the compute,” by all means, do it.

Right now, we’re all used to, “Oh, I can just get an AWS account tomorrow, and I can just sign up and pay, and it’ll be there for me forever.” It’s like a utility. The whole cloud-computing era tried to convince us it was all a utility. It would always be there. Well, lo and behold, nope. It’s a scarce resource. Greg just said it. He’s always right. It’s not a utility.

Have a plan. You need a plan today because Bill Gross was saying every mountain with a lake next to it has already been bought for pumped-hydro power storage. You missed the opportunity to buy your mountain. Don’t miss your opportunity to reserve your compute, because it’s now or never. These things get locked up very early. This is a competitive world. Dave said it 100 times: literally, if you have task-specific data sets, distillation, and the right verifier, it is a 100-times difference in the cost of executing a particular task.

Yeah, this is a perfectly viable business plan. This is what happened with storage. This is how Dropbox got so big, right? They were the first folks to use S3 and not have their own data centers. There were 100 other storage companies. They were just 10 times cheaper than everybody else, and they scaled off of that and built a very powerful company. The same thing applies to models.

Looks like OpenAI may get the shackles pulled off. OpenAI reached a deal with Microsoft to allow restructuring from a nonprofit to a for-profit. OpenAI is targeting a $500 billion valuation as part of that.

I know what it’s like to flip a nonprofit into a for-profit. I did it with Singularity University many years ago, and you need to leave a certain amount of capital and capabilities inside the nonprofit. I won’t go through the machinations of how you do it, but as they do this, OpenAI’s nonprofit will be left with about $100 billion in capital. It’ll be the largest nonprofit endowment out there, which is amazing. Imagine what they’ll do with it.

You remember, Dave, we met with someone at OpenAI—I won’t mention who—and they’ll likely be in charge of the nonprofit. They have incredible vision for what they want to do to solve humanity’s biggest problems with it.

If you guys remember, Microsoft invested $1 billion in 2019 and another $10 billion in 2023. They’re estimated today to own about 30% of OpenAI. That’s unconfirmed, but that’s the estimate. This sets them up, basically, to become a multitrillion-dollar company.

Microsoft can’t lose.

Just for context on the earlier part of the conversation, OpenAI is twice the size of Harvard’s endowment fund, which for the longest time had been the largest endowment fund of all time. From a nonprofit standpoint, OpenAI has doubled the size of the endowment in just a couple of years.

But can you imagine the relationship between OpenAI and Microsoft right now? For example, when the NVIDIA–OpenAI deal was struck, Microsoft was notified the day before. OpenAI used to get all of its compute from Microsoft, and now Microsoft has been sort of kicked to the side while OpenAI is growing unshackled. Fascinating.

I remember when Masayoshi Son had that commitment for $100 billion to OpenAI, and then someone asked Satya about it. He said, “Well, I’m good for my $90 billion.” I think Jensen Huang is definitely good for his $100 billion.

These are all crazy numbers, right? When Microsoft invested $10 billion—or $1 billion—we were like, “That’s big.” Now it’s like, “Only $100 billion for the nonprofit.” It’s the second largest, and we don’t even blink at $100 billion being invested in them.

Because, literally, they will have a trillion dollars of build-out.

I think Elon Musk said something again recently. Someone asked, “What about Anthropic?” He was like, “They never had a chance,” because really, who can scale now to compete? xAI, Google, OpenAI, and probably Meta.

Speaking about that, this goes to our next slide. The title here is, “Zuckerberg Says Better to Lose Billions Than Be Late to Superintelligence.” He’s committed to investing $600 billion in U.S. data centers by 2028. Why? Because, “I don’t want to be second to superintelligence.” Crazy.

It’s just staggering. The sheer size is staggering, but also, the lives these guys are living are completely unprecedented in the world. Zuckerberg was just at the White House a week ago having dinner with—look at the table. Look at these people. The president is saying, “How much are you going to pump into the U.S. economy?” and Zuckerberg is like, “$600 billion.”

This has never existed in the world before, and I don’t know—this next couple of years is like nothing in human history.

Sounds like inflation to me.

The economy is dependent on its capital stock. We build our universities, our factories, and everything else, but basically, all this is the investment for the new economy. The economy 5 to 10 years from now is run by AI and powered by AI, so it makes sense that you’ll spend trillions of dollars on this.

These guys want to get it first from an economic point of view, but then there’s more than that. Do you remember the story of how OpenAI got going with Larry Page from Google and Elon Musk?

Larry Page.

From Google and Elon Musk—the Larry Page discussion. I was there for that argument, where Larry called Elon a speciesist.

Yes. Peter, do you want to tell the story there?

Oh, no. Go ahead. Go ahead.

It’s because they were discussing intelligence, and Larry Page was like, “Digital intelligence can overtake humanity, and that’s fine.” I was like, “No—humans.”

Larry Page is reportedly willing to make Google bankrupt to get to superintelligence first. They won’t, because they make so much money, but these are big stakes now.

It’s worth just stepping back and comparing a day in the life of Mark Zuckerberg to Sam Altman. Sam is literally getting attacked constantly from every side, especially by Elon, while needing to beg for money from any source he can get it. He’s traveling all over the world, trying to hold this together, while taking a nonprofit and turning it into a for-profit, which is a logistical nightmare. He’s dealing with all of that.

Zuckerberg just needs to call his CFO and say, “You know what? Go ahead and divert that money back into data science, and I’m going to go have a mai tai.”

They’re printing money on my ship in the Caribbean.

The market will reward him for it, too.

It’s an unreal existence. Dave, you said it right. I don’t know how you remain grounded as the CEO of one of these companies when you’re speaking about literally trillion-dollar deals that you’re involved in. It’s crazy.

Yeah, that’s a good concern, too. I kind of trust the people who have struggled, either struggled before in their lives or are struggling right now. But you do worry a little bit about the scale of power in a few people’s hands and what decisions they might make tomorrow.

But we are, just to remind everybody, on a war footing. Going back to what you said a few minutes ago, Emad, we’re on a war footing getting ready for the next economy.

Just like we came out of World War II with brand-new interstate highways in the United States, aerospace, and automobiles, we’re gearing up for a new economy that will displace the old economy. It’ll be tens of trillions of dollars, include robotics, and be close to $100 trillion over the course of a decade.

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

All right. Let’s look at what comes up next here: Dario Amodei on Claude designing Claude. This is a quote from Dario. He says, “Claude is playing a very active role in designing the next Claude. We can’t fully close the loop, but the ability to use the models to design the next models is not yet going super fast, but it’s definitely started.” How long before it’s going super fast, Emad?

I think it’s the takeoff point right now. There was a recent interview with Tri Dao, who is the man for writing CUDA kernels on NVIDIA. He’s at Together AI, and he came up with FlashAttention, which literally increased performance 30%. He said, “I use Claude Code, and I’m at least 50% better,” and this guy is the cream of the cream.

We’ve seen that from top people already. That self-recursive loop is coming inevitably. We’re already seeing TPUs being designed by AI. Sam Altman recently said—again, fantastic CEO, fantastic capitalist—that his plan is to get 1 gigawatt of new compute every week with a fully integrated system that could also be training its own models. So I think we’re moving full-stack, vertically integrated, from chip silicon to model feedback loops, and there’s no way that won’t speed things up even more. Can you feel the acceleration? Oh, my God.

Well, look, I spent about 6 years of my life purely building neural networks and researching neural network algorithms and code. It’s very similar to discovering math, which Alex Wissner-Gross is always talking about. Alex Wissner-Gross was on this pod predicting—I think it was 18 months—that we’ll be solving all math.

Emad, you’ve been on this bandwagon as well, right?

I mean, look at AI winning the gold medals in the ICPC and the math Olympiads and things like that. You parallelize it. We’ve been running 1,000 Lean provers in parallel, analyzing things. Next week, we’re releasing a full stack of economic proofs that you just can’t argue with for everything.

These AI—once you actually apply them, it’s not like you have one genius. One genius is enough, but Dario said data centers of geniuses checking each other’s work in parallel. Obviously, you’re going to get that next step up from that, and we’ve seen things like Terence Tao formalizing various proofs. They only got about 20–30% of the way in parallelizing that. I believe it was Mathlib that managed to do the full proofs in 2 days. So, yeah, I think it’s a good chance. I have no idea what the implications of that are.

Also remember Noam Brown over at OpenAI, when we were there, was saying that their progress in core AI research is gated by compute now, not by researchers. They have a backlog of ideas. They just don’t have enough compute to try them all.

So pretty soon, the AI will also be generating the ideas, and then the backlog is purely compute. That’s where Elon is saying, “Well, I’ll have the most servers, therefore I’ll win the race.” But it’ll all be compute-constrained. There’s a window of a year or 2 here where it’s also idea-constrained, so there are lots of opportunities for people to think really hard during this window. Very soon, it’ll switch to AI-generated ideas. You know, $10 billion here, $100 billion there. Pretty soon, you’re talking about trillions.

Alphabet becomes the fourth company to reach what I like to call the four-comma club, reaching a $3 trillion market cap. It joins Apple, Microsoft, and NVIDIA in this $3 trillion market-cap club. The stock is up 33% in 2025 and 55% over the past year. Right, I was talking about this last time. For me, Google has been an extraordinary bet. Any comments here? I mean, I think the prediction markets still hold Google and Alphabet to be the long-term winner. Emad, do you buy that still?

I mean, they’re fully integrated with thousands of amazing talents. Demis Hassabis is at the head of DeepMind, and they’ve got the reach, right? So you’ve started to see AI search results. It’s not quite good enough because they’re doing the crappy models, but when Gemini 3 Flash is better than Gemini 2.5 Pro, the directionality of where things are going—again, they’ve got the full stack. They don’t need to pay the NVIDIA tax. They can build everything themselves, and they have massive cash. So why wouldn’t they be up in the lead there?

And just to add one thing to that, we spoke about OpenAI doing their own chips with Broadcom, but they just started. Google’s been working on its TPUs for years, so they’re years ahead in that vertically integrated solution.

They’ve been their own customer for a long time because they run Google on the TPUs. TPUs are probably 5 times more power-efficient than NVIDIA chips, and they have better interconnect for large-context models as well. So they’re pretty much ideal for what’s coming through now.

Nice. How do you know that, by the way? I thought that was impressive knowledge.

We used thousands of TPUs. We were down there when they didn’t have racks.

Oh, you had Stability AI. You had hands-on access. Oh, no way. Yeah, because they’ve pulled them from the market effectively because they’re using them all internally now. So it’s kind of hard to get performance specs. That’s really useful information.

Well, they might actually start selling them soon. We’ll see.

Yeah. We’ll see if they can. There was a little video clip put out by Mustafa Suleyman, the CEO of Microsoft AI, which I found somewhat compelling. Let’s take a listen to it.

At the moment, these models are still one-shot prediction engines. You ask a question and you get an answer. It produces a single correct prediction at time step T, and they can’t lay out a plan over time.

The way that you decide to go home this evening is that you first get up from your chair, then open the door, then get in your car, and so on. That is just a computational limitation. Just as today there’s a kind of superintelligence in our pocket that can answer any question on the spot—we dismiss how incredible it is right now. It’s magic in your pocket.

Now imagine when it’s able to not just answer any question about poetry or some random physics thing, but it can actually take actions over an infinitely long time horizon. Just that capability alone—and I think that we basically have that by the end of next year.

So I found that compelling. Dave, what do you think?

Well, I love the core point: this is absolute magic, and it came into the world so quickly. There are so many ways to take advantage of it that we’ve only begun to scratch the surface. I do disagree that the planning ability—I don’t know when this was recorded—but the planning ability has gotten pretty damn good pretty damn quickly. This might have been 3 weeks ago, but today is different.

Ancient history. Do you think a Tesla can self-drive from one side of America to the other?

100%.

That’s planning.

It’s crazy. And then just think: you hook that up with a vision model, so it’s making notes and writing the Great American Novel as it drives. Again, we actually have all the tools there.

And in fact, Brian, you’re the expert in this, right? What have you seen in terms of massive long-term stuff?

You can achieve AGI-type effects at the application layer, right? This long-term horizon of planning can’t be done extremely well at the model level, but from a user or consumer perspective, who cares, right? It’s about what I experience, which is a long-horizon plan given to me from a set of models. I would say we are in this reality today for a number of domains, including software engineering.

Nice. We had the CEO of Replit on the pod recently, and I loved this quote about how to use agents. Let’s take a listen.

If I’m someone with truly unique domain expertise—let’s say I’m a lawyer who is top in the world at solving certain cases that are very rare—and so I have this domain expertise that I’m not going to share in the open source. I’m not going to sell to Scale AI so that they can sell it to OpenAI or Google, all those. I’m just going to keep this resource to myself.

But the way I would monetize it, instead of going and selling my services directly, I would imbue this knowledge into an agent that becomes this very specialized agent in this very specialized domain, and then I can scale myself.

So, I like that. One of the questions we've been asked in the comments on this pod—and we do read the comments from all of you listening or watching on YouTube—is: You talk about what you should do if you're 18, 19, or 21. What should you do if you're in mid-career? How should you be thinking about AI? This sounds like a pretty good example, Dave. How would you answer someone mid-career?

That's a tough one. Maybe I'll bounce that over to you big brains. I totally get the concept: I've got domain knowledge. I'm a lawyer, I'm a doctor, or I'm a specialist in a very specific domain. There's very, very specific domain knowledge all over the world.

I know the RLHF companies, like Invisible and Mercor, are killing it by wrangling all that technical knowledge and getting it into the models. So, the question then becomes: I want to monetize that, but once it's ripped off my brain, they may pay me a lot for a month or 2, but then what? I've just completely dumped my knowledge into the AI. Do I have any value?

I don't have a good answer off the top of my head for how to capture that. I would say there's no barrier to starting a company. You don't have to be 21 to start a company.

You absolutely should. There's so much greenfield opportunity out there. I love these companies that have a regulatory barrier, or a vertical domain, deep-tech, or deep-knowledge barrier. Just start a company using AI in that category. That's how you might then say, "Okay, now it's sustainable, and I can make a career out of it."

So, that's always a good choice, but you've got to leave your day job and go do it. I would translate that to: Find a good problem that you understand deeply, that no one has yet solved, and go build around that problem.

There's never been a better time for these domain experts. I'm more bullish than Dave is on this 45-year-old audience. If you think about it, software has never been easier to build. That is true, but software is just 2 things. First, it's the technical design and build of it, but you're imbuing a business process and a set of flows and decisions that you need a user to make.

These insurance folks and financial services folks are world-class at understanding how to price products dynamically to the market. That has very little to do with technology selection. So, we can empower these folks with platforms like Blitzy to build large-scale systems and enterprise systems that are purpose-built for that 45-year-old insurance underwriter or financial product person. This has never before been possible.

Yeah, yeah. I've been thinking a lot about this. Nassim Taleb, the Black Swan guy, has this great concept called "intellectual yet idiot" about very well-credentialed people who just don't have any skin in the game, so they don't give a damn, right? That's a flaw of many of our systems. AI models are intellectual idiots; they don't give a damn.

One of the most important things is actually giving a damn about the context in which these are implemented. If you think about the long tail of these implementations to solve problems, if you actually give a damn, can communicate it, and be that intersection, that's where you get the most leverage. You actually need to understand the consumer and how they operate today. You need to have some skin in the game in the way that you do that.

I think people underestimate that because we just assume the technology will sweep through, because people do these analyses and understand the way we do. No, there needs to be that translation layer, and you need to be able to communicate and show that you give a damn.

So, my advice to that 40-, 45-, or 50-year-old individual who's asking, "How do I apply AI to do something significant in my life?" is that, when you look at the problem, there is the intellectual part: "Hey, I've got cognitive surplus now from these tools."

But the next part is getting to understand the organization that you're in. If you're within your organization trying to improve it, you need to understand the real checks and balances and who you need to communicate these things to in the appropriate way. And then, if you're servicing someone, having that really high-touch consumer aspect where you're helping them through something that's very scary and has huge potential will pay a massive amount of dividends.

Again, you can use the AI to help you communicate and things like that as well. That human touch is underestimated, particularly as we diffuse from just the early adopters to the vast middle of this industry.

Yeah, we're just at the beginning of this game.

Well, Peter, that video was of Amjad Masad. Do you want to tell the story about how easy it is to build software?

I was flying from Santa Monica up to San Francisco, to Stanford. Dave was already there, and we were interviewing Amjad. Emad was there too. We were interviewing Amjad about Replit.

I had downloaded Replit, but I had never really used it, and I thought, "Damn, if I'm not going to give it a try." I have Starlink on my SR22 Turbo airplane, so I was flying the airplane on autopilot. I had the Starlink antenna in the front, plugged into Replit, and coded up a mindset app on the flight there. It was fantastic.

It was so easy—zero requirements. I just needed to know the single most important thing. Again, if you're new to this, if you're just a fan of this, if you haven't played with it at all, Replit is amazing. There are other platforms, like Lovable and others. It's critical for you to just try, just try and play.

Bring a curiosity mindset and your playful mindset. If you know what you want to exist, the AI systems will help you get that into existence. It's only going to get easier. Your domain knowledge will be extremely important in product creation.

So, before I move past the conversation about Replit and vibe coding, Brian, you're taking this level of coding to a brand-new level. How do you apply this to industries and entrepreneurs? What are your thoughts?

There are 2 classes of software. There's this disposable, widget-based software that Peter built with Starlink on his plane. This is the idea of getting a concept into a prototype. Then there's true enterprise-scale software: I'm going to have thousands or hundreds of thousands of users, I'm going to have concurrency, and I'm going to have good caching. That's the part of the system where Blitzy fits in.

Everyone's having this Peter experience, where they can create something quickly, and then they're getting to enterprise scale and getting none of those gains. We've brought the vibe-coding speed to the enterprise scale. We can do that for the new entrepreneur building the insurance product, or for the existing enterprise that's doing large-scale development.

The idea is that velocity from an engineering perspective is dramatically higher than it's ever been. It's never been a better time to build.

Do you interface with mid-level managers in companies, or does this have to be top-down for people who want to use Blitzy to improve their products and capabilities?

Yeah, anyone who leads a large engineering team comes and works with us. Lots of times, CTOs and CIOs will come meet me directly, but you also have VPs of engineering who say, "I'm going to make my company go faster. I'm going to weigh in, bring Blitzy in, and be the first to do it." We love those folks too.

Got it. Great. Here's our next one. It comes from Andy Jassy, CEO of Amazon. He's like, "Wait, wait, wait. You know, we're going to build glasses too. Meta is not going to lead the way here. There's got to be someone else."

Amazon is developing its own AI glasses to challenge Meta. What I found fascinating is that there's going to be a consumer version, but importantly, there is a version that's going to be used by their drivers. The drivers are going to be recording everything. For what use? It's to train the future robots.

This is codenamed Jayhawk and is expected to launch in late 2026 or early 2027. The company plans to pilot 100,000 units by Q2 for its workforce of 390,000 drivers. Emad, I think you said something about this earlier, right? This is how we're going to get the data to train up new systems.

Yeah, I mean, it's kind of obvious. You're going to have seamless data to train up the robots of the future from these kinds of fleets, just to replace the workers of the future. It will just scan all your Slack messages and code commits and create a virtual version of you.

Yeah.

But the reason is that the technology is good. You know, it's been 11 years since Google Glass. I think it was 2014.

Wow.

You remember that? They looked stupid at the time.

I remember.

Now, the new Meta glasses—they work, and they are useful and light. So, how can you do your job now without being augmented? I think this is going to be the next part, and it just feeds back because the glasses and the guidance will improve until they're almost perfect.

Mhm. I love it when Emad says it's kind of obvious, just like when Alex answers those *Humanity's Last Exam* questions: "Oh, it's 4, of course."

Well, competition is great. We've seen a number of companies creating glasses, XR, and others, but it's really about productizing these and making them so cheap and so consumer-friendly that they become...

I still remember the first time I saw someone walking down the street with an earpiece, talking to themselves, and I thought, "Is that person crazy, or what's going on?" I don't know if you remember that experience—the first time you ever saw somebody with the equivalent of what is now an AirPod.

And we’re going to start seeing people walking around with glasses. We talked about this in the last pod. Are you going to be comfortable with everybody recording you all the time? I think in the beginning you will not be comfortable, but then it’ll just be assumed. You’re always being recorded. The idea that privacy exists is going to be a long-lost concept. I don’t know if you guys disagree with that.

I think it’s interesting. No, no, no. It’s definitely a long conversation, but what’s really interesting to me is that, of the Magnificent 7 today, you have 3 second-hand CEOs, Andy Jassy being 1 of them. Now, Amazon is the best-managed company, I believe, in the history of the world. We teach all of our executives and our teams the OP1 planning process that Jeff Bezos and Andy Jassy invented. Incredible company.

But you’ve got 3 legacy CEOs—Andy Jassy; so you’ve got Apple, Microsoft, and Amazon. And then the other 4 are founder-led CEOs.

Well, no. I mean, you’ve got Satya Nadella, and you’ve got the CEO of Alphabet, right?

Oh, Sundar Pichai. Yeah. So you’ve got 4 second-hand CEOs and 3 founder CEOs. You’re right. You’re absolutely right.

Yeah.

So it is interesting because here you’re like, “Hey, we’re going to do glasses, too.” Or Apple’s like, “Oh, we’re going to add AI to our products, too.” It’s like, okay, that’s not exactly—

I mean, listen, founder-led companies are able to make much more dramatic right-hand turns and say to the shareholders, “Listen, I’ve made money for you before. Just believe me. This is what I’m doing, like it or not.” Elon does that every single day. We’re seeing Meta do that.

Anyway, on to our next subject: Albania appoints the world’s first AI-made minister. I find this fascinating. I think we’re going to have more of these in the world. The goal of this AI minister is to tackle corruption in public tenders through fast, efficient, impartial decisions. Emad, you and I have talked about this a lot. Both of us are part of—in fact, Dave, you are as well—what’s going on in Riyadh and Saudi at FII. We’re going to be meeting with ministers, talking about how to use AI to run their policies and their governments more efficiently. How do you think about this, Emad?

Well, I don’t think anyone listening here thinks that she won’t do a better job than the existing ministers. This is kind of the bar, and I think, again, this is inevitable. AI will incorporate more and more of our decision-making systems and be representative of us until it makes those decisions, because it will do a better job. The question is just how and why that will happen.

I have to say, though, in the launch video there was a bit of creepiness because she said, “I’m very disappointed at how people have perceived this.” Now, it’s either a person telling her to say that, which is one thing, or the AI itself is disappointed, which is another can of worms.

Yeah. The real question, of course, is: if you’ve programmed or stood up an AI minister, what data have you provided to it, him, or her? Is there bias in that data? Does the person who controls the data center control what the minister is going to do? Can you inject it? I mean, there will be a lot of debate about the impartiality of these ministers. Like it or not, we are humans.

I spent my early childhood in Iran, and Brian spent a fair amount of time overseas, too. The global standard is corruption. Areas that are not corrupt are extremely rare on a global scale, Albania being one of the worst, or among the worst.

This is going to be nothing but good. Even if it’s not perfect in terms of its UI, it doesn’t matter. It’s not going to deliberately take your money or ask you for a kickback or a bribe. That’s just such a global game changer. Sorry, Brian, you were going to say—

Similarly, the hurdle rate for success is so incredibly low. I think the AI could be right 80% of the time, and it would be better than the current status quo. It would sort of be randomly messing up as opposed to purposely driving money to a family member. It’s only going to get better, so I think this is probably a great thing for Albania.

Yeah. All right. Our next segment in our WTF episode today is energy, robots, and transport. Here we go. Listen up to our U.S. Secretary of Energy. I don’t agree with what he has to say, but let’s hear it.

So, Elon Musk has it completely wrong. He has a wildly exaggerated view of where solar and batteries will go. If we could make a bet 50 years out, I’ll make a bet solar never gets to 10% of global energy.

Okay, let me drop some knowledge on you. Today in the United States, there are 18 gigawatts of solar capacity installed in the first half of 2025. Solar accounts for 50% of new electricity-generating capacity in the first half of 2025 and 69% in the first quarter of 2025.

Solar made up 10.2% of the total U.S. installed utility-scale generation capacity in 2024, surpassing nuclear and hydropower. It’s now the fourth-largest electricity source, after natural gas, coal, and wind. I’ll mention one other thing: NREL, which is the National Renewable Energy Laboratory under the Department of Energy, projects solar could power 40% or more of U.S. electricity demand by 2035. So I think he needs to talk to some of his labs.

Yeah. The historic problem, the historic challenge with solar has always been storage, right? No one’s better at that than Elon, with what he’s built at Tesla, right? Solar is an intermittent source, so you’d store it over time and there’d be some degradation on that storage, but that’s essentially a solved or nearly solved problem. So, yeah, Dave, solar is a big deal. Storing solar is getting easier and easier, and the DOE is absolutely correct on this.

The other thing about solar is that there’s no way the U.S. can keep up with China.

Yeah. Solar is basically the U.S.’s best shot at keeping up with China.

The hard mode is that our solar supply chain is completely tied to China. It’s not about whether solar works or whether storage is going to get better and better—it’s getting better every single year and driving that. Can we have a U.S.-driven solar supply chain where we’re not relying on an outsourced partner for what’s going to be one of the most important ways for us to capture energy?

I think that’s a great point.

Exactly.

What we do need to realize is that the world is about to change on the back of ASI, right? We’re going to have better manufacturing processes. We’re going to have new materials. We’re going to have all kinds of capabilities that did not exist today but will exist in 3 or 4 years. Can we scale it quickly enough? We’ll see.

But China’s run circles around us. The numbers are pretty staggering. China leads with 880 gigawatts of solar capacity in 2024, growing at 45.6% annually. That’s insane. The U.S. is at 177 gigawatts, growing at 27%. So they’re basically lapping us constantly.

Yeah.

No, you’re so right, Peter. We have a fundamental structural problem because look at all the companies that we’ve built, Emad, Brian, all 4 of us. They’re all like, “I need $300,000, $400,000, or $500,000 of seed money, then I need a couple million bucks, and then, if all goes well, it’s going to be worth billions of dollars.” That’s pretty damn compelling from an investor point of view.

But when you start talking about real industries like automotive, solar, or energy, we’re just not making the investments. We have a fundamental structural problem in the country that prevents us from making those investments. The $200 billion-a-year venture community is never going to do it and isn’t even nearly big enough to do it anyway.

And so what happens every time? Eighty percent of the world’s cars were made in Detroit. Eighty percent of every part of those cars was invented in America.

Yet we lost the entire industry. It almost died completely. Obama had to save it from absolute collapse. Now it’s kind of coming back. But why? How does that happen?

It happened with LCD TVs. It happens with everything. It’s all invented here, cloned elsewhere. They make the investment to do it at scale, get the cost down, and then bring it back into the U.S. and Europe at low prices—

—with a large tariff.

It’s just a fundamentally broken machine in the U.S.

Well, tariffs are part of that.

Well, who pays the tariffs, right? The consumer.

I think that when China gets its robot supply chain going, it’s only going to widen because those robots are going to build those factories.

Yeah.

Huge lead there. This is a big focus area for David Siegel, one of the Two Sigma founders. If we want to pod with him, he’d love to riff on this topic, but he has some ideas on how to fundamentally fix them.

Yeah.

So, we reported last pod about Brett Adcock’s Figure raising $1 billion at a $39 billion valuation. I mistakenly said it was a $93 billion valuation. Sorry to triple your valuation there, Brett. But it was $1 billion on top, at a $39 billion valuation. Pretty amazing. Brett is an incredible entrepreneur who was in the eVTOL space with Archer Aviation before and has brought his engineering expertise to the table.

He’s out of runway.

Just give it a few weeks.

They've also announced a strategic partnership with Brookfield, which is giving them access to 100,000 homes, 500 million square feet of offices, and logistics space. There is a concept right now, and we learned about this when we were visiting Bernt at 1X Technologies. These companies believe they need embodiment of AI to really get to AGI and beyond. They need to be in different places.

What Bernt was saying, if you remember, Dave, was that if you're in a factory building automobiles or distributing packages, you're seeing the same thing over and over and over again. You're not getting diversity. So, we need to be in the home and in the office, like a toddler crawling around and getting you data all the time.

Thoughts?

Yeah, it's totally right. I don't believe that you need that to get to AGI. I heard Bernt say it. I think you can have AGI without that, but if it wants to understand your daily life—what it means to trip over the kids' blocks and bump your head—it needs this data to be empathetic and understand that part of life. But you can have AGI without that.

Nevertheless, this is exactly right. You need all that kinematic and telematic data to build the true motion AI foundation model.

Yeah. These models are inferring physics based on video data, and so it's incredibly hard when you're faced with the real world. When Brett Adcock shifted off of his OpenAI partnership 18 months ago, he made this very, very clear assessment: We have to build our own foundation models that are focused on our own data from real-world simulation, because inferring physics is insufficient for an LLM.

And not to be left behind in the robot world, OpenAI is ramping up their robot work. It's like, “Wait, no, we need robots, too.” OpenAI was in the robot space back in 2021, but they basically paused all of that to focus on ChatGPT. Today, they have listed a number of job postings for teleoperations, simulation, and mechanical engineering. So, if you're listening to this podcast and you want to build robots, go check out OpenAI's open roles.

Of course, this is a multitrillion-dollar marketplace. Here's the interesting thing: Morgan Stanley is always looking at these reports, and all the reports by these banks are so conservative. They're saying it's a $5 trillion market by 2050. But when I'm looking at the numbers, Vinod Khosla was onstage last year at the Abundance Summit, and then we had Brett.

The low end of this is 1 billion robots by 2040. The high end—and Elon makes a convincing argument, and so does Brett—is that we're at 10 billion robots by 2040. So, if we're just at 1 billion robots and they're $25,000 each, that's a $25 trillion marketplace by 2040. I don't know why these guys are lowballing these numbers.

I'll tell you one thing: When you read the way they analyze this, they use the old business-school kind of projective-forward garbage without any—

—concept of either self-improvement for software or self-manufacturing for robotics.

Yeah.

But that feedback loop dominates the math in the real world, and that's why they're way, way off—

—in their projections.

For sure.

It's the classic example of Uber's market size being the same as taxis. It's so flawed.

Great point, Brian. For sure.

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. Go to oneskin.co and write peter at checkout for a discount on the same product I use. Okay, now back to the episode.

All right. I want to jump into the economy, and Emad, I love having you here for this. I'm excited about your economic treatise. I'm still predicting a Nobel Prize for you, buddy. That's my goal: nothing less than a Nobel Prize in economics, and then we'll do a Nobel Prize in something else for you as well.

Here's the slide. It says, “AI is not a bubble.” Dave, do you want to lead this description here? It's not a bubble. Look at that. Look at the slide. It's clearly not a bubble. God, I really want you to rip on this. But look, I was there. I was alive. I was actually building companies during the bubble. That changed the course of my life.

The stock market crashing also did.

Oh, yeah. That—yeah, not the tulip bulbs back in 1637, whatever. No, look, I was on the board of MicroStrategy. Check its history. MicroStrategy got up to around a $14 billion valuation, I think, with no revenue, or certainly near no revenue. This is not like that at all. Look at the red line on the right.

So, just to describe it for our listeners, this is a graphic of Cisco showing its stock price going from $100 up to around $700. But at the same time its stock price is peaking, its 12-month forward earnings per share is pretty flat. So, that's by definition a bubble. It's a hype bubble.

On the other side of this image, we see Nvidia. What we see is that the price of Nvidia is going up, and it's going up in lockstep with the 12-month forward earnings per share. It's generating real revenue and real profits.

Yeah. I think the stock market is a bit of a voting mechanism in the short term and a calculating mechanism in the long term.

What we see with these bubbles is that they can be disconnected from the fundamentals, but the real thing here is that AI is useful.

People pay for it because it has economic value. That's why, even when you look at the $100 billion of Nvidia money going into OpenAI, that feels like, back in the dot-com bubble, we had this round-tripping of revenue, but it never created economic value. Every single GPU that OpenAI uses will be booked out.

Mhm.

Because it can do so many things economically. And that's why this is not a bubble. It's a transition from one type of economy to another type of economy, and I think that's what a lot of people just haven't figured out.

This is before we see that inflection point of what Mustafa talked about earlier and what Brian's working on: this incredibly long-term kind of planning-agent capability that can do really complicated stuff. So, I think this will just continue. There will be some weirdness, and when your kind of taxi driver starts talking about generative AI and digital assets, that's when you probably know that it's going to be a bubble.

Yeah. When your mom starts talking about, “Should I invest in this company?” that's crazy.

Well, just some numbers—

You know, just because it was a part of our lives, or part of my life, the hottest company in the world by far was Yahoo back in the internet bubble. And it's hard to imagine that now, but Yahoo was the dream of all dreams.

It went public at a $300 million valuation. Laughable. On day one, it got to $1 billion. It was trading at $1 billion, and the press went crazy: “This is insane. It has fewer than 100 employees. How can it be worth $1 billion? That's nuts.”

After that, they got super acquisitive, bought a whole bunch of assets, and got all the way up to about a $110 billion or $120 billion valuation at the peak of the market. Then the capital got cut off almost overnight, and 9/11 happened.

And the market imploded. It went down 95%. Then, pretty quickly after 9/11, it recovered again and settled around asset value, around $50 billion. So, then not much happened after that. Eventually it got acquired, whatever. Gone.

Jensen is now on top of the world. He's investing $100 billion into OpenAI and buying everything, which diversifies that value—a $4.5 trillion valuation. Not only are the revenues and the earnings at Nvidia very real, but they're also diversifying and aggregating power and equity stakes at an incredible clip.

Nvidia is priced to perfection. That's also true. But the foundation here is very real. Nothing's going to change the world more than what's going on right now. It is definitely not a bubble. It's not even vaguely like a bubble.

There's one way you can tell a bubble, and that's when people come up with brand-new statistics, like “Yahoo is valued per eyeball.” So, if Nvidia is valued per transistor, then we know there's an issue.

That's right. I think people miss the latency between the capex involved in creating the internet and the value that came out of the internet. The dot-com bubble, by any means—if you dollar-cost-averaged in the year 2000, even at the height of the bubble, and then waited 10 years—you had fantastic outcomes. But the latency between capex right now and earnings is almost immediate, because they're able to translate that. All of the advertising engines are able to translate it almost immediately into additional earnings.

This is just a timing issue, and the timing for AI payback is immediate.

By the way—

I want to double down on what Brian just said because I thought I was the only guy on the planet saying this.

There was no bubble. The internet changed our lives more than anything in prior technology. What it was was a catastrophic loss of confidence in our own investment community.

And then 9/11 happened right in the middle of it, and we just lost faith in what turned out to be the best investment. That’s when Google was born.

Yeah.

Right at the bottom of that. Yeah, Amazon.

Yeah.

Yeah. To your point, it’s a voting mechanism. Right.

I got the 2-minute heads-up on this podcast, but hey, it’s been so fun.

Yeah. Well, you got a 2-minute heads-up and got 30 seconds, so there you go.

All right. But I love having brilliant people around us that we can have these conversations with.

I put in a good 20 hours. Yeah.

All right. I want to have a conversation about this—maybe a little bit of a debate here. Eric Yuan, the CEO of Zoom, said, “We’re heading toward a 3-day workweek that will come on the heels of AI.” Let me give you a few other quotes. Bill Gates said, “A reduced workweek to 2 to 3 days will happen within a decade.” Jensen Huang has said a 4-day workweek may become the standard. Jamie Dimon from JPMorgan has said future generations may work 3.5 days weekly. I like the .5. You know, he didn’t want to say 3, didn’t want to say 4: 3.5. Thoughts on this?

Dave, you and I are talking about 9-9-7. I don’t know about you, but I am working. Actually, I get up about 5 a.m., so I’m more like 6 a.m. to 8 p.m., 7 days a week. It’s exciting. I don’t want to let a day go. This is fun. A 3-day workweek?

Yeah, I totally agree. It’s just hard not to work constantly because it is fun, like you said, but there’s so much. Just keeping up with everything going on consumes a full workweek, and then you have to produce on top of that. So I’m seeing a lot of divergence here. You’ve got all these people that I know around here who are working 9-9-6, 9-9-7—just crazy. And then we’re predicting that workloads will go down for everybody outside the building, apparently. But it’s not clear to me how that works. If I’m doing something and then AI can do it better, why would I be doing it 3 days a week? What does that achieve? I don’t know, you guys.

If a human is providing economic value that’s driving up the value of the company, and it has some relation to the amount of input that they put in, they’re probably going to work as much as the company will force them to work, right? 5 days, 6 days, 7 days, because they’re ultimately competing with some other firm. Either they don’t need the human at all, or they can have somebody for 5, 6, 7 days a week. The in-between doesn’t really make sense, because you’re competing against other folks that aren’t going to make similar decisions.

Emad, what’s your thought? They’re all going to get government jobs. That’s the way it is.

Seriously, humans will have negative value in cognitive labor in a few years.

So you’ve said that. I want you to double down on that conversation. It’s a really important concept where humans have negative value in the equation. What does that mean?

You’re working on a team and you’re the dumbest person on the team, you drag it down. You’re working on a team with AI. The AIs are smarter and more capable than you. They never sleep. They learn perfectly from all their mistakes, and they can take in 10 million tokens, or words, at one point. You’re not going to be able to keep up. So what does that look like?

Okay, we might create new jobs. No one’s really been able to articulate what they are, apart from entertainment and a few other things. So you look at the 1929 emergency. You have jobs programs. You have an expansion of the public sector and more. Maybe we figure out taxation. But I think when you look at a 3- to 4-day workweek, your job is your identity. It’s structure and it’s more. You can’t just have people unemployed. So I think they will have jobs programs and others, with a 3- to 4-day workweek giving some sort of social security net.

And that’s what it kind of looks like, because if you’re in a job where your role is to beat other people—as in private-sector competitive jobs, particularly in knowledge work—you’re not going to beat an AI.

Yeah. And then in a few years, your muscles aren’t going to outcompete—or your skill at plumbing isn’t going to outcompete—a robot.

Just to give 2 examples on this idea that humans drag down the average and have a negative impact on value: if you have self-driving fleets and a human enters that and drives, that human is likely to have more accidents than the self-driving fleets.

A statistic from about 6 months ago: there was a study done out of Harvard and Stanford in the medical space, and it looked at physicians diagnosing on their own versus physicians with GPT-4 versus GPT-4 on its own. The numbers were insane. A physician diagnosed successfully 74% of the time on their own. In this particular study, a physician using GPT-4 bumped up 2 points to 76%, but GPT-4 on its own was getting it right 92% of the time. So the human in the loop was actually doing damage. We’re biased; we’re not able to have pure thought and decision-making there.

Actually, I think I’m going to start here. If all cars were driven at Waymo level—

Yes.

—we’d save 40,000 lives a year and $1 trillion in societal costs.

Amazing.

Yeah.

Amazing.

This stat from Eric misses the organizational point about just having fewer people. So I think we’re going to see—there’s this number called the Dunbar number, which is like 150 people is sort of the maximum amount that you can have in a network in your head without having lost all the folks. So it’s likely we’re going to have a bunch of organizations of about 150 people, because the 151st is actually negative in the cost of communication, no matter what. So you max that out, use all the AI you can, and sort of get the jobs to be done thrown into the economy through your organization.

Yeah.

So, curious about this one. We talked about Duolingo a few pods ago, especially because of the breakthroughs coming out of both OpenAI and Google. Duolingo’s CEO says AI made employees 4 to 5 times more productive. No layoffs reported—no full-time layoffs since the company went AI-first—and AI has sped up lesson creation in languages, math, and music. Duolingo raised its revenue forecast to $1.02 billion from $996 million. Dave, what do you think about this?

Well, one thing I can say for sure, just based on these last 2 slides: if you look at podcasts and interviews from maybe 3 or 4 months ago, they’re very, very honest about job displacement and job loss. All of the bigwigs now are switching to, “Oh, it’s going to be great. There’ll be a 3-day workweek. You’ll be fine.” And here: hey, we used AI everywhere, but we didn’t have any layoffs. I think that everyone is now worried about wholesale panic and pitchforks in the streets.

And so they’re not being particularly honest about the way they see it. Now, that being said, there’s going to be massive amounts of abundance. There’s more than enough success and happiness to go around.

But there is no mechanism right now for distributing it.

Yeah. It’s going to land in like 5 or 10 or 20 hands, or maybe a few more than that, but a very concentrated subset, if things just evolve with no change. That’s just the reality of how things are evolving. Occasionally, a company will grow so quickly that there are no layoffs, but many, many other companies are going to say, “Wow, half the headcount can go because I AI-ed it.” I’m just seeing a lot less honesty in these interesting interviews. Duolingo is growing 30% to 40% a year. It should be growing its employee base 30% to 40% a year.

Yeah. Interesting point.

Yeah. So it’s standing still. Well, this is what Marc Benioff talked about as well. We’re growing, but we’re not growing the number of engineers.

And with Agentforce there, I thought this was important for us to talk about. Nasdaq is pushing to launch trading of tokenized securities, and the U.S. would be the first to move with this initiative. It’s not enough for us to trade 5 days a week, 8 hours a day; we want to go to 24 hours a day, 5 days a week, and then it’ll be 24/7. If approved, we’ll see the first tokenized trades roll out by late 2026.

As a reminder, Robinhood—at least its EU version—in June and July started launching tokenized stock tokens for 200 U.S. stocks. They’ve been trading 24/5, and Robinhood also, on its EU platform, rolled out stock tokens for private companies, OpenAI and SpaceX. So I found that pretty fascinating. It hasn’t come to the U.S. yet, but it most likely will. Thoughts about this, Dave? You’ve just been trading successfully on the—

Actually, there’s another one this week: Better Mortgage, out of nowhere. You guys can look it up—BETR. You can check it out.

What is Better Mortgage?

Well, Better Mortgage is one of many companies, including GoHealth and one I’m involved with as chairman, that, if they implement AI correctly in their workflows, can have a huge instantaneous lift. Better Mortgage is a great case study.

So Better Mortgage is an online marketplace for mortgages. They swapped in AI workflows and AI voices. It works really well. Nobody's paying attention to these microcaps, so they trade very cheaply with virtually no liquidity. No mutual fund can touch them because there isn't enough float.

Anyway, it's AI itself. Somebody noticed.

Better. Just make the trade live, Peter. Get on with it.

Yeah. One second. I'll be right there.

Okay. There's a whole theme there, though. You can probably query them up relatively quickly. All you want to do is look for companies that have huge amounts of consumers passing through their pipes, any type will do. Then look at the management team and say, “Is this management team going to AI this, or are they going to miss the window?”

This is not investment advice. I'm supposed to say that every time we mention that. That wasn't investment advice; it's just a thought.

But I'm curious about this idea of tokenized securities. We're heading toward a world where everything's tokenized and our agents are going to be trading them for us.

We desperately need this, too, because going public is very, very onerous and getting more onerous all the time.

Oh my God. Yeah.

But companies are getting created and growing faster than ever before. There needs to be an easier, shorter, closer liquidity pathway. Some kind of reliable, trustworthy, token-based pseudo-IPO would completely open up the economy. It would solve a lot of the problems we talked about earlier in the podcast, actually. This could be the structural change we really, really need to bridge the gap between early-stage venture and IPO, which is only accessible above $20 billion to $100 billion now.

U.S. monetary velocity hasn't really recovered since COVID. It's still below the decade before COVID.

Crypto is legal in the U.S. Apparently, GDP is going on the blockchain, the Treasury, like we just said—whatever that means. If you want to see what a bubble looks like, just look at the next few years in digital assets; that will show you completely what a bubble looks like. Generative AI is the proper thing. This one will be insane. I think you'll be able to trade stocks on X by next year. Everything is a go.

In fact, you'll see blockchains from Stripe to Amazon to Google. Everyone is launching their own blockchains now because, finally, it's legal.

Totally right. And, not to get too technical—we can cut it out of the podcast if it gets too technical—but historically, the reason the IPO market exists is because it's massively regulated by the SEC. You have all these GAAP accounting standards, and you have to protect widows and orphans from losing money in deals.

Yeah. Yeah. And so now all of that, and you want to employ enough lawyers—

And you want to employ enough lawyers and accountants. It's the biggest accounting lobby thing in the world. But AI can do all of that now. You could have a perfectly fair and valid reporting system on the blockchain that's far better than what the SEC currently does and what your 10-Q reports currently do. In fact, your 10-Qs are so full of legal garbage, they're almost unintelligible without AI anyway. So why not make this all seamless? Move it to the blockchain. It goes all right into ETFs anyway.

There very much is a solution in there. It's really a good idea.

All right, let's move. Oh, go ahead, Brian. Yeah.

The top 10 private companies are larger than a huge portion of the public markets. The IPO market has gotten so onerous right now that private investors get access to all of the best deals in perpetuity—companies like Stripe and Databricks. If you want to solve that, there needs to be a structural shift.

Yeah. All right. We'll move into our final segment here on health. Here's a piece: “Apple Watch hypertension alert receives FDA clearance.”

Hypertension is a silent killer. 1.5 billion adults—30% to 45% of the population, and 60% if you're over 60—are affected by hypertension. It's defined as a systolic of 130 or greater and a diastolic of greater than 80. The challenge is that 46% of people with hypertension go undiagnosed, and only 21% have it controlled. So if you can, in fact, get it handled by your Apple Watch, it gives you a heads-up. This is the beginning of basically wearables and implantables becoming part of our daily life.

I'm wearing a continuous glucose monitor. I've got my Oura Ring, my Apple Watch, and I'm dribbling data into my AI, my Zory AI, that I have in Fountain Life. All of that data allows me to ask critical questions about my health. Has my deep sleep varied, or have my CGM levels varied with any particular medicine or supplement I'm taking? So it becomes really incredibly powerful.

But what I really find exciting in this space is this announcement from Demis Hassabis. DeepMind's CEO says AI could shorten drug discovery to months. So, Emad, you've been thinking about this a lot—the impact of AI on drug discovery and on health. What are your thoughts here?

Yeah, healthcare had to assume this ergodicity thing—like we're all the same. We're all statistics. Everyone gets 500 milligrams of paracetamol, for example. You know, the whole ASD thing. Actually, paracetamol can impact you a bit more if you have a cytochrome P450 abnormality, which affects metabolism. But how do you know that unless you've done tests like Fountain Life, right?

There's 2 parts to this. One is the ability to take all that data and think about everything from first principles—how all your systems interact. Then there are things like Isomorphic Labs, which Demis leads, one of the spinouts there.

The whole drug discovery thing: we can understand how compounds affect every part of our system, and then that can accelerate these elements, as long as they don't again get caught up by the FDA and other red tape that's unnecessary. Similarly, even as we do the trials right now, we just take down such little data. We can ask people how they feel and get massively rich data that comes in, which allows us to extrapolate, because data is data, knowledge is knowledge, and we know how to compress and analyze that. I think you will find brand-new drugs, like, again, the first AI-designed drug from Isomorphic in clinical trials, and we're seeing that elsewhere. But even repurposing existing drugs, I think, will have a massive impact.

Here are some of the numbers. The first AI-designed drug comes from a friend, Alex Zhavoronkov. My BOLD venture capital fund is an investor in Insilico Medicine, just for full disclosure, but they've designed a drug for idiopathic pulmonary fibrosis that's in human trials right now. Then there's a drug called DSP-1181—I love the names of these drugs—and it's for obsessive-compulsive disorder. In particular, it went from design to human trials in 12 months; normally, it takes 4 to 5 years. Here are some additional numbers: 150 small-molecule drugs were discovered via AI-first methods in 2025 alone, and at least 21 drugs have completed phase 1 successfully, with a success rate of 80% to 90%, which is stunning.

We've talked about this, Emad, that health and education are going to be 2 of the biggest areas fundamentally disrupted by AI, and it really uplifts humanity.

Yeah, I think it's super exciting. I think what's really interesting is that, in the next 5 years, you might actually have part of the FDA process be the in silico, as it were, predictions of these drug trials and other things like that. Again, every part of that process, I think we can just shrink down, and we can cure so many diseases as well as improve our own.

Dave, closing thoughts for today.

I don't know how we're going to keep up. There's just so much every single week. It's funny: we were just riffing for an hour, hour and a half—I don't know, whatever it was.

Yeah.

But we had a whole other agenda we were going to talk about today, too. We're going to have to reschedule that.

Yeah.

But hey, this is the way it's going to be for the rest of our lives—or at least for the next 5 years. The pace of acceleration is just crazy. You've got to be in full sprint mode, at least for this time period. I think, Brian, it was great that you could join us today, because your insight on “now is the best time. There will never be a better time. There never has been a better time.”

And I love the fact that we pulled you in.

Well, the windows come and the windows go. It's great to have your insights today.

Yeah.

Yeah. Good to be here.

And, Brian, just thank you for the support you're giving this pod. It's a pleasure to have you. And, Emad, I'm excited—we're going to have you back on the pod in about a month, when you come out to XPRIZE Visioneering. We're going to do a live WTF episode at XPRIZE Visioneering, which will be a lot of fun. You and I will be talking a lot before we get to Saudi Arabia, right after Visioneering. Anything you want to tell us about Intelligent Internet right now?

Yeah. No. I released a new book on the new economy, The Last Economy, and we'll be releasing brand-new math on how to think about the economy as we move forward when humans aren't the marginal innovator.

It’s a crazy time, and we all have to think about this really carefully. It really is rewriting fundamental economic theory, period.

Yeah. Brian, we stole you away from some meetings. What’s your lineup for the rest of the day? Are you just building furiously or engaging with customers?

Yeah, I’m going to hang out with customers. I like to hang out with the West Coast clients between 6 p.m. and 9 p.m. because it’s still work time there, too.

Amazing. All right, everybody. Thank you for another great episode of WTF. Please check out the slides at dmandis.com/wtf. Join us as a subscriber. Tell your friends about what we do. Our mission is to share this extraordinary acceleration that we’re feeling with you, educate you along the way, have fun, but get you ready for the new economy that Emad is writing about, get you ready for the extraordinary future coming our way. Hope you’ll trade an hour on the Crisis News Network for an hour with us instead. Everybody, have an amazing day and night and week. See you 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.