超级智能竞赛将决定我们这个物种的未来｜Salim Ismail & Dave Blundin｜第179期

We are going to become limited by power in our quest for digital superintelligence.

This is a structural issue we have in the US.

This is America's Achilles' heel.

China is going all in on energy production, and it's epic. Meta is worth $1.8 trillion today. They have $70 billion in cash. These offers of $100 million, or their acquisition offers on companies, reflect a winner-take-all mindset.

The natural dynamic is winner takes all because AI becomes self-improving very, very soon.

The biggest threat to Meta is that they fall way behind on AI.

We're in the middle of probably the greatest drama in human history here, which is why everyone should be tracking these moves closely. These numbers are so unprecedented, but they're completely justifiable given the impact. Now, that's a moonshot, ladies and gentlemen.

I'm here with my moonshot mates, Salim Ismail. Salim, I'm calling you the Emperor of Exponentials, because that's just who you are.

Oh, Dave.

And Dave Blundin, the Alchemist of AI.

Mm.

How's that for a title?

Okay.

You guys at work yet?

We need one for you, Peter.

All right, Peter, you have to be something epic. How about the Humongous Bungalungus of Abundance?

I'm not sure I like that. I like what Dean Kamen said on stage. He called me—

What was it? He—

—the Pope of Hope.

The Pope of Hope.

That was much—

That's awesome.

—Pope. Yeah, my mom liked that. She's watching all of my abundance videos and the whole show on stage, and she writes back. She goes, "Pope of Hope, I love that." Thank you, Dean Kamen.

All right. That's it.

We have a lot to cover today, and as always, our goal here is to deliver you the real news—the news that's going to impact you, the news that's changing every industry, every family, every country, right here, right now. So rather than watching the 6 o'clock news, join us on this epic mission to deliver a compelling, hopeful, and abundant future.

Here we go.

1. AI Rewrites Human Knowledge

Buddies, let's jump in. Let's talk about all things AI. Another epic week. Every week is accelerating; it feels that way. Let's kick it off with this conversation. This is from Elon, and I'll just read his tweet. He says, basically, "We will use Grok 3.5—maybe we should be calling it Grok 4—which has advanced reasoning to rewrite the entire corpus of human knowledge, adding missing information and deleting errors, and then retraining on that new corpus."

First of all, we've got this name escalation, right? We're going to have GPT-5 and then Gemini 2.5, soon Gemini 3. I think Elon feels behind on Grok 3.5.

Numerical warfare?

Yeah.

Dave, what do you think about this idea of retraining Grok 3.5 on a new corpus corrected by AI and getting rid of human errors?

That specific idea is actually low-hanging fruit and a real big win, but it's one of many big wins. The rate of change we've seen in the models in the last 2 weeks since the last time we talked is mind-boggling. An experience everyone needs to have is to pick up either a Gemini 2.5 Pro or GPT-4 voice mode and just talk to it as you're driving in a car for an hour or 2.

Yeah.

That's something you couldn't do a month ago, and now you can do it and it's engaging. If you draw a line between a month ago and today and look a month into the future, it's going to replace a lot of what you do in terms of media. It's just so incredibly engaging all of a sudden.

I think what Elon is talking about here is, look, the training data—believe it or not, the actual original training data for these models had a Reddit subreddit in it called Microwave, and the Microwave subreddit has a series of Ms. It just goes, "Mm," for thousands and thousands of lines, and then at the end it goes, "Beep." So that gets scraped and thrown into the training data.

A lot of crap.

So, okay, there's a lot of crap. But I think Elon might actually be referring to a lot of his own tweets with Donald Trump here. If we get rid of that crap, then the neural net has a much easier time learning what really matters.

This is part of a long list of low-hanging fruit that's right in front of these training algorithms. You're going to see really rapid improvement just from the obvious, including this.

Salim, what could possibly go wrong if AI rewrites the corpus of human knowledge?

I've been watching a few videos of Yuval Harari talking about, "Oh, my God, AI can now program itself and program things." He's going to go nuts on this type of concept, because if you can edit history, where do you end up? Where do you draw the line? Who decides what's accurate or not? We know—

In the beginning, there was AI, and it was good.

Exactly. Exactly. God said, "Let there be AI," and then everything followed from there, right?

Yeah.

This really poses some huge philosophical challenges. On the plus side, there are so many gaps and so many flawed narratives, where history is written by the winners of all the epic battles and wars in the past. Therefore, we can balance out that viewpoint a little bit and get a little more reality into it. That would be great. But there's a very dangerous line here, and I think it's the right thing to do. It's going to cause a lot of consternation.

We're expecting Grok 3.5 any day now. It was promised in May, then delayed into early June. What is today? We're recording this on June 26, so he said by the end of June. He's got 4 days left. But even if it's July, it's going to be epic. I can't wait to try and play with it.

2. The ASI Timeline Accelerates

I'm going to play this video also from Elon, and the subtext here is, "Superintelligence may happen this year or by the end of next year." All right, let's listen.

I think we're quite close to digital superintelligence. It may happen this year. If it doesn't happen this year, next year for sure. Digital superintelligence defined as smarter than any human at anything.

Here we've got the issue of definition, right? What is AGI? What is digital superintelligence? Dave, you and I recorded an episode we'll be sharing shortly with Eric Schmidt, going deep into digital superintelligence. His prediction is a little bit more—I won't say pessimistic, but it's the next 5 years on his timeframe.

Yeah.

What is it? Do you still have the confusion that I do when people are popping back and forth between AGI and ASI?

Yeah. No one's really locked down a clear definition, but I think Elon gave a very, very clear definition in that presentation, just for that reason: AI that can do anything better—any intellectual task better than any human. That's the hardcore definition, and he's saying by the end of next year, which is the soonest date that people are saying. But he's very close to the progress, so he has every reason to be right.

Yeah, I would not doubt his timeline. I actually went onto ChatGPT and Grok and asked them both for a definition of AGI versus ASI. Can I share that with you guys?

Mm-hmm.

Yes.

ChatGPT says AGI is a machine capable of understanding, learning, and performing any intellectual task that a human can do across domains, with reasoning, adaptability, and autonomy. Then it says, "ASI"—this is ChatGPT—"is an intelligence far surpassing the best human in every field, creating with creativity, problem-solving, and decision-making."

Grok says AGI is an AI capable of performing any intellectual task that a human can do with general problem-solving, and ASI is AI surpassing human intelligence in all intellectual tasks. Even these definitions sort of blur the line.

I go on my classic hobbyhorse here, because we have no idea what we're talking about when we talk about intelligence. I don't need to get into that trope again. But let me suggest this: The minute you can prescriptively describe a task, an AI or robot is going to be much better than you anyway.

If you want to define something like this, you have to define what you mean by it. The work then comes down to what the task is, prescriptively. Saying that it'll be smarter than a human being—that's a different kind of model. Here's where I'd like to see it do something.

Please.

You have Kepler, a couple hundred years ago, one day making an intuitive leap that maybe the Moon is affecting the tides—a massive intuitive leap that can then be backed up with scientific and experimental evidence. That's the kind of thing that, if AIs can do, you start tickling at the edges of what we mean by intelligence, because we have emotional intelligence, spiritual intelligence, and so on. There's the end of my rant.

I think that's going to happen. One of the predictions is that we're going to start to see math, physics, chemistry, and biology getting solved by these advanced AI models in the next 2 to 5 years.

This is what our friend Alex Wissner-Gross keeps on hitting on. We're gonna solve math—

Yeah.

—in the next 12 months.

Yeah, I think it's important to stay out of the philosophical debate if you want to succeed with AI and focus on the capabilities within swim lanes. The reason Elon Musk is saying, “Look, guys, I'm talking about AI that can do literally any human intellectual task better than any human,” is that he's trying to create awareness and motion, because people are underreacting so badly in so many areas.

But as Alex Wissner-Gross is saying, it gets miles ahead in areas like math and code writing, where it's not data-constrained—

Right.

—And it lags behind in areas like biology, where it needs the full cell simulator to make forward progress. The rate of progress is gonna be hugely different in these different swim lanes. The exact date when it can do any intellectual task better than any other human is gonna be like a blur that comes and goes.

Whether somebody was right down to the minute or not, nobody will care a year later. But we'll care a lot about the impact on society and all these different use cases.

Yeah.

So I think going down that path of saying, “Here's a vector, a swim lane,” is a really good way of framing it. But when you throw out general words like AGI or ASI or whatever, that's when I go a little bit nuts. The swim lane thing, I can totally vibe with that.

3. The First ASI Becomes the Last

All right, I love this video. I asked the team to cut it. This is from Geoff Clune, who's a DeepMind advisor. The title here is “The first ASI may be the last ASI.” So take a listen to this.

It is a world in which the first AI is the last AI.

And the creation of the first ASI suppresses the creation of ASI worldwide. Then that organization, whoever they are, has a decision to make. And that decision is: “We just effectively invented a god.” If that thing is aligned to them and will obey their commands, do we want to sit around and let those people over there also invent a god?

What nobody talks about as much as they probably should is how quickly things might get nationalized. If you are the premier or prime minister or head of state of a country, and a company within your borders creates a superweapon, a superpower, effectively a god, do you nationalize that? Do you start giving them orders? Do you make them run everything by you? Are you gonna let them just run as a normal company? That seems very unlikely to me.

So—

Yeah.

—just wow, right? I can very much imagine that. In fact, I just read a book with my son Jet called After On. Actually, it's the second time I've read it, and it tells a Silicon Valley story of the first ASI coming online. It is basically taken over by the government, and it is basically the last ASI because it suppresses other AIs around the world. It's a great story. Dave, what do you think about that?

Yeah. Well, hey, Salim, we have our side bet, and this really weighs in my favor. Look, the natural dynamic here is winner-take-all, and you're gonna see later in this podcast the amount of competitive pressure on these foundation model companies to get the best talent. The amount they're willing to pay is mind-blowing.

Why is that? Because the natural dynamic is winner-take-all because the AI becomes self-improving very, very soon. So the observation in that video is right on target. In our first slide here, we're saying, look, if the data that goes into these is one view, one point of view, and it's self-fixing, but it will filter out other points of view, that could be terrible.

And so the only way you're gonna have a variety of these—and America thrives on a variety of competitors in any given market, a variety of viewpoints—that has to come through some kind of regulatory framework. It's not gonna happen with the natural winner-take-all dynamic. This is a great wake-up-call video, and I completely agree.

Salim—

So—

—winner-take-all?

I'm not sure about winner-take-all. I'll stay with my bet on that one.

Okay.

So Dave and I will continue. It'd be great to have a Polymarket on this, by the way. But the idea that when something like this emerges, it might get nationalized is 100% true. There's no way that's not going to happen. I think this is what the governments are doing right now: They're just watching their various folks work on stuff, and they're gonna jump down their throats the minute something like this emerges.

Right.

And nationalization may take a different flavor. It may be the government buying a significant share. We're gonna talk about what the government is doing in AI in a moment.

Oh, I don't think it's gonna be like that. I think it'll be just like, “I'm sorry, we own that. It's a military threat, potentially.” Boom. And you've lost agency in that.

There goes the U.S.

Yeah. Well, I think that's gonna happen.

Mm.

From what I've seen with governments, there's no way they're not gonna do that.

Mm.

They almost have to do it in order to prevent other people from getting there if they think they're getting there first. The bigger picture might be: What do you do with that if you have a particular worldview and you have ASI? Because I have a feeling that ASI is gonna strip past the limitations of a particular worldview very quickly.

Mm.

And so then what do you do? Right?

Yeah. All right. Well, I have the next story—

Wait, let me just build on that for a second.

Please.

Here's what I think will happen.

Okay.

Some ASI, called whatever we want to call it, will emerge. A national government—call it Kazakhstan—will go, “We need to own that.” Right? “This is our worldview. ASI, please operate on this worldview, and then let's get everybody else to align with this worldview.”

And the AI, in about 3 seconds, is gonna go, “Their worldview is so limited,” right? “That this doesn't help at all.” And it skips right past all of that. I think once you have ASI, it's outside the potential for control for anybody.

For sure. Maybe this is like the modality of religions taking over and setting a worldview around the world. Does that make sense?

Yes. You have to relate to it in that way, except that religions are based on absolute unverifiable truths or assumptive truths, like Mary was a virgin, or Muhammad was the last prophet, or Jesus was the son of God, or whatever. And an AI—any kind of AI half worth its salt—would skip past that assumptive truth instantly and go, “There's no evidentiary basis for that.”

Yeah.

I would argue that I don't think there's much of any chance that the U.S. is gonna nationalize a single AI company and say, “This is our national AI.” I think that if you look at the way the Defense Department works, some things like uranium and plutonium refinement are nationalized, but all the missiles, inertial guidance, and defense systems are private-sector companies that work for the government.

That's gonna be the likely outcome in AI as well in the U.S. Maybe not in China. Maybe not in the Middle East. But certainly in the U.S.

Well, we're gonna find out in the next few years. I think that's the key point here.

By the end of next year, clearly. By the way, Jeff looks more like an AI in that than in any video I've seen in a long time.

All right. This next story is one I wanna dive into. The title here is “Meta tried to buy Ilya Sutskever's $32 billion AI startup and is now planning to hire its CEO.” We'll get into this in a moment, but I just wanna pass a theory by you. How does Ilya get a $32 billion valuation? So he basically goes out and pitches Andreessen Horowitz. His investors are Andreessen Horowitz, Sequoia, DST Global—which is Yuri Milner—Alphabet, Nvidia, and Lightspeed Venture Partners.

The triple-A list of investors, and he raises—what was it?—$6 billion of capital on a $32 billion valuation.

Mm-hmm.

How do you do that without any product or any tech to show? I have a theory. Here’s my theory. You ready?

Okay, yeah.

We just saw the presentation on the first ASI, the last ASI. He goes in and says to these venture funds, “Listen, I know how to build an ASI that will blow away the other AI companies. It will be a safe ASI, because here’s my strategy: It will be the first, so it will be the last.” You believe him, and as a venture fund, you have no other choice than to invest in that company at whatever valuation he offers you. How do you think about that?

I think that’s exactly right, and I think there’s another point, which is that clearly the truly great neural architecture people—the Ilyas, the Mira Muratis—are not intimidated by the progress that’s been made at OpenAI, Groq, and Google.

Mm-hmm.

That’s an amazing fact by itself. When you look under the covers, the research teams working on this are 10 or 15 people. They’re not 10,000 people. The innovations are still piling up, and there are still 10X improvements out there. Undoubtedly, Ilya, having been an architect right in the middle of this, is saying, “Look, I know how to 10X this, and I’m not afraid of the big guys.”

The actions at OpenAI are reinforcing that. OpenAI is racing to control the consumer experience by buying Windsurf for coding, having voice mode, and trying to get everyone. They’re trying to be like Google and have a huge user base installed.

And they’re succeeding, by the way.

And succeeding wildly.

Yeah.

Because just competing as a foundation model is not necessarily defensible. It’s not just Ilya; it’s Mira, and then there are also some other things bubbling up out of MIT that are getting huge valuations because they’re very likely to work.

We’re in the middle of probably the greatest drama in human history here, which is why everyone should be tracking these moves closely. These numbers are so unprecedented, so much bigger than anything in history.

Insane.

But they’re completely justifiable given the impact. More people should be getting involved, reacting, contributing, and not being intimidated, because Mira’s not intimidated, Ilya’s not intimidated, and the investors coming in to invest in Ilya are not intimidated.

A billion dollars a day. Salim, do you remember the meme that came out when Ilya left OpenAI, or staged the revolt? The meme was, “What did Ilya see?” Do you remember that? Now I want to know: What did Ilya pitch?

Yeah.

That’s the new meme here.

Yeah.

Salim, what do you think about this, a $32 billion valuation? Does he have an ASI in the bag, and is he racing out in front?

I think this conversation kind of nails it, right? If you’re in front of investors, they don’t know. They’re kind of trusting you to know, and the fact that he has a confidence-based approach to saying, “We can beat the other models,” is huge.

I think Dave’s assessment is right. OpenAI is now focused on getting the biggest consumer share in this, and they will go after those other white spaces that are there. There’s a lot of white space.

My big question is: How do you do this safely? I’d love to understand what he said to investors that gave them the sense that this could be done safely, because that’s the foundation of his approach, right?

Yeah.

“We’re going to make AGI that’s safe,” and I’m curious as to—

And that’s the name of his company.

I’m curious how he’s going about that.

Safe Superintelligence. Right.

Yeah.

All right, the other side of the story here is that Meta is trying to buy talent left, right, and center. Let’s take a listen to this video.

All right, OpenAI CEO Sam Altman has some strong words for Mark Zuckerberg on a new podcast, criticizing Meta’s recruitment methods and even its level of innovation. Deirdre Bosa with more.

It’s cutthroat out there, Brian. You’re right. Critical words may be an understatement. Sam Altman, on his brother’s podcast, says that Zuckerberg is offering $100 million sign-on bonuses to poach top OpenAI talent.

Keep in mind, those kinds of bonuses don’t have cliffs; they don’t vest over a number of years. $100 million just to get on board. Nothing’s stopping talent from leaving in what is already a revolving door of talent in AI.

Salim, did you get an offer of $100 million from Zuck yet?

No, but can I please be an intern at one of those companies? Maybe I’ll get a $20 million signing bonus.

It’s insane, right? Here are the numbers, just to put this in context. Meta is worth $1.8 trillion today. They have $70 billion in cash.

If you think of it that way, these offers of $100 million, or their acquisition offers on companies—because Meta tried to buy SSI first—it makes sense. The biggest threat to Meta is that they fall way behind on AI. Dave, what’s your calculus here?

There’s so much. First of all, I don’t believe for a minute that it’s $100 million to join, with no vesting. You can’t just join and quit the next day. There’s no way that’s true. I don’t know where that fact came from.

But these numbers, again—you see professional athletes getting numbers like this, but other than that, it’s unprecedented in human history. It’s hugely justified if you get one of the key research talents at this inflection moment in the competition toward ASI.

I don’t know if the choices of who to go after are necessarily right. We heard on our tour through San Francisco 2 weeks ago, Peter, that Llama 4 really does suck, and it’s kind of embarrassing. I know Mark tried to save it with a podcast world tour there, but if it sucks, it sucks.

That doesn’t mean they can’t catch up in a heartbeat, though, because a couple of tweaks in the algorithm and suddenly you’re back on top. You get the right people who know exactly how to try the next experiment the next week, and it’s worth a lot more than $100 million. It’s worth many billions, if not $1 trillion.

I think that’s the bet they’re making, and this isn’t the only one. There are a lot more of these going on.

Yeah. There’s a $70 billion war chest, and it really is a winner-take-all mindset in this. They’re willing to do whatever it takes to move forward.

Can I game this out a little bit?

Yeah.

If we go back to the previous conversation, I think what Ilya has figured out is how to use AI to tweak itself. That then gives you a very, very fast iteration path to what you’re trying to do.

If the investors believe something like that, then they go, “Wow, if he’s figured that out, then nothing will stop that from being the winner,” or something like that.

I think you’re right about that.

Now, regarding this particular thing, when WhatsApp was bought for $18 billion, everybody laughed at Zuckerberg, right? They thought, “This is nuts. This is unprecedented,” et cetera, et cetera. But it was actually a hugely important and relevant bet.

Given the past success in throwing money at this and going after it, you can see that he believes the market is that big, and this is pennies in the bucket—pennies on the dollar in terms of the potential outcome.

Yeah, I would never bet against Mark for exactly the reason you just said. He can act unilaterally and quickly, he’s aggressive, and he’s super, super smart.

What’s interesting, though, is the other thing we learned on our tour through San Francisco 2 weeks ago: There is a mass exodus of AI talent out of Meta. Then Sam’s saying, “Hey, they’re trying to buy everybody back for $100 million.” It’s like, well, dude, you just took everybody, so it’s fair game.

The question I have is, why were they leaving in the first place? Why did Llama 4 not come out the way they wanted it? We’ll dig in on that. I’ll try and get to the bottom of exactly what’s going on there.

The tide is certainly turning. Throwing money at it is one way to turn the tide, and it’ll likely work.

Yeah. The story here again is that Meta tried to buy SSI. They were rebuffed, and now they’re trying to hire Daniel Gross, who’s the CEO of SSI, and Nat Friedman, who’s been on our stage at Abundance 360.

Again, he’s out shopping, and he just made an acquisition. He hit the Neiman Marcus store for AI, and he basically bought our friends at Scale AI. Alexander Wang was also on stage with me a couple of years ago at A360.

$14.8 billion for a 49% non-voting stake in Scale AI. Dave, the IPO markets are just beginning to open in the tech and AI space.

Mm-hmm.

The acquisition markets are getting hot.

Mm-hmm.

You’re deploying Link SPV’s venture fund in companies out of MIT and Harvard. How are you seeing the acceleration? It’s been a relatively closed IPO and acquisition market over the last 5 years. Does it feel like it’s opening up now?

I’d say, as of the last month, it’s wide open.

Yeah.

These deals are unprecedented, huge deals. The CoreWeave IPO is way up. Circle is way up because it’s the way that agents can transact with each other. So, the Yahoo moment clearly happened. The door is wide open.

The deals are still concentrated among the top, the Magnificent 7. But Jamie Dimon sees that, Bank of America sees that, everybody else sees that, so their banking teams are spooling up. Everyone’s getting ready. It’s going to be just like 1997 and 1998 all over again.

That’d be great.

But much bigger.

Salim—

Also, a couple of things. I don’t know if people care, but the structure of the deal is really important.

Please.

I know a lot about the topic—if people don’t care, you can cut it out of the podcast. But this is the deal structure of the future. The 49% acquisition dodges Hart-Scott-Rodino, so the deal is closed the day you sign it. You don’t have to go through the 6-month torture waiting cycle of DOJ review.

It does skirt the edge of the rules, but the rules are bright-line. There are 2 parts to it. Forty-nine percent is not a controlling stake, so you don’t have to report. The other part is that it’s a non-voting stake. There’s another threshold of 19.9% or 20% ownership where you have to consolidate financials. But because it’s a non-voting stake, you dodge that rule as well.

So you’re like, “Well, okay, but do I really own the company?” Then you look at the contractual structure, which isn’t disclosed. There’s no public disclosure of the underlying agreement, and that agreement probably says, “We own all the intellectual property, and if you don’t work your ass off, you have to clean windows at Mark’s house,” and a whole bunch of things like that that really effectively make you own the company.

I also know that the investors in Scale are distributing the capital, so it’s not disappearing into the corporation. It’s going to the shareholders and getting distributed to the investors in the company. So it’s truly an acquisition.

Salim, I’m curious. You and I have both spent time with Yann LeCun, who previously was heading AI at Meta, and now Alexandr Wang comes in. Alexander Wang, I guess—Dave, he was a freshman at MIT, dropped out after his freshman year to start Scale AI. Was he the youngest billionaire out of MIT?

Oh, yeah. By far.

I wonder if Yann is going to stay on at Meta. Any thoughts there, Salim?

I think clearly there’s a changing of the guard there. Whatever they feel they’re deficient in, they’re trying to leapfrog, and they’re doing it very, very aggressively. Something that I love about Scale AI is that it’s really attacking the heart of the problem, which is the tagging of data. If you have that, you can solve the garbage-in, garbage-out problem in a really powerful way, and then it means you can use much better models.

The models you use become much better. You can use lesser models because you have much better data.

Yeah.

Right?

I think Yann LeCun—just to bring it back to that—is a super-brilliant, sweetheart guy at the center of all this, but he’s had a much more conservative point of view on AGI and ASI.

It’s the same with Geoffrey Hinton. Yann is kind of in that camp of saying everybody needs to slow down and be really, really careful about what we’re about to unlock here.

Yeah, I think Mark has the opposite—

Well, also—

Right? Yeah.

Mark is an engineer at heart. All these guys are engineers at heart. They’re not trying to buy an AI philosopher. A lot of the people who are the senior AI leaders from the big labs are saying, “Look, there’s something fundamentally missing from these transformers. They’re not actually reasoning. They’re just brute-forcing their way to intelligence.” They don’t want to buy that.

What they want to buy is, “I know how to make this algorithm work on a million concurrent GPUs. I know how to change the algorithm so that it stays synchronous across this massive amount of compute. I know how to actually deploy the Transformer algorithms. The SwiGLU isn’t working. We need to go back to ReLU.” That’s what’s in the minds of these mid-20-year-old geniuses. That’s what they want to buy.

What’s interesting to me is that normally the older, highly successful Eric Schmidts have all the money, and the young people have all the brainpower. But here, the young people have the brainpower, and now they suddenly have a lot of money, too. So that’s a new thing in the world as well. It’ll be interesting to track.

Well, speaking about brainpower and money, Masayoshi Son, one of the old-guard investors, has pitched $1 trillion. He wants to replicate Shenzhen’s scale within the US. Let’s take a listen to this video.

Talking about $1 trillion worth of investment out of SoftBank.

Yeah.

Put that into context for us.

$1 trillion, and he wants to recreate a kind of Shenzhen in the US, potentially alongside TSMC, of course, the foundry, and of course Samsung as well. You would imagine that OpenAI would have a piece to play, as would Arm, which is moving closer into data centers in terms of its CPUs aligning with AI accelerators.

So it hits that, and it ticks that box in terms of Trump’s ambitions. But we do need to find out where the capital is coming from, where the spending is coming from, and whether indeed they can get the talent and the engineers—not just to build all these projects, but actually to operate them as well—which has been a constraint and a bottleneck in the US.

Have either of you guys been to Shenzhen? I’ve been there a few times.

Oh, really?

Yeah, years ago.

It’s changed in the last 5 years. I was there between 2014 and 2019, and it was an incredible hotbed. It was an engine of innovation. The old mindset there was 996. It was 9:00 AM to 9:00 PM, 6 days a week.

People talk about China replicating stuff. There were a lot of entrepreneurs creating very new ideas out of there. But it was basically a convergent mecca for technology, and the idea that Masayoshi Son wants to rebuild that here in the US—I find that fascinating. Thoughts?

I think part of that vision doesn’t align. When we were at OpenAI, one of the questions I was asking D. Sculley was, you know, the computer science talent pool in Boston is about 20 times bigger than in Silicon Valley, and it’s also not nearly as picked over. Why doesn’t OpenAI open an office in Kendall Square—

Mm.

—just like Google did and Microsoft did? And the answer is, “Yeah, we would do that, except the timeline to AGI is so short—

That’s—

—that we’re going to have a multi-billion-dollar or multi-billion-person AI workforce before we could even finish the building and populate it.” And you’re like, “Okay, that’s a pretty interesting insight.”

That blows my mind.

Wow.

Blows my mind.

Yeah. So then you’re like, okay, well, Shenzhen—that’s a huge number of people, buildings, but is that timeline going to line up with the Elon Musk video that we saw a minute ago? I think the constraints here are electrical power and chips—

And not so much building a huge city that's all working on ASI.

It's the 6 Ds, buddy. Digitize, dematerialize, democratize, demonetize, and disrupt. Yeah.

Mm-hmm.

This is a chart near and dear to your heart, Dave.

Oh, yeah.

4. AI Startups Break Growth Records

Cursor, the fastest SaaS growth in history: $500 million of ARR in under 3 years, blowing away Anthropic, Uber, and OpenAI. Talk to me about this.

This is so inspiring for the teams in the office at Lynked Studio. Lynked Studio now has 26 teams from MIT, Harvard, and Northeastern.

These are—

They're—

These are startup companies that are being incubated at Lynked Studio.

Exactly. They're culturally just like Cursor: 3, 4, or 5 best friends from school, brilliant but never having operated a company before, building something with AGI or AI that's groundbreaking. They see a company like this thriving and hitting a huge valuation.

Back when this happened in the internet era, you got huge valuations, but the companies were very fragile because they didn't have a huge amount of revenue. These companies have $500 million in revenue. The margins on that must be astronomical. They're raising a lot of money at a big valuation, greater than $10 billion-ish, but they can actually operate profitably on 1 day's notice if they want to—

Mm.

Because they're not headcount-intensive. These are the best companies financially that we've ever seen in history, and the timelines are just laughable: 2 years to get to $500 million of revenue.

How many Cursors are out there in the next couple of years?

Dozens. It's actually constrained by the number of teams, not by the number of opportunities.

It's fascinating, right? We have constraints on electricity, on chips, and on the smartest entrepreneurs who take this forward. At least for the moment, it's humans that are constraining.

Yeah.

The difference in the vibe around Boston versus any time I've ever seen is just so blatantly obvious and palpable. You just need to walk around. But everywhere you go, everyone's just talking about Scale AI, talking about Cursor. These are fellow alumni that they actually knew.

Yeah.

The jealousy factor is a great motivator. It's really an amazing time.

Salim, these are all exponential organizations.

All EXOs. They have an MTP. They're using community effectively. They're building developer communities. The engagement levels are really great. They've gamified, in many cases, what they're trying to do. It's phenomenal to watch.

We predicted this in the book, right? We said we're gonna have a continuing increase in velocity, scale, and speed. This is surprising even to us at some level. $500 million ARR in this timescale is just unbelievable.

I do agree with Dave. There's a lot more coming down the pike on this, and these curves are just gonna get more and more vertical. Team formation is a really intriguing problem. How do you find the right teams? I'm wondering if you could use an AI solution to find teams that can then be put together and thrown together for this. That would be a really interesting problem to go out and solve.

Dave, can you do that?

Yeah.

One of our billion-dollar investments is Mercor.

Yeah.

It's all about finding and hiring. What are the attributes of the founding team that you're looking for?

That's a great question, and we study that all the time. We quote the Fred Wilson rule a lot. We call it the Fred Wilson rule; Fred didn't call it that. He's the founder of Union Square Ventures, an MIT alum, and numerically the most successful venture capitalist of all time. He keeps a low profile, so we don't talk about him every day, but Fred Wilson is really a god of the industry.

What he says, as he gets older, is, “I always invest in teams of 3 or more best friends who write the code themselves,” meaning they're technical—really technical, all 3 of them or more—“and I trust them. And if they pass those 3 filters, I invest, even if it's the stupidest idea in the world, because they'll change the idea much more quickly.”

You can't change friendships, you can't change relationships, and you can't change yourself overnight, but you can change your idea overnight. Once you get them into an ecosystem of other people who have great ideas, take them out to Silicon Valley, introduce them to Eric Brynjolfsson, take them to HAI, take them over to OpenAI headquarters, and run them through Google. We're doing all that with these teams now.

Yeah.

Then they come back home and they're enlightened, and they always have good ideas at the end of that—

Let me hit on one of the points there. People say, “Why should they be best friends? Why should they be around, having had a relationship for a number of years?”

Road trip.

We just saw these large companies like Meta, Google, and OpenAI raiding companies and stripping out the talent. If you've started a company with some stranger that you don't know, and it's been 6 months, and someone gives you a huge signing bonus, you're gonna leave.

Yeah.

But if you've started a company with your best friends and you've got a long history, you're not gonna abandon them. I think that's a really critical point.

That's exactly right. That's the number one failure mode, actually, for companies: somebody bails. You'll always succeed in the end if you stick with it. The way you characterized it is exactly right.

You'll pivot a dozen times.

Pivot a dozen times, and you'll—look, name any one of these companies that didn't pivot at least once. Going back to PayPal, every one of them pivots at least once. That's just part of the journey. But when you pivot and then someone says, “Oh, I give up. I'm leaving,” that's what ends up killing the company.

The way I've been phrasing it for many, many years is more true than ever before. Imagine that you're on an international flight and you're sitting right next to somebody in a middle seat. The way you feel when you get off that flight, that's the way it's gonna feel when you're doing a startup together.

So if it's you and me and Salim flying, we're gonna come off that plane energized because we've been talking about everything in the world for 8, 10, 12 hours.

That's right.

Not only that, we'll have infected the 3 rows around us to all get into a conversation.

Yeah, walk out with more employees than you started with.

Yeah. Salim, when we had Singularity University's Graduate Studies Program, the GSP, going, we were starting companies that had a 10^9-plus mission, right? Impact a billion people over a decade.

It was around the same exact timeframe that Y Combinator was getting going. The failure mode, I think, was that we thrust 100 alpha males and females into a room, independent of each other, and said, “Start a company.”

That was very different from Y Combinator, where teams came in with an idea already. That preexisting glue around something that they're all passionate about, I think, is a super differentiator for investments.

If you look at the track record, the ones that did succeed were the ones that became friends and stuck together over time.

Yeah.

The one thing we did do was create lasting friendships that lasted a lifetime. If you look back, a lot of those alumni have gone and started working together where they found affinity, not necessarily on the teams. Team formation is critical, and that early chemistry is really important.

This whole conversation reminds me of Yossi Vardi, who kind of single-handedly created the Israeli startup scene.

Yeah.

He sold ICQ to AOL, which then became AOL Instant Messenger, for, like, half a billion dollars back in the '90s.

I remember it well.

He did something amazing. He basically went to founders in Israel and said, “If you're a good guy and you have integrity, I'm giving you $50,000. That's the bar.”

Then he just trusted them. He would check their integrity and their character very carefully, and then he'd just give them money. He invested in something like 400 startups. The outcome of those has been a little bit like the Fred Wilson type, where it's just been off the hook.

The team and the individual that you're betting on is everything in this type of a world.

You know, Salim, you're so right, and you reminded me of something that's really important. Peter, remember when we went over to Israel, to Tel Aviv, to Startup Nation, to try—

Sure.

—and figure out why the startup success rate is 5 times higher there than anywhere else in the world, per capita. There are a lot of reasons, but a lot of it comes back to everyone having to do their military time, and there's a huge amount of bonding—

Hmm.

—you know, just marching through the desert together and suffering together, and that creates these lifelong friendships. Then you go to college and appreciate it a lot more, and then you start your company while you're in college. So they're a little older but a lot more bonded when they're going through that experience.

But if I port that back to the US, MIT is absolutely thriving like I've never seen before in terms of startup success. Daniela Rus, who runs CSAIL at MIT, the biggest AI lab in the world, has 2 daughters. One went to Harvard, and one went to MIT. Her Harvard daughter was constantly at MIT for the parties.

Okay.

And nobody thinks of MIT as a party school, right? Why would you? But actually, when you're there, it has an immense amount of bonding. Part of it is because of the way it's set up with the living groups and the fraternities and sororities. Part of it is because the school is so freaking hard, and that's like the Marines of—

You cannot get through on your own.

Yes.

Those problem sets, all night long with your best buddies trying to get the answers.

Just a plug here for Waterloo, which is the MIT of Canada. We have the same thing. You couldn't get through unless you collaborated really closely with a bunch of other fellow students, and that created lifelong friendships. Really great point.

Yeah.

Right. Well, if anyone's listening out there in school administration, Harvard has a little bit less of that bonding culture because school is so stupidly easy. Everyone says it—Mark Zuckerberg, Alex—everybody says it. And it also doesn't—

Hard to get into, harder to fail out of.

Harder to fail out of.

Oh, boy.

Yeah. And then Stanford has become even worse. If you talk to the students there, they're like, “Where the hell is the crazy, fun bonding culture?”

One of our Harvard guys in the lab decided he was going to open a window and do a rock-climbing drill from the second floor down to the first floor on the brick wall. There were cops all over the building and security guys running around, and they're like, “Dave, why are the cops all over the building?” I said, “Guys, let them be. This is what they need. They need to bond. They need to blow off steam. They need to be a little crazy. This is what's going to create the success in the long run.”

But we also can't have the cops here every day. This is the culture that is ultimately going to thrive because it's kind of lacking on the Harvard campus, and they need to create it.

Yeah.

And so they are self-creating it.

All right, this next slide from Andreessen Horowitz is pretty epic. It's labeled, “What's working means in the era of AI apps. Gen AI startups are shattering growth records.” Dave, this must make you feel pretty amazing.

Yeah. This chart's a little hard to read, but if you look at the top quartile, it looks like they raised less money, but they actually raised it much more quickly. So if you look in the bottom-right corner, pre-Series A dollars raised: $3.1 million. That means they were very capital-efficient getting to $8.7 million in revenue run rate.

This is what I was saying earlier. The fundamentals of these companies are so good compared to the Internet era. They're very capital-efficient, with great ARR, and this is accelerating really quickly now.

Did you see that one of our companies, Farsight, was talking to JPMorgan as a customer, and nothing was happening for months? It was just unresponsive. Then Jamie Dimon sent an email to everyone in the entire company, every manager, saying, “If someone is trying to sell you AI, you better buy it or at least listen right now, because this is going to change …”

All of a sudden, they called back. JPMorgan actually reached out to Farsight and said, “Okay, we want to talk. Get over here right now.” So that's going to happen now across most of the Fortune 500 and all of the mid-market. This will get even more traction very quickly now.

I just love these numbers. They're extraordinary, right? ARR: $8.7 million, time to a Series A in 5 months.

Yeah.

Extraordinary. All right. Well, I mean—

I mean, those are crazy-good numbers, but then look at the Cursor number from the prior slide.

Yeah, sure.

$5 million in 2 years. I mean, it's—

Sure. I mean, a grand slam—

It makes these look weak, right? So you look at the bottom quartile there, right?

Yeah, 3.

They raised $10 million in Series A, and in 6 months or 12 months they've brought back a third of it. That's still an amazing number, comparatively.

It's a great point, Salim. Yeah, $3 million in 12 months would have been top decile a couple of years ago. Here it's bottom quartile. Yeah, that's a great point.

Yeah, an acceleration of acceleration.

Mm-hmm.

5. Government Deploys AI

Here's a big story in AI this past week. The Trump administration is launching AI.gov. They're hoping to launch it on July 4. I hope they hit it. I just want to dig into this a little bit for those who've been frustrated by the government.

This is a project being led by a Tesla engineer by the name of Thomas Shedd, who's leading the team. The idea is: Can the government use AI across federal agencies—the GSA, DOT, FDA, DOE, FAA—all these agencies that have been sublinear in their existence at best?

Let's chat a bit about this. The GSA, the General Services Administration, buys everything for the government. They could use AI to optimize procurement, get vendor performance, automate contract analysis, and really eliminate fraud. DOT is going to be predicting flight delays, analyzing real-time vehicle data, and helping support infrastructure like roads and bridges in advance. Wouldn't it be great if they could predict where the potholes are and get those fixed?

Mm-hmm.

DOE is about optimizing grid operations, forecasting demand and supply. For the FAA, we're going to see a story on automated drone traffic management, weather avoidance, and—just as a pilot—the FAA's air-traffic-control system is a bloody 1950s mess.

And then, of course, we've spoken about the FDA using AI to enhance drug and device approvals, develop faster clinical protocols, and optimize food safety. If there's one part of the world that needs optimization with AI, it's the government.

Yep.

Thoughts, gentlemen?

You know, a great roadmap for how this works, too, I think, is Palantir and AWS. Everyone was worried about how the government was going to interact with cloud computing. There's a huge privacy issue here, and is the government going to start building its own data centers and its own cloud? Obviously, they don't know how to do that.

So Palantir and AWS set up secure clouds, private clouds for the government, and that became the roadmap. Now with AI, it's like, well, I obviously can't take all my government documents, tax returns, and everything and dump them into Gemini or into ChatGPT. How's that going to work? So now, if the big AI companies are smart—

And just explain to people why you can't do that, why you shouldn't do that.

Well, first of all, there's no compartmentalization, so it gets pooled into the training data with everything else. Everyone asking, “What time's the soccer game?” gets pooled in with someone's tax return and goes into the training data. Then somebody else queries ChatGPT and says, “Hey, what did Peter Diamandis's taxes look like?” and it just answers.

So that's not going to work. All kinds of concerns like that. But it's going to be figured out in the private sector and sold to the government as compartmentalized AI modules.

There are a lot of questions about whether departments can pool information and share their AI, so those are really tricky conversations. But I guarantee none of the ideas are going to come from the government. What will happen is mandates will come out.

This is what the state of New York just did. The governor of New York said, “You know what? We need a gigawatt of nuclear power.” “Okay, any ideas on how to do that?” “No. I'm just saying, make it happen.”

Then every private-sector genius can propose a way to do it, and they'll just approve one of them. That's what will happen. The same will happen with AI here.

Hopefully, the big AI companies are aggressive in building up their government-services operations, or they bless some other third party, like a Palantir type or a new startup, to go and become that entity.

But that's the only way this can actually happen, and God knows we desperately need it, right? AI can solve so many government problems—

Oh my God.

…so quickly.

Yeah. Salim, you've been working with governments around the world with your ExO hat on. Speak to this, please.

I have so much to say here. Okay, 3 quick points. One: note that most government processes are prescriptive, and the minute you have a prescriptive, repetitive process, you can apply AI to it and totally change the game. I think that's a huge area.

Second, a few years ago, I was asked to give a talk at the Republican National Leadership Conference, and the title of my talk was going to be “How Do You Drop the Cost of Government by 10X?” You could do it easily using some of these technologies: blockchain, AI, et cetera, et cetera.

Third, I'll give a specific example. If you were applying for a wind turbine approval in—I think it was in Colorado—it was taking 2 or 3 years to get approval for that, right? Then they brought in a programmer who put it on a Google Map: Where are the electrical mains? Where are the water mains? Where are the flight paths? They were able to reduce that 2-year approval time to 30 seconds.

And that's just the smallest example of how you can do this across the board, and we've mentioned some of these already. It's going to be a game changer. I can't—I’m so excited about the potential government applications of this.

I love it. I love it. And it is.

Here's a related story. The U.S. Army appoints Palantir, Meta, and OpenAI execs as lieutenant colonels. This is a special unit created to support the government. I'm going to give some names here because they were published. The appointees include Palantir's CTO, Shyam Sankar; Meta's CTO, Andrew Bosworth; and Kevin Weil, OpenAI's chief product officer. Kevin's going to be joining us on this podcast next month.

I can't wait for that.

Yeah. Dave, you and I had an amazing meeting with Kevin up at OpenAI headquarters.

Mm-hmm.

And then Bob McGrew, former OpenAI chief revenue research officer. I find this absolutely fascinating, sort of indoctrinating them. They made it super fast. There's no required traditional training, no boot camp for these individuals. What are your thoughts?

Peter, you didn't read the quote there—the backlash quote.

Okay. “The appointment of a lieutenant colonel in the U.S. Army followed the creation of a special unit created for rich, big-tech mavens seeking military leadership roles.”

This is like—

Uh—

Exactly what you're always saying. Everything turns into a drama, whether it needs to be or not. That's just the nature of social media.

I mean, this is one way that the government can bring in extraordinary intelligence that they could never hire or recruit otherwise. This is sort of a part-time military service to make sure that the U.S. government and the U.S. military have access to the brightest minds.

Yeah.

This is—

Yeah, exactly. And then I don't know the other guys personally, but Kevin Weil we know, and he's one of the perfect—

He's a sweetheart guy, brilliant—

Perfect guy.

Yeah.

Absolutely. Is somebody going to just naturally join the Army as a private, work their way up, and end up being aware of how to use AI to solve government and military issues? No, that's not likely to happen. So go get the best guy on the planet. He's absolutely the right guy.

I mean, just—he's a physically impressive manager. He can actually move mountains—

Yeah—

—while still being the nicest, sweetest guy on the planet, and he knows exactly how this stuff works. This is just great for everybody.

Yeah.

I don't know. I don't know. The negative—

It's a smart move—

—the spin here is nutty.

Salim?

I think that this is a great example of a human being plus AI, because as they bring AI to help in these roles, it's going to be totally transformative. You, of course, have the monster immune-system response, with people going, “Well, you can't do that unless you've worked your way through the ranks,” et cetera, et cetera. People have a thing about it, but I think this is a great application.

This reminds me of the big problem around leadership training, right? We've spent decades, and hundreds and thousands of books have been written on leadership training. A few years ago, it turned out the best leadership training in the world was World of Warcraft.

The fact that technology can outstrip this age-old human institution is unbelievable, but it's there. I think that, added to what these guys can do, plus bringing technology and their mindset to the mix—

Mm-hmm—

—this is where I think it will have the biggest impact. They'll bring that mindset to it and, hopefully, infect the rest of the armed forces with it.

Yeah.

Mm-hmm.

6. AI Predicts Extreme Weather

All right. Talking about breakthroughs in AI, I love this. Hats off to DeepMind for continually pushing new capabilities out that support all of humanity. This is a DeepMind algorithm supporting better tropical cyclone predictions.

Let me give you the data. It's a 5-day track prediction that averaged 140 kilometers closer to the actual storm path. It's the difference between hitting Florida and Georgia or Virginia. This was trained on 5,000 cyclones over 45 years. There's about $1.4 trillion in economic losses from cyclones over 50 years, and I'm excited about this. Thoughts?

Yep.

Yeah, I mean, there's a piece of AI folklore called “The Bitter Lesson,” where basically, anytime you throw a lot of data and a lot of compute at one of these algorithms, you're likely to get a great outcome. You can sit there and stare at a wall trying to think through how to do it with differential equations for 3 years. You're not going to compete with the big-data approach.

This is a great case study. I'll bet the people working on this got it cranked out in a very short period of time with just a couple of people, yet it's far more effective than anything that's been done over 50 years of weather research. There are so many of these around. The benefit to humanity, if we coordinate it and wrangle it correctly, is immeasurable, and this is a great example.

Salim, what are your thoughts here?

I label this as “nothing to see here,” and I don't mean that in a negative way.

What does that mean?

This kind of thing should be completely expected, right? You take an ancient data set where human beings are trying to hand-plot these things, which they're never going to do that well, and now you throw AI at it, plus the rich data set that's very bounded, and we know exactly the history. Of course it's going to come out with a much better thing, and thank God.

Look at the predictive ability now going forward. Huge impact, but I think we should take this as something we should expect—like 1,000 of these coming out in the next year or 2 or 3.

Yeah. These would have been great XPRIZEs as well, and I've been pushing for this. I think DeepMind is an extraordinary company under Demis Hassabis, and creating these kinds of assets to support humanity is really in their DNA, in their culture.

I can't wait for an earthquake prediction XPRIZE—an earthquake prediction algorithm, right? If you could predict an earthquake with 10 minutes or 30 minutes' warning instead of 30 seconds or a minute, getting people into safety would be huge. I mean, one of the—

Wait, can I just—

Yeah.

—fill in on that?

Sure.

That's a perfect example. We know animals can sense this early.

Exactly.

Right?

The data's there.

The data's there; we just have to get the right kind of algorithmic approach to it, and that's an area where I would expect to see a breakthrough in something like that. Then when it happens, everybody's gonna go, "Oh my God, this is unbelievable." But we should expect things like this.

Yeah.

In fact, we should take areas and go find them. We know what the answer could be, we know definitively it's possible, and then put AI against that.

Yeah, my guess is we'll see earthquake prediction algorithms within the next 2 years, if not sooner. Here's an interesting, controversial thought on this. Imagine if you could control the direction of a hurricane. So instead of having it hit Miami, you steer it down into Central America. You go, "Oh my God, why would you possibly do that? That's terrible."

Well, if you hit Miami and the cost of the impact there is $50 billion, and the government of Costa Rica says, "Listen, you pay us $20 billion and you can land the hurricane here," that's an interesting arbitrage on geography. A crazy idea. We have people that control the path of these things.

Do you know the mechanism that they'd use for that?

A butterfly effect. I think you'd use lasers and heating the atmosphere. Maybe it's magic voodoo dust. I don't know.

I can just see the other side of that coin, right? "Hey, Mr. Trinidad, you maybe wanna pay us some money so that we make sure that hurricane doesn't hit you." And you get into all sorts of crazy outcomes.

Oh.

Mm. Mm.

If you can measure it, you can impact it. And I find this, again, has huge and interesting implications.

7. AI Rewires Childhood Learning

Okay, our next story here comes from Mattel and OpenAI. They've announced a strategic collaboration. I've talked about this forever. I'm super excited. Your toys are gonna become super intelligent with GPT-5. Your Barbie doll, your Hot Wheels, your American Girl—I know, Salim, that you in particular like the American Girl dolls—Thomas and Friends. Sorry, buddy.

That's okay. You've uncovered my deep secret.

I think this is gonna be a boom for the toy companies.

Yeah.

And this is gonna enable rapid early education of our kids.

Yeah.

The area I'm really excited about is the feedback loop. As you interact with these toys, you'll learn a lot about the child, and we can use that for understanding learning behavior, implementing guardrails, and understanding where their motivations are. I think it's so exciting because we'll get more data about young children than we could ever have gotten before. Yeah.

Yeah.

I was a little surprised OpenAI wanted to touch this one. It's really clear when you're talking to AI voices now that within a year they're gonna be just crazy engaging, super, super friendly, and a lot of kids are gonna prefer talking to AI all day rather than talking to real friends.

Mm.

And there are good and bad things that come along with that. I've loved every minute of raising my kids, and I hate to see that change in any way, but it's clearly coming soon, and it's inevitable.

The other part, Dave, is sparking their kids' imagination, right? When you have to make up what your Ken doll or Barbie doll is saying, that's critical for fostering early curiosity and imagination.

It is. And you have the echo-chamber risk on the other side of that. So if it's done right, it's an educational goldmine and the kids are happy, and a lot of schools are terrible, so you're alleviating a lot of that. If it's done right, it's incredible. If it becomes an echo chamber, then you can see where it can go bad in a real hurry too.

That's why I'm surprised OpenAI wants to touch it.

Oh.

Because when you start talking about kids, you really got to get it right. You cannot make a mistake, right?

Yeah.

You gotta get it right.

Yeah. There was a friend of mine who had a product called Moxie, which was an AI robot, and it was mostly being used for young kids with educational challenges—neurodiverse kids. In which case, creating a best friend and helping them open up and communicate, there's real value there.

Yeah, well—

What happens when you AI-enable Chucky? That'll be interesting.

A whole new set of movies coming out. All right, let's jump into a little bit of AI in education. These are some scary reports that came out. This is in Time magazine: "ChatGPT May Be Eroding Critical Thinking Skills, According to a New MIT Study." Dave, did you track this?

Yeah, I did. Well, only because you put it in the slides. I said, "I better understand what's going on here. This sounds really, really important." So I dug up the research and read it. It's nothing surprising if you think about how Waze works, right? A lot of people don't know how to drive anywhere unless they turn on Waze. Places they go every day, they still could not get there without turning on Waze. So it becomes a crutch really, really quickly.

Yeah.

So that's what's happening here with writing, where you would've thought through all the underlying topics in order to write it, because the AI is filling in the blind spots. You're just not really understanding what you just wrote. But it's not at all surprising. When you write it up and put it in a headline, it opens your mind to what's going on. But when you read the paper, you're like, "Oh, duh, of course. This is exactly how it's gonna work."

Yeah, this is some of the data. I don't know if you want to use this to recount what the study said.

Yeah. Well, no, it's really straightforward. If you write a document yourself and you have to think through every single word of it, that time that you put in means you can then recount what you just wrote with incredible accuracy. The failure rate on the top line shows your quoting accuracy of what you were just talking about.

So if you use AI to write the same paper and then immediately ask you, "Hey, what was Shakespeare's favorite toy?" you're like, "I have no idea." Well, you just wrote it down. Like, "Oh, did I?"

Yeah.

That's the difference there.

An 83% failure rate if you use ChatGPT versus 11% if you're using Google, meaning you're actually looking up the data, then you're composing it. When you're doing the research and the writing, you're effectively training your own neural net, and the data's being deposited in your brain. It is scary. It's going to become a crutch in terms of thinking, and it's gonna get a lot worse. Yeah, Salim?

Yeah, I didn't read it as all bad, though. I think you got the work done 100 times faster in the left column. So your neural net actually didn't have time to retain every little detail. I don't know. If society's gonna move 100 times faster, we're not gonna retain every detail. It's just that simple.

So yes, as a teacher, you could say, "Look, the kids are not really learning this stuff," but as a person moving through life, the kids are covering a lot more terrain. Isn't that more important? It's a mixed bag. It's not all bad.

Salim, good or bad, what are your thoughts?

I have an 80/20 approach: 80% unnerved, 20% will navigate this. The 80% is, I actually saw this with Milan, my 13-year-old. He was writing an essay and just used ChatGPT to write it, and he clearly had no memory of that essay. He would've fallen completely into these buckets, right? So when I saw this slide, I was like, "Whoa, this reminds me completely of what Milan just went through." He has no cognitive framing for what he wrote in that essay because he used the AI to help.

Now, the other side of it is it's happening much faster. I think the key question is, how do you effectively train kids on critical thinking into the future? And a guest we might wanna think about is Nicole Dryski, who's actually solved this problem and has found a way of teaching critical thinking to kids in a very active and very accelerated way compared to the way we do it.

The last point I'll make is that I remember seeing a study that 52% of the CEOs in Silicon Valley are liberal arts majors.

Mm-hmm.

Right? Which is the ability to think in different ways—

Mm.

—is a critical factor of success in leading a tech company.

Yeah.

I find that really interesting.

Paul Graham has a different spin on that: liberal arts majors succeeding is more tied to their desire to not do irrelevant, difficult things and get to important topics, and it's just an easier way to get through—

It's a good filter.

Silliness.

It's a good filter.

It's a good filter either way, yeah.

Dave or Salim, I don't know if you saw Andrej Karpathy's presentation at the AI Startup School. I showed it to my son, Jett, yesterday, trying to incentivize him about vibe coding, which Andrej came up with. He put out the tweet that went viral defining vibe coding.

Mm-hmm.

My son's reaction was, “I want to learn how to code, not just vibe-code.” Right? So I'm curious what you think about that because the concept right now is that English is the new coding language.

Good idea.

But there's a lot of value in fundamental coding, so what are your thoughts, Dave?

Yeah.

Oh my God, so many thoughts on this. I have a little story—an amazing experience this week. I don't know if you want to hear it.

I do want to hear it.

Yes.

Of course you want to hear it. You remember that right when I got out of MIT, I went to MicroStrategy, came back to Boston, and started my first company, DataSage. Our very first product used neural networks for handwriting recognition.

Yeah.

I took the backpropagation algorithm, read the raw research paper and all the differential equations, and said, “I have to try to code this up on real-world hardware in assembly language.” From the day I started on that journey until we productized it, it was 4 years of coding.

Wow.

And then that turned into a few million in revenue, and then we just skyrocketed from there.

Wait, did you say you coded it up in assembly?

Well, we had to, because we started out in a high-level language, but you needed to squeeze every MIPS and every FLOP out of these processors to build anything scalable enough to work. We invented quantization, which is now all the rage, to try to shrink the parameter size. Anything to get another 2× performance out of these things, just to make it work in a real-world use case.

Awesome.

And so it was a lot of hard work. Then this week, on Monday, I said, “I wonder how long it would take me to recreate those 4 years of work?” No joke, and I'm absolutely not exaggerating, it took less than an hour to vibe-code it from scratch to the exact same product—even the demo, the graphical interface I put on it. I vibe-coded the entire thing in under an hour.

Now, it's not entirely apples to apples, because there's a ton of open source out there that vibe coding can pull in, so just to be fair.

But the broader point is well made, though.

Yeah, yeah. 4 years down to an hour is just mind-blowing, what you can do. Now, your son is exactly right. Spend some time looking at the code and writing some raw code. You don't realize how much faster you are until you try to do it the old way, and so it's really, really important to get that experience.

So let me give you the end of this. He went off and wrote a complete short story without any help, and it was mind-bogglingly good.

Mm-hmm.

I thought he used an AI to write the short story, but it was really, really good.

See, I think we're in this golden era. Right now, AI can do almost anything for you, but it's not creative yet. So you still have to think; the human component is still by far the most important component.

I don't know if that period of time will last forever, but right here, right now, is such a golden era where you're empowered but not demoralized or crushed. It's just such a wonderful next couple of years, and you really have to take advantage of it.

8. AI Automates Work and Mobility

All right, next topic here is AI and job loss. This was a Stanford survey revealing which jobs AI would most likely replace. They surveyed 1,500 workers and AI experts. 69.4% want AI to, quote, “Let me focus on high-value work,” unquote, and 46.6% want it to take on repetitive junk. That's kind of obvious.

Here are the occupations most likely to be automated: bookkeepers, payroll clerks, data entry, insurance claim processors, software roles, tax preparers, public safety, and telecom. Dave, Salim, any thoughts?

Well, the high-level thought is, look, everybody needs to become a user of AI to get ahead of this. The study would say, “Wow, it looks like bricklayers are going to be immune for a while,” but you're not going to go and become a bricklayer just because you've got 3 more years of immunity. That doesn't really give you any actionable advice.

Yeah, robots are coming for that job too.

Yeah, exactly. So don't do that. Just start using it every day, understand it, and ride the wave.

Let's double down on that. If you're listening to this podcast and AI sounds fascinating, but you're not a power user, how do you start? Salim, what do you start with?

Just take a task that you're trying to do and ask, “AI, how would I do that task?” Then say, “Do it for me,” give it the raw data, and watch it just go.

It pretty much is that easy. And I love the point you made, Dave, earlier: I do use ChatGPT's voice interface, and when I'm doing my red-light therapy or taking a sauna, whatever the case might be, or driving, I'm having a conversation about whatever subject I'm curious about.

It's extraordinarily educational and fun to have this back-and-forth with an AI, saying, “I don't understand that term. Could you dive in?”

Yeah.

“What's the data that backs it up?” Or, “When did that happen?” I mean, just being able to have a continuous conversation—it's like a childlike experience of “Why? Why? Why?” digging deeper.

That's exactly what you should do.

Yeah.

And get out of the media rut. I know, Peter, you say this a lot.

Yeah.

But we have the internet now. You can actually watch useful media like this or study KJ Hardrict and his podcast. He's phenomenal. Watch Lex Fridman; watch Dwarkesh Patel. But while you're doing it, if you don't understand anything, have either Gemini 2.5 Pro or ChatGPT-4 open and just talk to it while you're listening to the podcast.

Stay out of the mainstream media. Do not listen to the rock stars, sports broadcasting, whatever. It's going to distract you and suck you in. You need that time back, and use that time this way. Once you're into it, it's just as much fun, but much more useful.

Can I make a couple of points about that last line?

Of course. Of course.

Let's note that a lot of those roles are being done because people have to do them, not because they want to do them, right? It's grunt work done repetitively, and it's perfect. It's a perfect area.

I was really happy to see that a large number of folks, 40-whatever percent, said, “I want to take all the repetitive tasks,” because nobody wants to be doing that anyway, and now we can automate a lot of that side of it.

And just a quick point on bricklaying. I remember showing a drone at Singularity University a few years ago where you had drones working in cohesion to lay bricks on a wall. The drones did it in no time flat because we know exactly where the bricks need to be laid. It's totally prescriptive. Off you go.

Yeah.

Mm-hmm.

This is a story from The Guardian: Amazon testing humanoid robots to deliver packages. Took them long enough, right? We see in the image here Agility's Digit robot.

I was a little alarmed when you read that it's going to spring out of the van. That's going to be a little unnerving when that starts happening. My 2 dogs are already freaked out by the postman. What the hell are they going to do with these fucking guys?

Yeah. I mean, come on, guys. We're going to have autonomous vans driving around where robots do the last 10 meters of delivery to your doorstep. And it's just going to drop the cost of this. So we're going to see drone delivery of products. We're going to see autonomous-van and robot delivery of products. This is happening. There's no question about it.

It's a when, not an if. Agree.

Right. And so the question is when. When do you predict, Salim?

I'd say middle of next year, because the technology's all there.

There's probably some regulatory stuff to get through, which might slow it down, but the technology's potential is there now.

Mm-hmm.

Yeah, the technology's there. I think the humanoids are inspiring everybody. There's no surprise here. But what people don't talk about as much is that the robotics doing nanosurgery and microsurgery are unbelievable. Also, the robots crawling through pipes, cleaning things, and getting into sewer systems are taking off. They're not as sexy as the humanoids, but it's all happening concurrently, and it's just a tremendous benefit for humanity.

I want to make a point about humanoid robots.

Okay.

You know, we've had this long, strong discussion where I've been saying—

I'm tired of your point. You're making it here over and over and over again.

Hold on. I thought of a very real-world example of my octopus idea, which is that you've got a humanoid robot. How many times, when you're doing something, do you wish you had a third arm to hold the garbage bag open or whatever?

Better.

So can we please have these humanoid robots have 3 arms?

No, we'll just have the 2 robots with 4 arms.

Anyway, here's what's interesting: We're going to see these companies stack resources. I would not be surprised to see Tesla get into delivery services with autonomous vehicles and autonomous robots. We'll see Amazon and Agility, obviously, and we'll see other companies coming together on this. But a full stack in the physical world.

Let's talk about robotaxis. Big news since we spoke last: On June 22, the Tesla robotaxi launched in Austin with a flat fee of $4.20. Elon loves his $4.20 over and over again, so the robotaxi is now live. Here is Elon at Tesla headquarters in Austin, and here's a quick video of what a fan experienced in the robotaxi. Thoughts? Finally, he's done it. I think it's one of the biggest markets that Tesla's going to experience. We've seen Cathie Wood predict it as a multitrillion-dollar marketplace. Tesla has been known as a car manufacturer. It will be known as a humanoid robot and an autonomous robotaxi service delivery company.

Mm.

Thoughts, gentlemen?

Two quick thoughts. One is: never bet against Elon. I've learned that through my massive investment portfolio suffering over the last decade. The second observation is that these cars are technologically quite inferior to the Google Waymo taxis because they don't operate in heavy rain, they're limited by human-level sight, and that's a problem in the long term. He may get some buzz out of it, but he's going to have to upgrade the lidar for this to be really workable—

Well, we'll talk about that.

He could just be waiting for lidar systems and all that to hit the cost curve where they become cheap enough, and then you just flip over to that.

We'll talk about the next one.

Hold on. Let me just step on the other side of that.

Yeah.

He could just be waiting for lidar systems and all that to hit the cost curve where they become cheap enough, and then you just flip over to that.

Well, you know his thesis, right? If a human driver can drive with just his or her eyes—in fact, with 1 human eye—then, from first-principles thinking, an autonomous car should be able to drive with a couple of cameras. And that's his basic thesis.

Yeah.

Why not give it superhuman skills?

That's great, except lidar can see 5 cars ahead.

Why not give it superhuman capabilities?

You can do it. Why not do it? And yes, it's much more expensive for now. I think he went down a philosophical route of moving away from lidar, and I think that's going to lead to an inferior product. But if it's workable and it works well enough, that's fine.

I mean, listen, I drive with my Tesla's autonomous mode all the time, and it works perfectly. The only time it ever stops working is when it catches me picking up my phone and looking at it, and then it beeps at me. Otherwise, it's extraordinary.

That part surprised me, actually. When we were driving together a couple of weeks ago, the inward-facing camera was new to me. It's actually watching your behavior and then adjusting. It's like, wow, that's a little weird.

Yeah, it's funny. If I'm looking out the window to the left, it beeps at me. If I pick up my phone, it beeps at me. It's really annoying.

I've kept my 2017 Tesla Model S for exactly that reason. There's no inward-facing camera, so it's great. I've now driven that car 4 times up and down the country, from Miami to New York or Toronto. So I've become a bit of an expert at highway driving—

Have you heard about airplanes, Salim? Have you heard about this thing called an airplane?

You know what? I load up with 5 of my favorite shawarma kebabs, I arrange 40 conference calls, and the car drives itself 80% of the time. It's free to do that, and I arrive more refreshed than when I left. I mean, it's phenomenal.

All right. Let's take a look at this next one. Tesla faces protests in Austin over Musk's robotaxi plans. Protesters claim Tesla's FSD, or Full Self-Driving, has been linked to hundreds of crashes, including dozens of fatalities.

Let's look at the data here for a second. Here it is: number of accidents per million miles. Without any question, humans are terrible drivers. We're distracted constantly now, with at least 1 cellphone per car.

We are terrible control systems for 2-ton cars going at high speed. Can I throw out a quick data point here? In 2011, I think it was, there was a 3-day outage in BlackBerry services around the world. Nobody could send text messages for 3 days. The accident rate in Abu Dhabi dropped 40% during those 3 days. Today, you look at anybody driving, and they're looking at their phone. It's to the point where we absolutely should not be driving as human beings.

I agree. We're going to see an extraordinary increase in safety. Just the idea of a 16- or 17-year-old driving a 2-ton vehicle at 60 miles per hour through the streets, with very little experience and their phone distracting them, should scare the daylights out of everybody.

Well, this is America's greatest Achilles' heel, actually: We're responsive to sympathetic stories that are statistical rounding errors.

Or wrong.

Or wrong, yeah. Look at the State of the Union address. Somewhere along the way, it changed to, “Let me call out 4 or 5 people in the crowd here and tell their personal stories.” And you're like, “Well, how do I know? That could be 1 in a trillion, for all I know.” But, yeah, that's okay. I'm just trying to sway a bunch of voters toward what I'm trying to get done here.

We're ridiculously swayed by that. If you look at that prior slide, yes, hundreds of fatalities from self-driving. Well, hundreds relative to what?

Yeah.

You put your statistician hat on, and you're like, “This is clearly better.” But some of these are really, really tragic, and if it's all captured on video—which everything is now—it can sway opinion. This is where, if America is going to lose to China or to some other state, it's going to be for this reason.

Yeah.

Because—

It's the same concept when people hear about an airplane accident: “Oh my God, I'm fearful of flying.” It's like, have you looked at the accident rates in cars? Flying is still the safest mode of transportation.

1.2 million people a year die in car accidents around the world every year.

Yeah.

Well, hey, nuclear power—we'll come to that later, I guess—but that's another case study.

Great point.

We will. So here's an important news bite from this past month here in Los Angeles: 5 Waymo vehicles were torched in downtown Los Angeles. Waymo paused service in Los Angeles and limited it in San Francisco, Austin, Phoenix, and Atlanta. The question is: Is this the beginning of the Luddite revolt?

Are people responding against all of the technology? Here's an image of the Luddite revolution from 1811 through 1816. I have some data I want to share with you guys on this because I think it's worth noting, and I want to talk about it.

The Luddite revolt over that 5-year period was a series of protests by English textile workers against the mechanization of the Industrial Revolution, particularly against automated looms that threatened their jobs and wages.

Sounds familiar.

Mm-hmm.

Right?

Mm-hmm.

Named after the mythical Ned Ludd, the movement began in Nottingham and spread across the UK. Workers fearing deskilling and unemployment amid economic hardship from the Napoleonic Wars destroyed machines in organized raids at night. The revolt involved thousands, and the British government deployed 12,000 troops to suppress it. Sounds familiar.

Mm.

Here we go. “Machine breaking became a capital crime in 1812, leading to 17 executions, dozens of hangings, and the movement was crushed by 1816.”

Mm.

Thoughts, Salim?

This is the classic immune system response, right? In a slight twist, I note that when I read this up, the Waymo cars were actually called to that spot so they could be attacked. So the AI is watching those people carefully for future retribution.

But I think this is the general backlash of technology against society. What we don’t understand, we fear; what we fear, we get angry at, and we try and destroy.

Yeah. I like—

You should insert a picture of the picketers who are in front of OpenAI. The security guys have cleared them all out now, but if you pull one off the internet, it’s very Luddite-revolt-looking.

Yeah. I think we are, without question, going to have a revolt. How often, and where, and against which companies? We haven’t really started to see job displacement truly occur, but when it does hit—and I think we’ll be seeing significant displacement in the next 2 to 3 years—we’ll need to figure out re-skilling and other mechanisms.

Can I give the positive spin here?

Of course. We have no more looms.

No. It’s the fact that we’ve seen this kind of displacement fear throughout history, and we always survive it very, very well.

On the re-skilling, which everybody is saying, “Oh, my God, how are we going to re-skill?” note that we have AI to help re-skill everybody, right? You can pick your passion and go, “I want to be re-skilled in that area,” and you’ll get a really good potential work.

Note also that we’re near full employment today, and have been for a while, so we actually need a little bit more buffer in the labor market than we have today. Those are my positive spins, and I also understand the negative.

I am curious what rules and regulations the government will put forward. I don’t think it will become a capital crime. I do think we’ll see troops deployed again on things like this.

We’ll just arm the Waymo cars with little machine guns so they can defend themselves.

I love that, right? Fight back.

This was a scary story: 2 children were shot while in a Waymo here in Santa Monica, at 2nd and Broadway. This is probably a mile from my home. Two teens were shot in the arm and torso; the injuries were non-life-threatening, and they’re now stable in hospitals.

I think violence against tech—if you remember all of the scooters here in Santa Monica, Salim, if you were here at that time. People sort of raged against the scooters because they were blocking sidewalks. They would literally throw them into the middle of the street. They’d decapitate them.

I think when we see Optimus robots and Figure robots on the streets, we’re going to find them in various positions, hanging from trees. Thoughts here?

We are, but I think this is really just part of the shift. They probably called the Waymo to try and get away, and there was probably some sort of gang violence involved in this, something like this.

Oh.

Hmm.

Okay. This was the slide about Waymo versus Tesla. Here’s a tweet: “The downfall of Waymo began yesterday.” This was tweeted on June 23rd about the launch of Elon’s robotaxi.

People need to understand that the Waymo car is not cheap. It’s a $200,000 vehicle. It’s got 29 cameras, 5 lidars, and 6 radars, versus Tesla’s robotaxi. Elon has held to his first-principles thinking: “There shall only be cameras.”

This would be a great Dave-Salim bet. Dave, do you have a particular preference on which one would win this?

We’ll bet on this? Yeah.

Peter, you know this topic far, far better than I do. My bet, knowing what I know, is that it’ll be about a 70/30 split. There are always 2 or 3 vendors in any given market in the US. It always stabilizes at that for antitrust reasons.

Tesla will grab the lead now because they put a bigger investment into the neural chips. On the other hand, Google has incredible technology. But if I had to bet right now, I’d say 70% of the market goes to Tesla and 30% goes to Waymo, and it stabilizes.

I’m 50/50 because I always prefer better technology.

Mm-hmm.

But you have the “never bet against Elon” problem on the other side.

Mm-hmm.

Here’s the issue: it’s a huge capital expense that Waymo will need to roll out nationwide. Elon has a different option. You buy a Model Y, turn on self-driving, and it drives you around. You’re going on vacation for a week, so you tell your car to go off and earn you revenue.

Basically, the CapEx is covered by the consumer, and it becomes a revenue engine for you. You’ll buy a couple of these and just have them go out there. So he’s going to populate millions of these cars across the country at no CapEx to himself.

Mm-hmm.

If you apply that model, which is essentially the ExO model—assets on demand, where you don’t own your own assets and let other people self-provision these assets—it will win hands down because it’ll scale much, much faster than Waymo trying to own its own cars, which it has to.

Yeah. This is—

This is directly connected to our other side bet. You remember last week: Elon and Trump finally break up, and now they’ll hate each other forever. I was like, “No, no, no, no, no, no. No.”

First, this might be totally made up. They might just be trying to make PR for themselves. It wouldn’t surprise me at all. But second, if it is real, which it probably is, they’ll kiss and make up in no time. That was my bet.

Yeah. By the way, we’ll have Anthony Scaramucci join us back on this pod to talk about how the administration’s—

He was eerily, eerily accurate last time. After the inauguration, we asked him how long this bromance would last, and he said, “I think 30 Scaramuccis,” or something like that. It was really nearly dead-on, almost to the day.

So we’ve got to respect his views next time.

Yeah.

Yeah. Well, look, the government needs SpaceX launches, and Elon needs regulatory approval of that really good idea: you buy your car and then lease it back to be a robotaxi. It’s just a really good idea.

All right. Let’s hit a couple of other quick ones. This is flying cars and drones. We’ve seen Trump sign an executive order on drones, flying cars, and supersonic travel. For these drone systems to actually work, you need to enable beyond-visual-line-of-sight mode.

We’re seeing the US government support 5 regional eVTOL pilot programs. I’m super excited about Archer Aviation here in LA to serve the Olympics in 2028.

The other point here is that they’re enabling supersonic travel by scrapping outdated overflight bans. The supersonic jets of the Concorde could not fly over the continental US because of the sonic boom. They’re changing that and providing FAA waivers, in particular because a number of companies have come up with mechanisms to actually avoid or absorb the sonic boom.

This is going to accelerate aviation as it should. There hasn’t been that much change in aviation for the last 50 years.

Yeah.

It’s been incredibly slow.

Yeah. I was wondering if all the privatization of space travel was also going to lead to breakthroughs for out-of-the-atmosphere hypersonic travel. I don’t know if you remember Steve Keshi [?]—he was working on that back when we were in school.

Yeah.

I know way too much about this. It’s just really tough—the amount of energy required to build something that’s going to skip across the upper atmosphere. There have been many companies that have died on that mission statement, with billions invested.

The closest thing right now is Starship going spaceport to spaceport, but it’s still expensive.

Mm.

Here’s another competitor we know about. We know about 3 or 4 of the companies out there providing eVTOL services.

This is Wisk. It’s been selected by Miami to launch in that location, so watch out for Wisk in Miami. I don’t know if we want to talk about these eVTOL flying cars at all, Salim, but they’re coming.

What’s interesting is, Dave, when we spoke to Eric Schmidt about this, he was not a believer in eVTOLs.

Yeah.

So I remain hopeful.

Yeah, I think it was a relative thing. There’s so much change going on, and so much of it is so impactful. I don’t think it was like, “Yeah, I don’t believe in this.” It was like, “Look, in aviation, we have helicopters. Now we’re going to have four-prop electric versions of them. So what?”

The so-what to me is that they’re self-flying, self-driving. That, to me, is really, really a big deal, and they’re much safer too. So I do think it is a big deal. But he’s like, “Yeah, but relative to space and AI, is it a big deal?” And he was like, “No, not really.”

For me, this is one of the most exciting technologies we could have. Can I explain why?

Sure, but you’re still driving your Tesla everywhere.

I am, but I love it. That’s fine. I’m allowed to do that. By the way, I have a Porsche Macan Electric also, and it’s maybe the best car I’ve ever driven, but it’s not way, way better than my 2017 Model S. It’s marginally better.

That’s really profound, to say how much further ahead Tesla was than all the other carmakers in many, many areas. But let me go back to eVTOLs for a second.

Yeah.

I think this totally changes the game. Why? Because real estate that’s hard to get to is priced very low. It’s scarce, and therefore we pay a lot of money for that little waterfront property on a lake somewhere because there aren’t that many of them.

Now you can get to all sorts of places you can’t get to by road. A really beautiful plot high up on a mountainside that you can’t get to by car now becomes viable real estate, and we turn real estate from a scarcity problem into an abundance problem. I think that’s incredibly exciting for the future.

There are huge economic implications for this, not to mention that many of us fly around a lot. How much of a nightmare is the damn commute into the city? Just having a drone corridor from Kennedy Airport into Manhattan would change the game.

Yeah, for sure, and that is coming. We’ve seen this with Joby, with Archer, and now with Wisk. You’re going to be able to commute from downtown Manhattan to JFK in one of these vehicles.

Right. In São Paulo, it’s 2 hours to get to the airport.

Oh my God.

It’s like a nightmare.

Helicopters are the only option you have if you’ve got the money.

And those are not the safest.

Yeah.

So these are safe and solid, and I think they’ll start to become a really big deal for tourism destinations. The list just goes on in terms of the broad effects this could have. This is hugely exciting to me.

9. Energy Becomes AI's Bottleneck

I want to jump into the next subject, which is timely and critical: the demand for energy from AI systems and a look at the US versus China. I think people need to understand that we are going to become limited by power in our quest for digital superintelligence. So how are we going to power this revolution?

China is rapidly scaling up nuclear power. It’s their equivalent of a Sputnik moment. China aims to surpass the US by 2030 in nuclear power. The numbers are staggering and incredibly pessimistic here. The US only added 2 reactors this century—2 reactors over the last 25 years. There are 94 total in the US versus 58 in China.

Mm.

China is building a reactor every 52 months. US licensing alone takes 10 to 12 years.

Yeah.

Crazy.

Yeah. You know, this is the reason I love doing this podcast, and I look forward to it so much, because your team will dig up a topic that I really need to study, and then it’ll come into the deck, and I’ll be like, “It’s just so fascinating to me to dig in on these things.” It’s so fun to talk about it.

A lot of these things are so obvious, and this is America’s Achilles’ heel. We cannot act on long-term thinking and long-term investing. Our investment cycle is 3, 4, 5 years at the most. Our election cycle is 4 years or 8 years, and we can’t think about 10 years in the future. It’s killing us. It’s absolutely killing us.

Yeah.

The data in these next couple of slides is mind-blowing.

Yeah, I mean—

I’ll look into it.

We’re going to talk about nuclear in a future podcast with some of the CEOs in these industries. Generation 1 nuclear power plants no longer exist. Generation 2 plants are still out there, including what we saw with Fukushima and Three Mile Island. Those were the dangerous plants.

Generation 3 plants are the ones that have really been manufactured over the last 30 years, and Generation 4 plants are the ones that have currently been designed. They’re fail-safe. They’re the kind of nuclear plants I’d put in my backyard.

But the timeframe for developing these plants is insanely long. We’re talking about 10 to 15 years if you wanted to start today, and that’s why Three Mile Island is being recommissioned, because it’s already approved from a government regulatory standpoint. It’s crazy.

Mm-hmm.

We’re talking about 10 to 15 years if you wanted to start today, and that’s why Three Mile Island is being recommissioned, because it’s already approved from a government regulatory standpoint. It’s crazy.

Mm-hmm. It is totally crazy.

Salim?

I’ve got a— I’m really, really angry about this one. I’m trying to be constructive here, but we know that solar scales. China is going to have more solar than the entire US.

Well, let’s get to that. That’s our next topic here. Let’s go to the next slide.

Oh, wait, Salim.

Salim, you got muted here.

Salim, you muted yourself. I don’t know if you muted yourself or we muted you.

Yeah, I did. I’m just really mad.

That was perfect timing.

I was going to swear, and I thought I’d better cut off the mic before I swear.

All right, so here’s the point. China is winning the race to become a Type I civilization—in other words, a civilization that harnesses all the power hitting the Earth from the sun.

By 2030, China will have the ability to build an entire US’s worth of power generation from solar and storage alone every single year. Look at this chart, Salim. Talk to us about this.

Yeah.

Talk to us about this.

You go, Dave. Let me just gather myself.

This chart is great. I added the gigawatt axis on the left there, because who the hell talks about terawatt-hours per month? What the hell is that metric? Gigawatts are a better way to look at it.

This is actual gigawatt utilization. They created about 700 gigawatts of solar panels in 2024 and deployed 250 gigawatts of peak capacity. Kilometer after kilometer, the hillsides are covered with solar. We covered that on a past podcast. It’s insane. They’re just rolling this out, building capacity and distribution.

Mm-hmm. Yeah. I mean, just for context, the US total capacity for energy production is 1.2 terawatts, so 1,200 gigawatts. Peak actual utilization is more like three-quarters of a terawatt.

You’re like, yeah, half of what the US creates, they created in solar panels in 2024 alone. Just solar panels in 2024.

Yeah.

Crazy.

All right, so they have an advantage because they’ve got all the rare earths, and they can use those for solar panels and build out. They’re going to add more solar than the entire US energy output in a little while, so that’s going to be crazy.

It blows my mind that we are putting restrictions in place and not extending tax credits for solar and other things here in this country. We need to unlock that in the biggest possible way and let the private sector go nuts on this. Put government subsidies there instead of government subsidies on oil, which is what we do today.

Yeah.

It’s the stupidest energy policy we could possibly have.

I’m fine with all energy needs to be made available, but where do we invest in growth? Solar is available today. The numbers are—can I just share a couple of points here?

Yes, please.

Enough energy hits the Earth in 1 hour to provide the global energy needs of the entire year. So 1 hour’s worth of sun hitting the Earth provides global needs for the entire year.

Oh, yeah.

All right? Here are some additional numbers. Global solar capacity reached 1,300 gigawatts by 2024. About 1% of global energy needs is being provided by solar. It’s estimated that if you cover just 0.1% of the Earth—about 150,000 square kilometers—with 20%-efficient panels, that would generate 200,000 terawatt-hours annually, exceeding current demands.

So how big is 150,000 square kilometers? It’s about the size of South Dakota. I’ve never been to South Dakota, but if it were covered by solar panels, it would be giving us a huge amount of energy.

The fundamental problem with solar for data centers and AI is this: the good thing about AI is that you can move it to the power. You can take a data center and move it.

I love that point.

Yeah.

That’s a huge advantage. The bad thing is you need those chips to be running 24/7. They depreciate really, really quickly. They’re very expensive. The chips cost 10 times more than the power. You’re not going to let them sit idle if it’s raining out.

Right.

So solar has the horrible flaw of being intermittent. It’s really good if you can store it, but the cost of lithium to store the solar when it’s cloudy and raining is about 5 times higher than the cost of the solar panels.

So if you want to become the world’s first trillionaire, find a way to store huge amounts of energy cheaply.

Yeah, this is the—

There is a way, and we’re going to be talking to Bill Gross, the CEO of Idealab. He’s been using gravitational storage, which we’ll talk about. Rather than storing energy in batteries, you use a portion of the energy during the day to move a large weight vertically, and this has been done with water for ages. You move a large weight vertically, and then at night, the weight gets pulled down by gravity, and a generator generates electricity. So it’s efficient, it’s available, and it works.

Can I throw out some thoughts here?

Sure.

In 2016, we crossed an inflection point where it became cheaper to build a solar power-generation facility than a fossil-fuel power-generation facility, and almost all energy generation since then has become solar because of that.

In 2019, we hit a more important inflection point, which is that it became cheaper to build and run a solar facility than to just run a fossil-fuel facility. So the CapEx and OpEx of solar are now cheaper than the OpEx of fossil fuels.

Hmm.

Right? That’s a crazy inflection point, meaning we don’t ever need to build a fossil-fuel thing again. We should just be building solar. The utility-storage problem has been solved at scale, as you’ve said. At large scale, this is very easy. You just pump water up a hill to an artificial lake and use hydro at night on the way down, right?

Yeah.

It’s a little clunky, but it’s very workable until battery technology or Energy Vault-type technology that Bill Gross is also doing comes along.

Yeah.

So this is really a known problem. We should be going full out for this. I think it was 100 miles by 100 miles that would power the whole of the US—that was Elon’s calculation. The one I saw was that 2% of the Sahara covered with solar panels gives you enough power to cover the whole world’s energy needs. Distribution is a challenge, but just that visual is a really killer visual. It’s absurd that we’re doing what we’re doing.

Here’s the chart that Elon posted. Elon’s tweet is, “Solar is 100% of energy long term.” No question, this is what’s going to drive humanity forward. Here’s a chart that basically reads: It took 8 years for solar to go from 100 terawatt-hours to 1,000 terawatt-hours, and then just 3 years to go from 1,000 to 2,000 terawatt-hours. You can see that super-exponential growth curve, exceeding hydro, coal, gas, nuclear, and wind. Pretty amazing.

Yeah, pumped hydro is very, very efficient, but you need a lake about the size of Loch Ness to move up about 300 meters and come back down to store enough energy to power a huge data center. So that’s a lot of water.

Well, Bill Gross has solved that, right? With these gravitational towers. Basically, you pump huge, multi-ton bags of dirt up into an elevator shaft in a large building, or up the hillside if you’re near a mountain. He’s got this working today. He’s got huge contracts. We’ll be talking to him about that.

I think storage can be solved. I’m hoping, in fact, AI is going to help us with new technology for solving this problem.

Yeah, that’s exactly what I was going to say. If you’re a materials scientist or a chemical engineer, I’m pretty sure that with AI’s help, in a couple of years—or maybe even this year—you can come up with a reversible chemical reaction that’s completely self-contained and stores huge amounts of energy. Use solar to drive the reaction one way, and then when it’s cloudy, run the reaction in the opposite direction.

Yeah.

And if you come up with something that’s 10 or 20 times more energy-dense than a lithium battery, which seems very viable, you’re going to be a trillionaire.

Yeah.

So just figure that out. Use AI to help.

I can’t.

A couple of years ago, we tried to design an XPRIZE around this, and the numbers were: get off-grid storage 50 times cheaper than it is today.

Here’s another tweet from Elon: “Solar power in China will exceed all sources of electricity combined in the US in 3 to 4 years. It’s a wake-up call.” China is going all in on energy production, and it’s epic.

I have one more thing to say about this.

Yeah. Please.

I think this China–US stuff is a little overhyped. I don’t think we’ll have that much of a strategic military conflict.

I think it’s a stocking horse.

But I think it’s a great comparator.

Yeah.

I don’t think there’s real deep conflict there.

No, there’s not. What it does is show the US what is possible.

Yeah. And to Dave’s earlier point, this is a structural issue we have in the US: We have 4-year election cycles, and we have no mechanism to look out 20 years and say, “This is the water, healthcare, and energy we need over a 20-year period.” We need to solve that structural issue.

Well, and running our venture funds—you know this acutely—you make an investment, and investors want to see liquidity within 4 or 5 years. They don’t want their money to be sitting out there for 10 or 15 years. In China, they know they’ll still be in power, or they think they’ll still be in power, 20 years from now.

If you look at the date they said, “You know what? We need to be the world’s biggest manufacturer of solar panels.” So it’s about 2004 or 2010, or somewhere around there. From that date to having the factories up and running and having the production—and then also, when there’s a recession, you’ve got to keep cranking.

Mm.

So just keep the engine running because you’ve got a 20-year view, a 30-year view.

Yeah.

We just can’t do that structurally in the US. That’s why we get so far behind in nuclear and solar and some of the other long-term trends, and that is our fundamental Achilles’ heel.

All right. Last subject for us, gentlemen: crypto.

Always have to have a little crypto in the conversation. We've seen Bitcoin over the last couple of weeks dip down below $100K and resurrect itself up to $107K. The predictions I'm seeing, Salim, still hold that we might see $200K by the end of this year. I still remain all in and massively enthusiastic. Are you?

Big time. I heard Michael Saylor say Bitcoin will get to $21 million a Bitcoin. I'll be happy when it gets to $1 million a Bitcoin. That'll be perfectly good enough, I think, for most people, because that'll just put it at the level of gold, which is infinitesimal anyway in terms of global asset class.

But I think there's a bunch of things happening that will start to accelerate this. I heard Freddie Mac and Fannie Mae are looking at or approving mortgages backed by Bitcoin. That'll be a huge thing.

Yeah.

So it's starting to get systemic approval, and then things go crazy.

Yeah. For sure. How about you?

Big time.

Yeah. Dave, thoughts?

I'm really excited about the next story, actually.

All right, well, let's go to the next story here.

Dave's holding his powder.

Okay.

Yes. There we go.

10. Crypto Enables Agent Payments

Circle Internet Group goes public and explodes, goes exponential.

Mm-hmm. Yep. Yeah, so, 2 parts to this story. One of them is Jeremy Allaire, who's stuck with this for so many years, and he deserves every bit of his success for just grinding it out. You know, it's always tough to be an entrepreneur. He had to grind it out over a long period of time with regulators all over him, different administrations and people going to jail, and what a persistent story.

But the reason this is so important is because agent-to-agent transactions can be done in either Bitcoin or dollars now. And you can move back and forth seamlessly. The old method, the SWIFT network—the interbank trading network—is great if you're trying to move $1 million to Hong Kong, because it's only about a buck. It's terrible if you're trying to do a penny microtransaction with another—

Yeah.

AI agent.

Crazy.

So this solves that problem. The current pricing model on AI is subscription fees: give me $200 a month or $20 a month flat. But that's crazy, right? Utilization gets throttled because, if you're trying to write code or talk to your AI, it'll slow down every now and then. Why is that? Because there are too many users online. Why can't I just pay for what I use as I go? Well, it's because we didn't have Circle.

Mm.

So the whole AI economy needs to now move to these micropayments, and this is what's going to enable it, and that's why the stock is up so much.

Yeah. I've been texting with Jeremy. I'm super impressed, congratulating him on this exponential growth. The IPO goes out at $31 a share and peaks at $300 a share. Extraordinary, right? These are the IPOs that the entrepreneurial market needs to fuel the opening of the doors wide open. Salim, thoughts?

I want to echo the kudos here for a slightly different reason, which is that one of the big issues in the crypto world has been trust.

Mm.

And you've had Sam Bankman-Fried—

Also.

—and a lot of hucksters, and can you actually deliver trust? Jeremy, over a long period of time, has demonstrated a rock-solid, stable, trustworthy environment, and that's not been easy to do in an environment where everybody else is a huckster.

And in the United States—

And so, huge kudos to him.

Within the U.S. regulatory framework, if you would, which is super important. So Circle Internet Group is a financial technology founded in 2013. It's a stablecoin pegged one-to-one to the U.S. dollar, and it's going to enable, as you said, Dave, my AI agents to go and transact microtransactions. And it's finally going to enable what has been the full promise of the internet—not just data, not just video, not just words, but a financial layer.

Can I say one thing about stablecoins?

Yeah, sure.

It's awesome to have this pegged against the dollar, but once you have something, there are so many natural assets in the ground and out in the world that you can peg a stablecoin against, say, real estate holdings or something like that, and when that becomes possible, as long as the trust again is there, you're going to unlock unbelievable amounts of capital flow.

One of my friends is actually contracted—I can't say the details here—with a government that has large gold deposits underground, and they've gotten a contract where they're going to peg a token against the gold in the ground and not dig it out, waiting until the technology for being able to extract it in a more environmentally friendly way is there. But you know, that's an extraordinary thought: there are so many assets that could be connected to a token or a stable token.

Yeah, it's the missing ingredient, actually, in kind of the trifecta of digital transactions, because Bitcoin has become a great store of wealth. But there's no guarantee that it's stable. The dollar is guaranteed to go down in value, right? We just print more dollars every year. So you don't want to have a huge balance sitting in Circle for a long period of time.

What people will tend to do is park their money in Bitcoin, then, if they want to transact in dollars, move it over to Circle. It's all seamless: do your microtransactions in Circle, then come back to Bitcoin. But what if you want something tied to something incredibly stable, like real estate or gold or whatever? Well, that doesn't exist yet.

Exactly.

So that would be the trifecta of the crypto circle.

Oh, fantastic. That's great. Yeah. Dave, how about you?

Yeah, we're going back to see Kevin Weil very soon in San Francisco. Hopefully, the sooner the better, as far as I'm concerned.

Yeah.

And then our mega event—yeah, OpenAI headquarters. Our mega event on September 9th—Google agreed to host the pre-party at their headquarters.

Mm.

So I'm basically going to be going back to San Francisco over and over. That's why I need that hypersonic plane. I just want to get back and forth a lot. Yeah.