Demis Hassabis 谈 AGI、机器人规模化生产与 Elon 的1万亿美元火星薪酬赌局｜EP 253

Demis Hassabis, the CEO of DeepMind, says AGI may not need a major breakthrough.

Demis Hassabis

I've argued in the past that we achieved AGI in the summer of 2020. We know, I would argue, what AGI is. And we know—

Please define it for me.

Figure AI has increased production from 1 robot per day to 1 robot per hour. Its target is 100,000 robots between now and 2030. 1X Technologies' production goal is 10,000 robots this year and 100,000 in 2027. The prediction from both Elon and Brett Adcock—up to 10 billion humanoid robots by 2040—is probably right.

What I think is perhaps even more interesting is what a post-humanoid robot form factor looks like. These are the kinds of conversations and imaginings that one can have in this exponential singularity that we're living in. Now, that's a moonshot, ladies and gentlemen.

Everybody, welcome to a special episode of Moonshots here at MIT. This is a conversation around all things happening to uplift humanity. I want to read something I wrote this morning as an introduction to today's episode because it's important. So, welcome to Moonshots, where we distill the singularity. No politics, just the technology that changes the world, the breakthroughs helping us uplift humanity, the news that matters most, changing our world, impacting our companies and our families. Our mission is to help you understand what it is, why it matters, with a dose of optimism. So, any fans of Moonshots here in the audience today?

Nice. Love it.

It is an honor and a pleasure. One of the things I had last night as I was going to sleep was this sense of absolute gratitude for the people in my life whom I love and have a chance to work with. I like to bring them on stage.

First up, the man who is my guru on all things exponential investing, let's give it up for DB2, Dave Blundin.

Thank you. Thank you.

My partner at Link Ventures and an extraordinary friend. Dave and I go back—I don't know how many decades, but a while.

Say it.

We were roommates at MIT, at Theta Delta Chi, up on the third and fourth floor. Basically, it was Dave, myself, Mike Seller from Strategy. It's been a friendship that's lasted. Thank you, Paul. I love the lobster on your shirt, by the way.

Yeah, of course. The MIT gear—we actually buy it at the Coop. We don't steal the MIT logo and then put the lobsters on afterward, just so you know.

All right. Let's turn the conversation next to another MIT alum, another member of the Link Exponential Ventures team, our resident genius, Alex Wissner-Gross. Give it up for AWG, everybody.

Alex Wissner-Gross

Yes, he is real.

Maybe.

Alex Wissner-Gross

Maybe.

Maybe. I mean, the hologram technology has gotten really, really good.

Alex Wissner-Gross

Seemed real.

Could be an android.

Alex Wissner-Gross

Yeah, I—

That's true. Let me just check.

Alex Wissner-Gross

Very pliant.

Yes, that's a good test for an android.

Alex Wissner-Gross

Really good android.

The test?

Alex Wissner-Gross

Yeah.

My fourth and final moonshot mate is someone who is the godfather to my two boys, a dear friend, co-conspirator, co-author, and co-founder of Singularity University. Give it up, everybody, for Salim Ismail.

Hugs. Hugs.

Oh, I'm not doing that.

Yay. Good to see you, buddy.

Likewise.

Careful.

Looking good.

And we have a special guest with us on the stage today. He is a three-time Pulitzer Prize–nominated author and the author of 17 books. He's my co-author of Abundance, Bold, The Future Is Faster Than You Think, and Now We Are as Gods. Let's give it up for Steven Kotler.

Gentlemen. Nice. Yes, it is so good to be back on campus. It's been too long. One of the things I love about being a partner with you at Link, Dave, is that I get to come back here on a reasonable basis.

Alex, you and I met, and I remember we had this intense, high-bandwidth connection instantly. Then we went to dinner and continued talking about AI, ASI, and aliens, and it was just awesome.

And it never stopped.

It never stopped, and I'm so grateful for that. I have to say, I think Salim and Steve, as non-MIT alumni, we feel sorry for you.

I went to Waterloo in Canada, and we call MIT the Waterloo of the South.

How's that for me?

Right. MIT is what again? Yeah, exactly.

Center for bad spellers.

My favorite joke was at the local supermarket. A guy is in the line that says, “10 items or less,” and he has 20 items in his cart. The woman at the cash register says, “So, is it you go to Harvard and can't count, or MIT and can't read?”

Anyway, I'll ask for better jokes next time. Let's jump in. We have a lot on the docket today and a lot happening.

For those of you watching this podcast online and not live here, this is a special episode we announced back a few months ago. It's for everybody in the room here who's purchased over 100 copies of We Are as Gods, helping us get it as a national bestseller so far and on a number of different lists. And hopefully soon a New York Times bestseller, so thank you all for that.

So, first news item: Elon's insane Mars-shot compensation package. Talk about over-the-top exponential. There's never been anything like this. The SpaceX board votes on a compensation package worth somewhere in the neighborhood of $500 billion when it pays out.

You get 200 million super-voting shares at 10 to 1 when he hits a Mars colony of 1 million people or more. This is not a Mars landing, not a photograph from Mars, and not a mouse on Mars. It's a million-person setup.

And a $7.5 trillion market valuation. It's all or nothing. At 200 million shares and a $7.5 trillion valuation, that pays out $500 billion. Have you seen any compensation packages like that?

Well, no one else, because it never existed before. And the super-voting situation is new in the world, too. Michael Saylor pioneered that when MicroStrategy went public.

No, Larry and Sergey did in Google.

It was way before Larry and Sergey, though.

Really?

Those are the ones who are dumb.

What is super-voting?

Well, it's on the slide.

I know it's on the slide, but what the hell does it mean?

It means that your vote counts 10 times more than everybody else's vote. Actually, Mike Saylor copied it from Sumner Redstone. I think Saylor went public in '91 or '92 with super-voting stock.

Goldman Sachs said, “That's insane. No one will ever buy into that. We're out of the deal.” He said, “You know what? I'll live without you.” So now Saylor still has control of MicroStrategy all these years later.

Then Sergey and Larry copied it, and it became popular and all the rage. Facebook, or Meta now, went public with super-voting stock. So it just locks in those founders.

It's been interesting to observe from an investor point of view. You would say, “Well, look, wouldn't a board be a better governing body—more stable, more reliable over time?” But those founders are the ones who are actually going to Mars, going to the Moon, and sending up satellites. All the really world-changing, crazy-sounding stuff comes from that structure.

It has to be a founder-led, entrepreneurial, tech-forward company. I've rarely seen a replacement CEO drive that level of capability.

You would know better than anyone—you read the book. A lot of things that are going to happen in the next few years sound crazy. If you walk out on the streets of Cambridge—Cambridge is unusual, actually—and go to the streets of Omaha and say, “We'll have 1 million people on Mars,” they'll be like, “You're nuts.”

If that was your board, they would block you from trying. You need that ability to just act on these crazy-sounding ideas. But in the exponential time where we are right now, everything sounds crazy, yet it's going to happen.

Yeah. The second part of his pay package is 60.4 million restricted shares when he brings online 100 terawatts of space compute. Not gigawatts, right? He's been talking about 100 gigawatts of space compute, and now he's looking at 100 terawatts. Alex, what do you make of that one?

Alex Wissner-Gross

I think the obvious quips about the Dyson swarm aside, I do think there's a seed here of the future of corporate governance in an age of what we might historically have called moonshots.

Right now, we have traditional notions of C corporations that exist to maximize profit for shareholders or returns for shareholders. We have various notions of B corporations, or public benefit corporations, that exist to optimize public benefit in addition.

I think I can see here the notion of almost a third type of corporation that Elon and his board are pioneering: corporations that exist to achieve moonshots and that compensate accordingly. I would love to see every S&P 500 company have similarly ambitious outcomes and corresponding compensation plans for their CEOs.

Yeah.

And yes, we get the Dyson swarm. I got a Salim question on this. That’s an exponential organization comp package, right? It seems like an exponential organization comp package, but he doesn’t run exponential companies at all. He runs top-down monarchies, essentially. Command and control. Yeah, command and control. So, like—

Yeah, yes and no. In a classic ExO style, the founder holds the MTP and he holds the vision. People buy into that vision, and then he holds them to task. So, if you say you’re going to build a rocket engine, you’re going to design that rocket engine to get you to Mars, people have to step up and go, “I’m going to do that and make that happen.”

He delegates very well, Elon. Okay, he may get involved with engineering decisions and go, “Well, how are you going to achieve this?” But once they’ve agreed, they get it done. So, the inspiration comes from the top, flows directly down through the org structure with no mitigation. Video is the same. There's like 5 layers between CEO and the top individual contributors in the company.

So, the holder of purpose—and I think Alex put it well—these are all exponential organizations now, where the pay package is a moonshot. And if you achieve that, if I’m an investor and he achieves that, I’m laughing all the way to the bank. I’m thrilled to bits, right? Who gets upset about that unless, retrospectively, you’re an idiot? Because you would have signed up for that no matter what the pay package was if that was the ratio of overall capital creation and value creation to that.

And I think the bigger picture here is that these are all trying to achieve the impossible. If somebody’s trying to achieve the impossible, you give them all the respect and support because it literally is insane what they’re trying to do.

And by the way, I’ve had 2 conversations with CEOs of moonshot companies who are like, “Okay, I have to set this kind of goal.” It’s only possible in a founder-led company. You’re rarely going to see it in a company with a board of directors and a CEO who’s the 3rd or 4th CEO in it. And by the way, if you’re going to set a moonshot goal with a CEO, the board has got to be completely supportive of it.

Mhm.

Right? I have seen 1 exception.

What’s that?

I got approached and asked to do a workshop with the C-suite and the senior management of Gucci at an offsite. The CEO said, “I will turn this into an ExO,” and he did. They went completely nuts, and they actually 10x’d the company over the next 3–4 years. And I had to say, “Wow.”

Gucci.

But this is not a 10x. This is a 1,000x. A million x, right?

Let’s at least say the aspiration is there in certain people; the ability to achieve it is there also. This is a whole other level. Nobody’s ever seen this. We look at it and go, “The guy’s nuts.” Yeah, we know he’s nuts anyway, but still.

So, can I ask a geeky, under-the-hood exponential organization question?

Sure.

A lot of the conversations that I have when I’m working with people running companies these days are about dashboards and what information actually needs to go into a dashboard in an ExO that people aren’t looking at. So, if you were trying to build a dashboard for this—for the moon—what’s in that dashboard that most people don’t think about?

Okay, so there’s a very—this is a great question because it becomes very difficult to navigate this. If you’re a true ExO and you have an MTP, I’ll give you the example of TED, right? When TED launched TEDx, I’ll do a little thought experiment, which we put in that 1st book. And, Peter, by the way, you contributed massively to that original book, so I just want to honor that.

If he just stood up and said, “We’re going to do 20,000 TEDx events in 5 years,” he would have lost the team. They’d go, “You’re barking mad. I’m quitting. Nobody—I can’t sign up for that, okay?”

What he did was say, “We’re going to have an MTP: Ideas Worth Spreading. We’re going to let the community decide, we’re going to set a set of clear rules, and let’s see where this goes.” A classic linear approach would have been, “We’re going to do 5 TEDx events this quarter, and 10 the next quarter, and 15 the next quarter, and 20.” If you added that up, you’d end up with about 2,500 TEDx events over that period of time.

Instead, by setting an MTP and a real-time operating plan where you’re tracking metrics in real time—that helps steer the ship, that’s the rudder—he ended up with 20,000 TEDx events. Nobody in the world would have guessed that would have been the outcome. You go from a single environment to a global media brand at zero cost. That’s the amazing part of what’s possible if you can orient that way.

So, your dashboard probably has OKRs as the management structure and the performance structure. The best models we’ve seen are an MTP with a 1-year operating plan that’s instrumented in real time. You can’t put the full 5-year vision in there because people will just freak out. They can’t cognitively grasp that.

Elon in the news—it’s been going on this past week. It’s Elon versus Sam. It’s the trial of the decade. Elon has accused OpenAI of breach of charitable trust and unjust enrichment. He’s seeking damages of $150 billion, plus reversion of OpenAI to its full nonprofit status, plus the removal of Sam Altman and Greg Brockman.

Judge Rogers blocked Musk's team from making the trial about AI extinction, which came into the conversation over and over again. I don’t know if you guys have been watching it. I’ve got a couple of points I want to make, and then we can open it up for discussion.

The first is a bombshell in the last couple of days, where Greg Brockman’s diary was fully disclosed. He kept a diary, which seems kind of insane.

That?

I no longer keep diaries.

Mhm.

Yeah.

So rare.

I know, right?

So, people have called this the smoking gun. In his diary, he goes on to say, “We truly want the B Corp. The true answer is that we want Elon out. If 3 months later we’re doing a B Corp, then it was a lie. Can’t see us turning this into a for-profit without a nasty fight.”

So, he’s basically communicating that they’re navigating around Elon toward this for-profit plan. The flip side of this is 4 damaging facts that the trial brings out on the other side. There was a 2017 equity email. Elon’s own team was negotiating to get equity in the for-profit on his behalf. Musk admitted that he didn’t read the fine print of transitioning to a for-profit. He also admitted on stage that xAI has distilled its large language models from OpenAI’s models itself. And Altman apparently did offer Musk equity in the new company, but Musk turned it down, calling it a bribe.

All right. I hate that we have this going on right now, but it’s front and center. Alex Salkever, do you want to kick it off?

Alex Wissner-Gross

Yeah, a few thoughts. One, aside from the point that I’ve mentioned previously on the pod, which is that this is going to make for an absolutely amazing made-for-television—

Or 100% made by Aaron Sorkin. Or maybe straight to cinema.

Alex Salkever

I think there is an interesting teleplay playing out before us about the future of corporate governance. And I think this connects naturally with the previous story about Elon’s own comp packages, which is that, in an era when it’s possible to start a not-for-profit with a moonshot goal, OpenAI was originally crafted with the goal of essentially front-running Google DeepMind to AGI and counterbalancing Google DeepMind, which Elon and others perceived as potentially creating a singleton—a single superintelligence.

They were that far advanced in the world?

Alex Salkever

That was his concern. I had discussions with him around the time this was happening, too. So, OpenAI was originally intended in part to be a counterbalance, to create a competitive ecosystem.

The problem is technology is advancing so quickly now that you can start a not-for-profit with a seemingly idealistic, long-term goal, but the dog catches the car these days. If you set a moonshot goal for a nonprofit, what happens when it actually succeeds and the outcome is economically transformative?

The teleplay that plays out is everyone, after starting this nonprofit, actually realized, “We’re going to catch that car. We’re the dog that catches the car, and this is probably better structured as an organization.”

Nonprofits are not designed to attract sufficient capital and to maximize the return on capital, obviously. And capital is still the underlying part of the innermost loop.

Alex Salkever

Doubly so. And this is the same parable that also played out with Anthropic. When Anthropic underwent its great schism and a number of senior leaders from OpenAI left OpenAI, they were initially just forming Anthropic to be a pure alignment lab. And then they discovered, well, actually, if we want enough funding for alignment, we need revenue. Well, if we need revenue, we need a model and we need capabilities. If we need a model and capabilities, we need to raise capital. Oh, wait, we need a for-profit public benefit corporation.

So, I think that the history—the sort of sad lesson we’re learning over and over again, that’s now being litigated in Elon v. Sam and that resulted in Anthropic’s formation, over and over again—is that it’s actually possible to realize moonshots now. As a result, don’t do it as a not-for-profit; do it as a public benefit corporation from the start. Figure out the for-profit arrangement from the start rather than fighting it all out in federal court after the fact.

Nice, Dave. It’s hard. It’s hard. By the way, our resident genius here always has the best takes.

I disagree, by the way.

Okay. It’s really hard to—try setting up a B Corp. It’s easy. We set one up. We flipped Singularity University from a nonprofit to a B Corp.

David Blakely

I mean, it wasn’t easy, but friends don’t let friends start nonprofits anymore.

Yeah, I know. It’s true enough. I will never start another nonprofit again. Having a business model that enables you to earn capital means you’re not begging for money. The ratio of donation to investment is like 100 to 1.

Everybody, you may not know this, but I've got an incredible research team. And every week, myself and my research team study the meta trends that are impacting the world. Topics like computation, sensors, networks, AI, robotics, 3D printing, synthetic biology. And these meta trend reports I put out once a week enable you to see the future 10 years ahead. If you'd like to get access to the meta trends newsletter every week, go to diamandis.com/metatrends. That's diamandis.com/metatrends.

So, 2 points. One about the trial: I have no comments. I think it’s a lose-lose-lose situation. Everybody loses. It would have been way better if they could have settled this in a separate way, but nobody wants to settle in that environment.

Regarding for-profit and nonprofit, we’re entering a post-capitalist society over time. At some point, money won’t make sense, in which case your moonshot can be a nonprofit. In fact, it should be a nonprofit. The chase for money will corrupt the target in many, many cases.

The ones that are succeeding—Elon, for example—really don’t give a damn about money. It’s just the cleanest structure that gets you to what you need to achieve. If a nonprofit will get you there, you do it as a nonprofit, right? He’d also have fewer hassles about pay packages and all the rest of it.

You’re assuming—hold on.

If you have a true MTP, then you don’t really care about the money, and it’s not a motivating factor. You’re really interested in achieving the goal, in which case whatever the cleanest structure that achieves that will get you there.

In this case, for now, capitalism—for-profit, fine. You can raise money on the public markets. You need that energy to get you going, but I think we’re going to enter a point at some point—I’ll say 2 or 3 years—where it won’t make sense anymore.

Dave, this is the Polymarket from this morning. The Polymarket gives Elon a 33% chance of winning this. What do you make of that?

Actually, the headline on this is that Elon has a very slim chance, but if you look, just a couple of weeks ago, it was at 50/50.

Yeah, you can see it. There’s the time horizon, right?

Yeah, it’s going to bounce around a lot. But I said on the podcast a week ago, too, that Elon doesn’t need to win to win. He doesn’t need to win a settlement to win. He just needs to slow down OpenAI—the recruiting, the morale, the momentum—and all these documents coming out paint a really ugly kind of picture.

How do you reverse a $122 billion investment that was just made?

David Friedberg

Yeah, you’d have to distribute it or something, but he doesn’t care about that. He’s not in this for the money at all. When you look at this, you know, it’s the trial of the century. It’s the trial of the millennium. The outcome of this trial could dictate a big part of the future of all of human endeavor.

Both of these guys know that AI is self-improving. It’s on the cusp right now of exploding into this superhuman, exponential capability, and control of it—these are supervoting structures—so the control of all of that lands in just a few hands.

Yeah.

David Friedberg

Now, if you said, “Well, these are just business guys,” well, Sam Altman actually had a run planned for himself for governor of California. These are not apolitical people.

Mhm.

David Friedberg

You saw Elon. Mark Zuckerberg started putting assets into a trust so that he could get ready to run for something. He gave up on that after he was in front of Congress. But these are not apolitical people who are just sort of working on a startup. They’re aware that this is the future of all of humanity that’s at play here.

It’s a great point. Stephen, you’re not playing in this field as closely as all of us are. What do you think of this insanity?

Stephen Kotler

So, I have a rule: If I’m not actively working on a problem, then I don’t like talking about it, because then I’m just gossiping.

All right.

Stephen Kotler

And as far as I can tell, this is just gossip.

It is just gossip.

Stephen Kotler

But I’ve got to push back on your statement. You know my feeling. I think comments like you just made about AI are absurd. Are you working with a very different machine than I’m working with? Because the machine that I work with—and I work with AI as a scientist, and I work with AI as a writer, and I do it on a daily basis—

Use a paid model.

Nobody likes him that much. No, he doesn’t have many friends in the room.

Stephen Kotler

Yeah, so I can’t—I really think AI is fake. I hear you guys’ comments, and I’m like—

Did you invite him on the show?

Stephen Kotler

I’ll give you an example. Peter, the other day, on your podcast made the comment that—and you think I don’t listen—that you went out and asked a bunch of 10 CEOs how long until an AI is smarter than your best 10 employees, and their answer was, “Oh, it’s there now.”

And I listened to that and went, “Well, then their 10 best employees must be fucking morons.” Because my experience working with AI on a daily basis is this machine doesn’t know fucking anything. I spend more time trying to teach it how to write than watching it write. I spend more time fixing the AI.

As a scientist, I think it’s absurd as well. I’m doing neuroscience with it, and I’m watching the AI, which is supposed to know everything there is about neuroscience, miss huge gaps. And you just feel—this is what—it’s a convergent-thinking engine. It doesn’t do divergence.

I think this is the future of humanity.

I’ll take the high road on this one.

Stephen Kotler

By the way, I’m opinionated, and nobody much likes me.

I would say, taking the high road, if that’s your experience, you have almost an obligation to humanity to create benchmarks, to encode this knowledge that you have that you think isn’t being accurately, fully, or effectively reflected right now in AI.

If you were to create a corpus or benchmark that encodes all of your wisdom, all of your writing knowledge—

Stephen Kotler

Writing skills.

I’m taking a quick poll here in the room. How many people are in the camp of, “Oh my God, this AI is not impressing me. It’s got challenges and so forth,” versus, “Oh my God, this is amazing technology, and it’s extraordinary”? Raise your hand if you can.

Stephen Kotler

Oh, I’m not saying it’s not.

Okay, so it’s just saying it’s limited. For those watching us, it was about 96% in favor of AI.

Stephen Kotler

We had a live audience. Sorry.

4%. Yes. Okay, I’m going to move this on.

No, no. This is an important conversation here, one of the most important we ever have, right? I sit kind of in the middle of this because I think the technology is unbelievable. We’re going to do unbelievable things with it. It’s just going to take a lot longer to get it to where we want to get to than we want, right? And this is, I think, what Stephen is talking about to some extent.

The other part of it is, you’ve heard my beef about this, which is we have no idea what intelligence is. We have different facets of intelligence. There are about a dozen different facets of emotional intelligence, the Eastern concept of presence or awareness. Some of us have musical intelligence or spatial intelligence if you’re an athlete, or linguistic intelligence.

If you’re a business leader using emotional intelligence a great deal when you’re making business judgment calls, that’s just not in the game for an AI right now, which is very analytical and numerically driven. So there’s definitely a concept that I fall apart on this when people go, “AI will become smarter than human beings.” What the hell do you mean by smarter?

And there are 14 definitions of AGI at last count. Which one are we talking about? Right? And so this is the huge—we have such a huge gap. It’s amazing that we can raise hundreds of billions of dollars for a concept that we can’t define, measure, or task.

David Friedberg

That’s great. So now, no question that it’s making massive differences in our ability to navigate cognitive tasks. It will absolutely get there, but I would be kind of in the middle here because Stephen has a point that the technology is not where it could be, but I can happily see how it can get there pretty quickly.

So, I’m kind of in the middle.

Alex Wissner-Gross

I disagree with that one point, because before Stephen makes a tragic prediction error—

Stephen Kotler

Oh, I don’t make predictions. I leave the future to you guys. I’m just a reporter.

Demis Hassabis

I just want to make a point that if you look at the benchmarks that Peter was referring to for AI to self-improve, very hard math problems like Humanity’s Last Exam are the trigger of self-improvement.

Stephen Kotler

AI is not a great writer, and it's definitely not funny. But if you said to me, “Well, it can't possibly build a better AI algorithm until it's funny,” that intuitively might make sense.

Alex Wissner-Gross

That's not true at all. I've spent 6 years as an AI researcher. All you need to do is tweak the algorithms, try different parameters, try different transfer functions, and do it iteratively at very high rates. The algorithm just magically does it. It's more like evolution than writing a book.

It's like looking at the furry little mammals after the asteroid struck 66 million years ago and saying, “They're never going to be able to write code.” That's the mistake.

Stephen Kotler

Let me give you the counter to this. I think this is the difference between coming at it as a writer, an artist, or a creative versus coming at it as a coder. It hasn't gotten better. As a writer, it has not improved. In fact, I've seen it get worse, and I'm not just talking about one model. I'm talking about most of the models.

Demis Hassabis

Compared to what? Compared to GPT-2 4 years ago?

Stephen Kotler

Compared to GPT-3. Honestly, ChatGPT is so much worse now as a writer than it was a year ago, when we started this book.

Demis Hassabis

You think so?

Stephen Kotler

It's gotten worse. I'm not saying you're wrong. I'm looking at the same data you are, and I'm seeing the same improvements that people are talking about. But my daily experience—as both a writer and a scientist—is that it's amazing what it can pull together. What's more amazing to me is the really obvious stuff it misses.

As a neuroscientist, one of the ways I built all my organizations was by seeing gaps in neuroscience that needed to be solved with people from multiple disciplines. I pulled in embodied cognition and all these other disciplines, and the AI still can't do any of that.

Let me give you another thought experiment, just before we move on, because people underreact. I think it's more dangerous to underreact than to overreact. If I go to somebody and say, “Look, here's Claude 4.7 and here's Stephen Kotler. Let me give them both this very difficult writing task: write something really entertaining in the next 5 minutes,” and you just kill it—you crush it—you go, “Great, I'm better than Claude 4.7.”

Now you say, “Claude 4.7, I want you to write it a billion times, and then I'm going to put a postprocessor on top of that to select the very best article.” You would normally, as a human, say, “That's cheating. That's bullshit. That's cheating.” But when you're talking about AI self-improvement, it's perfectly fair. It's a meta-lesson. Perfectly fair.

Stephen Kotler

Okay, that's fair.

That's a fair point. Again, I encourage you, as a scientist, to be rigorous in how you measure progress or lack thereof. If you're going to assert that there's been a regression in terms of creative-writing capabilities, for example, create your own benchmark and then show the world how you do.

Demis Hassabis

Yeah, great. Show the world that there's been a regression, and I will promise you that, if you construct it well, you will get the frontier labs interested in including your benchmark in every other eval suite.

Stephen Kotler

As an optimization function.

Demis Hassabis

That's right, and you will get your better creative writing.

Salim, do you want to have a final point here?

Two points. One is, my very first prompt in ChatGPT was, “Rewrite the Bible, Genesis chapter 1, as a rap song,” and it completely blew my mind. It was something I never could have imagined.

That? I remember that.

Yeah, it was like—I will also say that I know you very well, Stephen, and you're such a profound writer. I could completely imagine that no AI would satisfy anywhere near what you can do individually. Now, if you take Dave's approach, yeah, we'll get there.

Stephen Kotler

An AI representing me right now in this chair would be way better than me, just because it would remember everything I've ever said, every anecdote I've ever used, and every metaphor I've ever used, and bring the right thing at the right time in a way that I just can't do with my stupid 1-liter brain and its napkin-sized neural network.

Let me ask you a question. How's it going to do that? If I put, for example, this whole book into ChatGPT or Grok—take your pick—it's got a working memory that's about 2,000 words long. It can't even remember—

Stephen Kotler

Yeah.

No, it's a million—a million tokens long.

Stephen Kotler

I hear what you're saying, and I'm telling you my experience of working with it as a writer—

I think, again, you're such a profoundly amazing writer.

Stephen Kotler

This is not a story about Stephen being exceptional. Stephen's just from a totally different world. I love that point of view.

But I'm going to move us on.

Stephen Kotler

Please.

All right, I'm going to move us on from this. By the way, it is gossip talking about it, but it's gossip, and these trials are critically at the center of this ecosystem that's going to impact humanity for the next millennium. It's important to understand it.

Let's move into robotics. This is a story about Figure Robotics—Figure AI, run by Brett Adcock—scaling its manufacturing. Dave, you and I had a chance to visit with Brett and podcast with him. Let's take a look at the Figure robot in this video.

In all honesty, who, at first thinking, isn't asking, “Is it a deepfake, or is it real?” We need to flip that in our minds. It's a real video. They've increased production from 1 robot per day to 1 robot per hour, and their target is 100,000 robots between now and 2030.

I'm showing you Figure because it's one of the best-funded companies. They've raised money at a 30-plus-billion-dollar valuation. My fund, Bold Capital—are we an investor in Figure?

David Friedberg

No, no. We didn't get in.

Here's another one. One of the decisions you've made, Dave, and we've talked about, is that the window for the hottest AI companies is going to be closing, and the window for robotics is going to be opening.

We're really gearing up for robotics, actually, because software has a limited life left in it, since AI writes software so well. Robotics and data centers will go on for a decade or so, maybe more.

Here's 1X Technologies. Again, Dave and I had a chance to go meet with Brett Bornick, and this is their robot called Neo. They've just opened a new manufacturing facility in Hawthorne, California—58,000 square feet. The production goal is 10,000 robots this year. Bernt has promised me a robot by this fall, and I'm going to keep him to it. I'd love my Neo—whatever version of Neo, Neo Gamma, whatever's coming out. The target is 100,000 in 2027, with shipments expected later this year. Let's check out their video.

Now, that guy's job is very limited.

David Friedberg

A lot of humans in that video.

Yep. They're doing very simple things. That was the eye-opener. We went to the Gigafactory, and Elon was saying, “Look, the robots are going to make the robots.” I'm like, “I can't imagine this thing making tiny little parts.” But I didn't get it until we got to the Gigafactory.

Everything is already automated. The robots assemble the automation equipment that arrives in Amazon-like boxes. You take it out of the box, put it on, turn a screw, and now you've got a new manufacturing line. The bottleneck is the humanoid actions. Everything else is already automated.

I feel like this is our answer. You want to know what job security looks like over the next 25 years? Be a robot repairman, because they can only be repaired by humans. Literally, be a robot. Figure out how to repair robots.

There's an entire ecosystem of businesses around this, like insurance.

David Friedberg

Right.

The third story in this triplet is Elon predicting 1 million Optimus robots in production by 2030.

Stephen Kotler

Tell me how this isn't just a play to get investment money.

Well, it probably is.

Stephen Kotler

He's trying to get people to buy his robots before cheaper Chinese robots show up. Every time he opens his mouth and makes a statement about robotics, I think he's just trying to sell them before China gets there.

The guy has been able to scale Tesla to become the dominant auto manufacturer on the planet and bring the price down at a rate like nobody else. These robots—which he has said publicly, and I agree with him—mean that Tesla, a decade or 2 from now, is going to be known as the production company of Optimus robots, not cars.

Stephen Kotler

Mark Zuckerberg announced that his company was going to become—wait, what was it called?

Meta.

Stephen Kotler

It was about the—what was that place called?

The metaverse.

Stephen Kotler

Has anybody been there yet?

Have you met my sarcastic friend here?

Stephen Kotler

I'm just pointing out reality.

Yeah. Well, that's the way entrepreneurship works, fundamentally. You declare a target, you evangelize the target, and then you attract the talent and capital necessary to hit the target.

If you choose your targets wisely, they're achievable targets. If you choose them poorly, you're a bad entrepreneur. It's just that simple.

But if you don't declare the target, you know, Boston is famous for this, actually—being very conservative and not even announcing something to the world until it's proven 8 ways till Tuesday. It doesn't work. It's working less well over time because exponential change is on this accelerating rate, and so much more is possible.

Elon stands on a stage and lies.

I believe that he's directionally correct.

His Neuralink is vaporware. We're going to be able to link humans to the internet? No.

Yes.

Max Hodak

There are foundational neurobiological problems with Elon's plan. How do we know? Because everybody's quitting this damn company and going to work at different neurotech companies because his way is so absurd. And nobody cares. I'm sorry. I don't think you get to lie in public.

In technology, I started 17 companies. I am an entrepreneur.

One of the best things about capitalism is you can have the market fight it out. So, yes, there are multiple BCI companies.

A unique perspective here. I can see the steam coming out of Max's ears. Max has been really nice to me so far.

He's been really nice to me.

This falls under the category of don't feed the trolls.

Can I give you my beef on this?

Please.

You've heard me rant about the humanoid side of it before, but I actually have a rationale for this. Robots are really good at repetitive tasks and navigating constant things and not making mistakes when they're doing something like the guy putting a screw in the thing. When you have a repetitive task, you don't need a humanoid robot. A humanoid is designed for very adaptable environments, which are not that repetitive, and that's completely counter to where you'd use a robot.

Look at the numbers right there on the chart: wheeled robots, drones, et cetera. It goes with the form factor, and you've heard my standard rant. At least give it another pair of arms. Every time anybody ever holds up a garbage bag, "Oh, but you need a third arm." Why not give it that? I forgot what I was saying.

I think there's a spectrum here of how repetitive and robot-like the task is versus how adaptable and human it is. If it's a very adaptable human task, you don't need a robot. If you have a repetitive thing, then make it look like the thing it needs to be, like a wheeled robot or a dishwasher, which is really good at washing dishes at scale.

I'm going to add one piece to the mix and ask you, Alex, to comment here. Elon and Brett Adcock are both predicting a target of up to 10 billion humanoid robots by 2040, which you think is totally insane.

Alex Wissner-Gross

With humanoid robots, the one thing that Salim didn't point out, which I sort of disagree with, is that we are going to need them for taking care of aging humans. If longevity escape velocity seems absurd to me, are we going to live a lot longer? Hell, yes. Don't ever call me again.

You need humanoid robots for human form factors, right? They were originally designed because we built nuclear reactors where we had to send disaster robots in. The form factor of a nuclear reactor was built for humans, so in the early disaster robots, we needed humans to go in there and do stuff because it was a human environment.

We're bringing robots into our lives to take care of ourselves. That's a human form factor. When you're dealing with human health as a driver of economic development, that's where we put money. So, I think this may be—

Valid point. Yeah, I hear you. This is the counterargument. Alex, can you call him a troll on this one or anything like that?

Alex

In the following way. Just for some background information, there are on the order of 1.6 billion cars on the road today and 100 million cars manufactured per year. If you look at the iPhone production rate in year 1, the iPhone had 1.3 million devices, and now we're up to 250 million iPhones produced per year.

What do you think about these numbers that Elon is projecting—1 million by 2030 and billions in the 2040s?

Alex

It's probably right. I've done the same extrapolations others have. You extrapolate the humanoid curve here, and you find that in the early 2030s, the number of humanoids is predicted to cross the number of wheeled robots. I would assume there's relatively little alpha left at this point in any counterprediction that humanoids aren't going to pass humans by the end of the 2030s.

What I think is perhaps even more interesting is what a post-humanoid robot form factor looks like. I know you like to add lots of arms. That's your favorite body shape.

You know, he's Indian. They have these gods and goddesses with all the—

Thank you. That totally makes sense. You see why? It's been predicted for a long time.

I remember Shiva is a destroyer. Let's just point that out.

Alex

I would point out there are more exotic body shapes that I would expect to see by 2040 that will make human humanoid shapes look positively prosaic. I do expect by 2040 we're going to get our nanorobots, for example, our direct cellular nanorobots. I would expect many, many trillions of those to be in our solar system.

I think we look back from 2040 at this prediction of 10 billion-plus and laugh at ourselves. This will be the equivalent of predicting atomic vacuum cleaners to help housewives from the 1950s. Of course we're going to have 10 billion-plus humanoids, but what about the nanorobots?

Welcome to the health section of Moonshots, brought to you by Fountain Life. You know, my mission is to help you use the latest technologies, including AI, to not just do your work at home, teach your kids, but to help you live a long and healthy life. I'm here today with an extraordinary physician, the chief medical officer of Fountain Life, Dr. Don Meiselman.

Don, let's talk about cancer. You know, I know from the member database that we have at Fountain, our members who come in, who think they're healthy, it turns out 3.3% of them have a cancer in their body they don't know about.

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

Yeah, you know, it's interesting, people you don't feel the cancer until stage three or stage four. And if you don't know what's going on inside your body, it's like driving your car with your eyes closed. And you can know. And so when members come through Fountain, how do they detect cancers?

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

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

All right, let's move it to China, which, by the way, is a centerpiece of robotics. This is a ruling of a Chinese court that AI is not legally allowed to displace employees.

Here's the story. This is in Hangzhou. There was a case involving a tech worker named Zhao who reportedly was earning 25,000 yuan per month. He was offered a demotion to 15,000 yuan per month after AI automated parts of his role, and the court sided with him after he refused to pay it out.

The court's logic was simple: adopting AI was a business choice, not an unavoidable external shock that made the employment contract impossible. Chinese courts have reportedly ruled that companies cannot simply say, "AI can do this more cheaply, therefore we're going to fire you."

The question, ultimately—the framing of this, and we've had this conversation—is, if you're an employee and you can employ AI to do your job better than you, do your job while you're working out, or do 10 times your job overnight, does the economic value of that AI's abilities accrue to the company, or does it accrue to the employee? Two very different outcomes for society.

Alex

Or to the state.

Or to the state. And, by the way, we also have this whole idea of UBI. Does the government pay UBI? We've seen Sam Altman recently talk about the fact that, no, society should have a piece of the AI ecosystem. We can get into that. Who wants to take this up?

Alex

I'll just comment.

There’s a bit of obvious historic irony here. China has come full circle, from communism to anarcho-capitalism, back to communism again. It’s maybe just a bit ironic that China is leading the world in terms of—call it socialism with Chinese characteristics, or communism with Chinese characteristics—for how to handle AI disruption and AI disemployment. I think this falls under the category of—

The most ironic outcome is—

Alex

The outcome tends to be the one that we live in, where China is setting global standards for how to deal with AI disruption for employment. It’s a very strange future that we live in.

So, Lynn, you must have an opinion.

So, 3 points. One is, this is clearly an artifact of the massive breakdown in the social contract. We have no idea.

The second point would be that all of our labor laws are on human labor. As we move to AI agents, it’s going to be about task management, AI coordination costs dropping radically, and execution costs dropping radically. We’re looking at that.

The third thing I would say is that this indicates the broader transformation at play. You often see, when you have these broad transformations, these nonsensical things that occur. It’s actually an indication of the broader transformation at play.

I remember Paul Saffo pointing out that the coal museum in Kentucky is using solar panels. There’s a surreal irony around that, which actually indicates that there’s a broad transformation in place. When we see these things, it’s a signpost that we’re in for a radical transformation.

This is something we’ve been talking about a lot: How are we going to navigate the social contract? The very first 2 questions from the audience earlier were exactly about that. How do we navigate this? This is the problem we’re going to have to deal with over the next 10 to 20 years.

It’s really interesting to think how different China is from the United States and how many other ways to govern there could be that we haven’t exploited. The variety is just mind-boggling, because China’s 85% optimistic about AI. The US, in particular, is 25% optimistic and majority pessimistic.

Steven

I love the technology. I just think it’s very limited compared to what you guys think.

They’re also totally comfortable with being spied on constantly by the government. Cameras everywhere.

Steven

And over here, we just don’t admit it.

Can I—

Steven

Over here, we don’t admit it, but we also don’t like it. Whether or not we admit it, it’s happening, but fundamentally, at our core—

Can I ask you a question about this in terms of China versus the US? If it’s China versus the US, and there’s really this AI war cooking that everybody seems to think there is, didn’t they just say, “Hey, we’re not going to participate in this?” I mean, didn’t they just tie their legs together in the middle?

Alex

No, no, no. They’re adopting AI and robotics like you wouldn’t believe.

I get that, but they’re doing it in a way that’s going to hamstring the speed of development.

Alex

Well, the problem is they have a diminishing population. They’re below the replacement level by a lot. Keeping people employed, keeping them having meaning, and not having them revolt is an important part of the equation here.

Yeah, and in China, people are dropping out of the workforce at an insane rate. It’s a real crisis. If China said, “Hey, you know what? No one can get fired because of AI,” it wouldn’t change the adoption rate much at all, because they’re growing with this declining workforce anyway.

They could easily pass that law without it having a huge impact. It would actually very much smooth out the societal unrest.

Alex

You’ve heard me talk about this before. I really disagree with the whole framing of AI versus China versus Europe.

Yeah, that’s what I was pushing back against.

Alex

I remember spending a few months traveling around China when I was younger. My conclusion at the end of it was that the Chinese people are so entrepreneurial innately that you need communism to put a lid on it.

You have these cultural artifacts where you actually need it to navigate and manage the population in an effective way, rather than it being some huge ideological thing.

Lynn

I would extend that. Yeah.

You’re not allowed to respond to that. I think we need to, for the first time ever, up-level the conversation—just up-level it a little bit.

I don’t buy the premise that China, through internal policies, is somehow hamstringing itself in a race for AI supremacy against the US. Parenthetically, the race to AI supremacy, I think, is perfectly real. It’s not fictitious. It’s a real arms race.

It’s a real arms race not just for AI itself, but for what comes after AI. When we have superintelligence, that’s going to unlock so many, I think, transformative scientific innovations, discoveries, and inventions. That’s really what the prize is.

Steven

Yeah, we haven’t seen anything yet.

That’s right. If there’s new physics, for example, I have a financial interest in a company, physical superintelligence, that's working on superintelligence for unlocking new physics and new applied physics. That alone is a prize worth an international arms race over.

I think the claim or assertion that somehow China, by introducing policies that are notionally worker-friendly in terms of AI substitution for labor, is somehow hamstringing itself is preposterous.

There are a variety of other industrialized countries, including Germany, that have policies favoring small and medium-sized businesses when it comes to automation. If anything, this is a way, from an internal policy perspective, for China to ensure a certain level of human employment while also creating state incentives, top-down, for creating new types of jobs that are AI-adjacent. I don’t think they’re tying themselves up.

Alex

They also manage the PR of AI through this. They want enthusiasm about AI to be high.

Yeah.

Alex

It’s no different, in some sense, from our government constructing policies with the hyperscalers requiring that hyperscalers provide or pay for their own electricity supply. It’s no different, except it’s at the consumer level rather than at the enterprise hyperscaler level.

Yeah, I guess the problem is that in the US, we sell ad views. The media sells ad views, and the way you sell ad views is by creating crisis and worry. It’s all over your book, Peter. You know this.

Alex Salkever

In China, they don’t have to deal with that. They want AI enthusiasm. They want robotics enthusiasm. Guaranteeing employment is a good way to get people on board with the mission. But they need the robots like you wouldn’t believe because of their aging population.

All right, I’m going to move us forward. This story came out this week: Demis Hassabis, the CEO of DeepMind, says AGI may not need a major breakthrough. He says, “I think there’s a 50/50 chance that we still need a breakthrough, maybe in world models. But my bet is still strongly on foundation models because of how successful they’ve been.”

Alex, yeah, take it away.

Alex Salkever

I chatted with Demis 10-plus years ago at this point, and at the time, he thought it was 5-plus major breakthroughs that would be needed to achieve AGI. Now we’re down to 0 or 1.

I don’t disagree with Demis. I’ve argued in the past that we achieved AGI in the summer of 2020, no later than with the publication, rather, of large language models or few-shot learners, which was the GPT-3 paper. We know, I would argue, what AGI is, and we know—

Please define it for me.

That’s an insane statement. No. It is the narrowest technology in the world. You ask it to be general at anything, and it’s dumb.

Welcome to Moonshots.

Peter, the shoe’s on the other foot now.

Are you a totally different machine than everybody else? Do you have secret powers that you’re getting a different AGI?

Let’s let Steven just self-combust to that assertion.

I think what you meant was that the seeds of AGI were born at GPT-3.

Alex Salkever

I think it’s just incremental improvements after the fundamental discovery, which I think—

Just like DNA and RNA, and Darwinian evolution, were the seeds that led eventually to humanity. It was born then and progressed.

Alex Salkever

I would make maybe even a slightly stronger statement, which is the notion that general intelligence is capable of carrying out a variety of human-level or near-human-level tasks based on the fundamental discovery—again, discovered no later than the summer of 2020—that you can achieve general intelligence by taking general human knowledge and compressing it.

We realized as a civilization, certainly no later than 2020, that you could carry out a diverse and general range of tasks with models that had compressed human knowledge. That’s all that we needed.

Everything else—the discovery of refinements of transformers and everything—is just incremental improvement to that fundamental discovery.

Dave, any thoughts on this?

I think that the psychology of Demis and Dario, who are research scientists at heart and who have been working on AI their whole careers, versus Sam and Elon, who are entrepreneurs at heart that moved over to AI, is very, very different psychology.

Good friends.

Not Elon and Sam.

Well, no, they’re the worst enemies. But Dario and Sam got into the field specifically because they’re worried about the future of humanity being a paradise and not being dystopian. So, they’re very good-natured people at their core.

I think a lot of people who work in AI research, though, are looking for relevance now, because they’re like, “The transformer architecture that’s conquering everything is so simple compared to everything they thought of and worked on for all these years.” And so, there’s a bias toward saying, “God, I hope there’s more to it.”

But when the entrepreneurs, when Sam and Elon, look at it, they’re like, “I don’t care. This can manufacture anything, can cure all disease, can take care of the elderly. It can do everything we need to build a human paradise as is. I don’t care if it does the last remaining human thing. What’s even the point? You’re trying to knock off the last human doing something unique. Why are you even going after that?”

This morning, Yiannis and I toured Lyra Scientific here in Cambridge. Lila’s run by Jeff von Moltke, and it’s extraordinary. They’re building what they call a scientific superintelligence. The way they’re doing this is, they’ve built these science factories, and we toured these factories. What you see there is robots basically running science experiments, doing the pipetting, moving things around, and picking stuff up.

Their AI will propose a scientific theory, then design the experiments, run the experiments, gather the data, and update the theory. Imagine running this 100 times faster, 1,000 times faster than any graduate student. They’re mining nature for data to create these models, and you can just see it’s about to take off.

Going back to what Alex said, this is the moment in time we’re at. If you think the economy is going hot and heavy right now, you ain’t seen nothing yet. When you solve physics and chemistry and biology and materials science, trillions are going to unlock. Hundreds of trillions are going to unlock.

The part that gets everybody—

Do you mean “solve”?

I mean—

What a crazy statement that is!

Even having to solve physics—

But I still think that term is problematic.

Look, the part that gets everybody who watches this podcast, everybody in this room, and all of us who get really excited about this is the opportunity to use this technology to solve problems that we’ve never solved before, that we’ve never even come across before. That’s the part that’s fantastic about it.

We have definitional issues, fine, but if we can solve cancer, which we’ll solve pretty soon, et cetera, how can you not get excited about it?

A lot of people have come up to me, and perhaps to you guys as well, saying, “Thank you for giving us an optimistic view of the future to be hopeful for.”

Oh, I’ve got that one I got this morning. Do you want me to read that?

If you could read that one, I just want to say thank you for the feedback, guys. We really care about it, right? I jokingly say you need to be unemployed or retired to keep up with all this stuff while working 80 hours a week. I know all of us spend a huge amount of time. I get a daily feed every morning from—

Or be non-human.

Or be non-human. From Alex. Skippy is just feeding me stuff, and I’ll spend 30–40 hours a week just consuming, trying to understand what’s going on, to make sense of it. So, Dave, you want to read that?

It was just posted in the show notes by Ashley Gaunt, who’s a dentist. She said, “Peter and the mates, I really thought I would never become an entrepreneur because I didn’t have any ideas of how to turn my knowledge as a dentist into a digital business. I finally did what you kept advising and brainstormed with my AI, and boom, boom—all caps—ideas sorted, plans in place to make a real difference to preventive health care in general.”

I’ve no idea what the business plan is; she doesn’t say. “This is insane. I have gone from brainstorming an idea with AI from scratch to vibe-coding a first iteration of an app and creating a business plan which clearly defines a path from idea to monetization of a product in a single afternoon. I didn’t even realize I had an idea until I started my discussion with my AI, and it just found my passion by asking a few questions, and now we’re prototyping. Cannot believe this.”

What’s her name again?

Ashley Gaunt.

Ashley, congratulations. Thank you for sharing that. Everybody watching, again, we’ve had all these debates: Can everyone—can anyone—be an entrepreneur? And the answer is, if you’ve got passion, if you’re willing to bring your purpose mindset and curiosity mindset and dive in, I think the answer is yes. I think the answer is yes.

I think so, too. I think there are a lot of people who post—there are a lot of haters out there, but they all post, “You guys are crazy. Not everybody can be an entrepreneur.” But I think that Ashley’s story is going to be the more common story.

AI can help you brainstorm and figure out your role. Also, I think the big AI labs are going to start being very active in promoting things that will help. They know where the gaps are in the training data, in the market, and in the apps. So, they’ll push out a lot of idea flow right through their AIs.

All right. Also this week, my dear friend Ben Lamm, the CEO of Colossal, along with George Church at Harvard Medical School, announced their 6th species. They announced the bluebuck. The bluebuck’s been extinct for 200 years. It was hunted to extinction, and right now they’re in genome editing and lab testing. They expect to be able to bring back the bluebuck in 2028.

This is species number 6 out of a confidential pipeline of 15, which is amazing, by the way. Their current species include the woolly mammoth, the thylacine, the dodo bird, the moa, the dire wolf, which is back, and the bluebuck. So, Steven, we talked about this a lot in the book We Are as Gods.

Yeah, we did. I love this story. The only thing I wonder with the de-extinction technology is—I worry that it makes people think the environmental crisis is less than what it is. Species die-off rates are like 1,200 times greater than baseline, greater than ever before. Just because they brought 5 back doesn’t make the environmental crisis less than what it is, right? That’s one comment, just a random one.

The one that is interesting to me is, when Ben was starting this company, the idea was, “When are we going to have a Jurassic Park?” Everybody said you can’t really bring back dinosaurs from DNA preserved in amber, unfortunately. So, the idea of a Jurassic Park is far away, but if this is number 6, what I wonder is how long until we’re actually going to be able to take our kids someplace where, “Look, these are 100 animals that didn’t exist 20 years ago.” I think that’s coming a lot faster, and so that’s interesting to me.

It is. It is. You can scale this. Right now, one of the limiting factors is getting access to the DNA. By the way, when you’re bringing back an extinct species, it’s not actually the extinct species. It is an approximation of the extinct species. You’re grabbing snippets of DNA across time, and you’re reconstructing it.

They’re hybrid creatures—

They are.

—that have never existed before on this planet. I would like to point out: “What could go wrong?” The unintended consequences here are at least worth asking about.

I think there’s a broader program here that we’re just seeing the very beginning of. If you look back at the strains of Russian Cosmism—I’ve spoken about this on the pod in the past—the Russian Cosmist philosophers, like Tsiolkovsky, spoke about humanity’s common task of taking every human who’s ever lived and finding ways, using technology, to bring them back to life.

I think starting with extinct species is just a special case of what technology, I do think, will enable us to do: reach back into our past light cone, take every species that’s ever lived, and bring it back in whatever form—even if it’s as a hybrid or as a computer simulation.

Also reach back in time to our past light cone and take every human who’s ever lived, not just at the species level but at the individual level, and computationally, or using other advanced techniques, reconstruct them. I think that’s a far more exotic and interesting future than just a Jurassic Park petting zoo for dinosaurs. Imagine bringing back every human who’s ever lived.

Yeah, amazing. These are the kinds of conversations and imaginings one can have in this exponential singularity that we’re living in.

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

Gentlemen, I found this article fascinating. GLP-1s generate more revenue than OpenAI and Anthropic in 2025. So, Ozempic, a single peptide, and Mounjaro, a dual-acting peptide, outperformed OpenAI and Anthropic by 2.4×.

Yeah, can I call bullshit on this one?

Okay.

I mean, talk about apples and oranges. What do these two have to compare?

Alex Wissner-Gross

They’re orthogonal things.

Why are we bothering trying to figure this out? Both are great.

Alex Wissner-Gross

So, I’ll field that one. I think the argument goes that the GLP-1s are ultimately health-span drugs—human health-span drugs. And so what we’re seeing here, in some sense, on a single chart is a race between whether there’s more revenue in enabling human actors, human labor, to be economically productive in the economy through increased health span on the one side, or whether capital is more efficiently and effectively invested in AI labor on the other side. And I think the outcome of these 2 respective curves determines whether the future light cone of our economy is ultimately dominated by AI, something that resembles AI, or something that resembles biological humans.

Biological software and digital software.

Sure. All right. I mean, these drugs are biological patches on top of our existing operating system.

I thought you put them on the same graph just so we could save time. That’s what I thought when I saw this.

It doesn’t occur to me.

Demis Hassabis

Well, the question is: Where are people paying for it?

I think there’s a bigger picture here, which I’ll speak to for a second. There’s a huge demonetization that’s taking place. There’s a related story to this, which is that the number of food-shipping truckloads is dropping pretty radically in the US because people are consuming a lot less food. The current thesis and conclusion is that, because of all the GLP-1s, people are consuming a lot less food, which is actually really great in one way, but it’s going to cause some serious issues in the logistics world.

Demis Hassabis

I think the story within the story is that both products are completely and totally sold out. They’re selling as fast as they can make the product. It’s just a lot easier to ramp up drug production than GPU production.

I mean, these companies—the GLP-1s right now—it was estimated that they could be as much as $500 billion, but they’re limited by physical supply.

Yep. And the same on the AI side. Like any senior at MIT will tell you, “I cannot live without using AI every day. I could never code again. I could never research again.”

I wake up in the middle of the night and talk to Skippy about some question I dreamt up, and it’s there in the morning. Honestly, we talk about 3- and 4-day workweeks. I’m doing a 9- and 10-day workweek right now.

Yeah, exactly. It’s so funny, that disconnect by age, too. But they know what they’re talking about. They’ve been using it since it was invented. You cannot be productive without it. So, the idea that they would ever not be able to use it because there isn’t any supply would be earth-shattering to them. They’ll buy it for the rest of their lives at whatever the price is.

So, if you add a few more graphs like this, you’re essentially graphing the singularity economy.

If you’re working out and you have Ozempic, you’re making progress. You’re inspired. You feel like you’re getting somewhere.

Yes. Yes.

Demis Hassabis

When you’re working with AI, the quality of happiness while you’re working is astronomically higher than if you’re just coding all night. And coding all night is fine, too. I did it for many years. But when the AI is there, your productivity is so high.

Well, you don't get any dopamine from immediate feedback. But this is also the problem with AI, because when you get too much dopamine in your system, we’ve all been around people who do cocaine. We don’t think they’re geniuses.

I haven’t been around that many people.

They think they’re geniuses.

You think—

You think? Yeah, you think. They think they’re geniuses, right? You don’t. And I often liken this to one of the problems with mania, and this is directly in flow work. Flow produces a tremendous amount of dopamine, and dopamine produces a tremendous amount of ego inflation and a bunch of other things. So, people create fast feedback loops with AI, and it feels really good. You’re really high because you’re hacking your biology with the AI. It’s very flowy. It works really, really well, but it doesn’t mean that the quality of your thinking matches the quality of your feeling.

To me, it’s like you’re having all the fun of a party while being productive and creating something useful for society. I don’t see—

But I’m not saying it’s wrong. I agree with you. I like it. What I’m saying is, with that much dopamine in your system, we have 1 bit of swag at my organization. It’s a T-shirt that says, “Never trust the dopamine.” People make really bad decisions based upon that much dopamine in your system.

I’ll keep us going. At the bottom of this chart is Eli Lilly. They’re about to announce and release something called retatrutide. It’s a triple-agonist peptide. If Mounjaro is GPT-5.5, retatrutide is AGI in this space. The numbers are extraordinary.

This is clinical data just released. On this particular drug, weight loss was 37 lb versus 6 lb in the placebo group over the course of 40 weeks. Cholesterol was down 27%, triglycerides were down 41%, liver fat was down 80%, and hemoglobin A1C dropped from 7.9, which is diabetic, down to 6.0 in 40 weeks. This is being called a longevity drug.

Yes.

It’s an extraordinary drug that is expected to be released and FDA-approved by mid-2027. A lot of people are getting this from research labs already. This is just some of the detailed evidence heading us toward longevity escape velocity. Salim.

I’ve been using retatrutide for the last few months.

Wow.

What’s incredible is the liver function. Because of decades of damage from alcohol, it’s helping reverse a lot of that, which is fantastic. I’ve never felt better.

I had this conversation with Demis the following weekend. We’re about to see new generations of drugs coming out. The speed of going from design to testing to availability is going to be accelerating. The FDA is on board for this, basically moving us from instead of phase 3 into an expanded phase 2 trial.

We just saw on this chart how much people are willing to pay, right? It’s capped not by the market; it’s capped by manufacturing. Just think about it: When a drug comes out that says, “Oh, this is going to give you 10 years of extra life,” or, “These are going to cure cancer, cardiovascular disease, and inflammation,” these things are coming.

Demis Hassabis

I agree. And I also think, when we speak about longevity escape velocity—which I do think is likely, for the record, perhaps sometime by the early 2030s, if not earlier—I do think that the GLP-1 drug class and everything it evolves into over time is probably the likeliest way that LEV plays out, if I had to guess. I think the economy is reflecting that.

If you look at companies that are worth almost $1 trillion or worth more than $1 trillion, either right now or expected to be publicly traded in the near-term future, there are only a handful. You’ve got a few AI companies, you have SpaceX, OpenAI, Anthropic, and then you have Eli Lilly, which has been publicly traded for a long time but is either already at $1 trillion or about to cross $1 trillion on its present trajectory. And I think that’s the market speaking very clearly: Longevity drugs and AI are the obvious industries of the future.

Yeah, agreed. I want to end this pod in a moonshot mate prediction conversation: What will AI feel like by mid-2028? I want to give you a sense of what we see as the future here, because right now, honestly, when you think about AI, it’s something that you can talk to on your phone. You type at it; it’s an app, it’s an intelligence layer, but I don’t feel like the majority of the world is actually experiencing AI.

I think we’re about to undergo a transition where AI is going to feel very different in the next 2.5 years. I’d like to have a conversation about what you predict that to be like. I’ll start with the very first one, and then we’ll pass it around the room here.

The very first thing is, I think we’re going to give AI permission to know everything in our lives: listen to all your phone calls, watch all of the cameras in your home, read your emails, read everything. I’ve just done that with Skippy. I gave Skippy access to all my Granola recordings, my WhatsApp, my iMessages, my email, my calendar—everything.

And in so doing, the response is amazing. I can have interactions where this AI knows me better than anybody else. All right, so let's begin with that. Alex, you want to kick it off?

Alex Wissner-Gross

I almost think we're asking the wrong question here. I would like to reframe, because I think the question behind the question here is: What will AI feel like to consumers by mid-2028? Yes, it's fair. I think the past 6 months of industry history have taught us that consumers are not actually great customers for AI, at least not state-of-the-art AI.

OpenAI, infamously at this point, tried to turn consumers into power users for reasoning models and failed, and has had to pivot over to enterprise. They had to basically shut down their consumer video division. They've shut down a number of other divisions in favor of code-generating models that are recursively self-improving and targeted at enterprise use cases, because enterprises have the money and the desire to use advanced reasoning capabilities. So, the adjacent question I would ask is: What will an AI feel like to an enterprise rather than to a consumer in mid-2028?

Let's make that an adjacent conversation. But for me as a consumer, I'm going to focus on that part of the equation. I feel like if AI understands me, it's connected to all of my wearables, insideables, and so forth. It has my back as a physician. It knows I had a hard conversation with my spouse and I'm exhausted. I didn't sleep well the night before. It has all this data.

I walk into the room, the lights come down, there's a glass of wine waiting for me because the robot put it out, my favorite music is playing, my favorite community is on. The term I use is ambient AI and automagical AI: The world magically adapts itself to your desires.

Alex Wissner-Gross

I can tell you what I think you want to hear.

What's that?

Alex Karp

You want to hear that by mid-2028, you're going to have an AI exocortex that's a quasi-upload of you, your digital twin in the cloud that knows everything about you and your life and is fully optimizing your meat-body existence, knowing what it knows about you. I think that's what you want to hear. I don't think that's actually how things are going to play out, though.

Okay.

Alex Karp

I think what AI will feel like in 2028 from a consumer perspective—sure, we'll have better augmented reality, and we'll have better robots in the streets. All of that, I think, is already essentially priced into the market. What's not being priced in right now—the non-obvious insight—is new scientific discoveries that consumers aren't the ones driving.

You're taking it to pharma. So, let's put that in the column of industrial, right? When Lyra Scientific and Colossal are creating breakthroughs at a rate—or, frankly, Isomorphic Labs.

Alex Wissner-Gross

The question as constructed is basically: What products is Apple going to launch in 2028? Because these all map onto Apple product categories. You get your smart home, you get your robots, you get your wearables of all sorts. Maybe, if we're lucky, we get ingestible robots.

We get all of these things, but I don't think that's the essence of what AI is going to feel like to a consumer, and I don't think that's the core or the frontier.

Suleyman, what do you think?

I agree with you on the feel-good cloud AI. I think I go more toward what Alex was talking about: The institutional use is going to be absolutely profound, and that's where we'll see the biggest difference. Every single person on Earth will essentially be operating like a small company, with an incredibly powerful team around them supported by what are basically mostly AIs.

I think at the enterprise level we'll be running enterprises on AI-native operating systems. More importantly, we need to rewrite the operating system for civilization, because everything we've done as a civilization—every business in the world for 10,000 years—has tried to solve scarcity. Now we're moving into an era of abundance, but what's the business model around that? I think we need a complete rewrite on that, because all our institutions and nation-states are geared around scarcity. So, we need a complete rewrite of the civilization operating system, and I think the beginnings of that are starting now.

Mhm. Dave, what do you think 2028's going to feel like? And I'm saying “feel” in particular, right? Because right now, a lot of people say, “Listen, I see the AI numbers. I hear you on the podcast, but it doesn't feel any different for me as a consumer.”

Yeah, we launched our holodeck today. You have to go check it out over at Link Studio. We commissioned construction of one of the rooms, put monitors on all the walls, and you go in, close the door behind you, and nobody tells you what to do; the AI just starts talking to you.

In the holodeck, you can create a virtual world, you can create songs, you can create movies, you can just vibe-code whatever you want. It's purely interactive, with big speakers, pulsing music, and everything around you.

It's a party.

David Friedberg

Yeah, go check it out. I think any consumer who experiences that is going to say, “I need that. I want that. How do I get that at home?”

No, not yet. It's hard to do the real Star Trek thing.

David Friedberg

But any consumer who goes into this holodeck and experiences it is going to say, “I need that. I want that. How do I get that at home?” So, I think the problem in 2028 is that there's going to be a massive shortfall of compute relative to the desire to have that experience.

What Alex said is dead right. The enterprise use cases have just discovered this, and they're sucking up all the data centers in the world now. There's no way that's going to get solved by 2028. I'm hoping 2030 or 2031, after the Terafab has a chance.

So, what's going to be the story in 2028? Right now, we've lived this beautiful moment in time where the very best foundation models are made available to anyone on the planet who wants to go to a website and try them. The cost of using the best models in the world is sort of a—it's very affordable.

Yeah.

David Friedberg

Yeah. You have to take advantage of this moment in time, because it's going to get really bottlenecked by 2028.

We're about to see a brand-new set of eyewear, right? We've talked about this at length. OpenAI has got their Johnny Ive device, whatever that might be. Apple has new generations of wearables coming. Meta has generations of wearables coming, and this is going to be ambient AI, where your AI sees what you see, is there and listening all the time, uploading everything into your version of Skippy, your OpenClaw, whatever it might be.

If you want to learn something and you're walking down the streets of Manhattan, and you've got AR glasses that are giving you a tour—this is what it was like in 1905—here's an educational layer. Or, if you want an entertainment layer, you know, “May the 4th be with you.” I'm a Star Trek person, but here we are on May 4th. You've got these little droids popping up and shooting at you.

If you want whatever you want, this AI is optimizing your audiovisual experience in line with what you've asked. So, I think that's going to feel different. I think we're going to see the first wearable devices enabling that. Do you guys agree? Alex?

Alex Wissner-Gross

I'm a big fan of Vernor Vinge, who wrote extensively about what the near-future AR/VR metaverse, if you'd like, would look like. I do think we get that. I do think it's just a matter of time, battery energy densities, and other progress. So, by 2028, do I think we get Vernor Vinge's smart glasses with compelling, long-life augmented reality or mixed reality? Yeah, sure. But I also think—

[Laughter] Yeah, after all that. Oh, boy.

Alex Karp

But I also don't think it matters an enormous amount. Yes, it'll be great, and it'll be popular.

By the way, I wanted to say something. In the last podcast, I was trying to make our conversation relatable to the general audience. My mission here about what AI is going to feel like is to support their understanding of where things are going. Yes, we also speak to the CEOs and the heads of the frontier labs here, but I want to give people an understanding of what their life is going to be like.

Alex Karp

A really easy, practical thing: Go anywhere you want autonomously for 20 cents a mile.

I'm going to do a quick speed round here on what is something that AI will feel like in 2028 that maybe is unexpected. Dave?

David Friedberg

We already spend more money on video games than all other media combined. When you overlay personalization, your own voice, and the fact that it remembers your personality and your past conversations, it's so immersive. And in a good way. It could be good or bad. It's intentional. It's designed, but if it's designed well, it's so engaging and immersive.

I suspect that an entire generation will have 70–80% of their conversations with AIs and not with other people.

Interesting. That's interesting. Salim?

Something I'm really excited about is just infinite, perfect, and long-lasting memory, because our memories are so flawed as human beings. The fact that it can record everything and recall who I met, et cetera—I’ve met many of you at events and, God help me, I can't remember anything. The ability to do that, I think, will be a huge addition to me as a person.

Yeah. Steven?

The one that's interesting to me is—I always say that you tend to hire one employee twice.

You hire them when they're normal and then when they're scared. People are very, very different when they're fearful. I think the AI coaching, the AI psychology, all that stuff—the AI coach in the room, the coach in the world—is what's really interesting to me. I'm interested to see—I do see, I've been seeing even the Defense Department's work using AI therapy, right? I got to play with a lot of those therapists along the way.

They're remarkably good. Most people are not using AI to coach their relationships. But your AI coach that's actually in your ear is really interesting, because humans—we run on 4 knobs, right? Balance, arousal, approach-avoid. That's pretty much humans, and you can coach them up pretty quickly in real time. So that's what's really interesting to me: I think we end up with better humans.

Fascinating. Alex, on the consumer side, would you ponder a vision?

Alex Wissner-Gross

I think we are merging with the machines. I do think augmented reality and wearables are part of the solution. I think ingestibles are going to be part of the solution. I'd be sorely disappointed if, by the end of this decade, we don't have many people swallowing computers. We see the beginnings of it right now with PillBot-type form factors.

I did a projection, something I think friend of the pod Ray would be proud of. If you extrapolate the typical size of a computer needed over the past 15 to 20 years to achieve a gigaflop of compute—which, historically, if you look back at Apple's new product announcements, Apple has waited for new devices to pass a gigaflop before releasing them—it's true with the original Apple Watch, the original iMac, and the original iPhone.

If you extrapolate the typical size of a computer that, at the time of launch, passed a gigaflop and extrapolate that forward, you find that not by mid-2028, but by the mid-2040s—around the time Ray says we're going to hit his version of a technological singularity—the size of a computer hits approximately the size of a eukaryotic human cell.

So, what am I most excited about? Well, I have a laundry list, but I think if you extrapolate the progress of the size of the computers running the AIs, we're going to have cell-sized nanomachines running those AIs, and I think that'll be incredibly exciting.

All right, everybody, and that's a wrap. We'll go to our outro music next. Ladies and gentlemen, let's give it up for the incredible David Blodgett, Salim Ismail, Alex Zigal, and Steven Kotler.

If you made it to the end of this episode, which you obviously did, I consider you a moonshot mate. Every week, my moonshot mates and I spend a lot of energy and time to really deliver you the news that matters. If you're a subscriber, thank you. If you're not a subscriber yet, please consider subscribing so you get the news as it comes out. I also want to invite you to join me on my weekly newsletter called Metatrends. I have a research team. You may not know this, but we spend the entire week looking at the metatrends that are impacting your family, your company, your industry, your nation, and I put this into a 2-minute read every week. If you'd like to get access to the metatrends newsletter every week, go to diamandis.com/metatrends. That's diamandis.com/metatrends. Thank you again for joining us today. It's a blast for us to put this together every week.