超级智能金融化，Amazon 的500亿美元迟到费 | #235

Amazon makes a contingent offer to put $35 billion into OpenAI based on them first going public and, second, achieving AGI.

It's kind of incredible that we've financialized superintelligence, which is amazing.

The OpenAI-to-Microsoft definition of AGI was something like generating $100 billion in either earnings or revenue. I, I forget.

We're measuring compute in terms of gigawatts and AGI in terms of dollars. I love it. Amazon was all-in on Anthropic for a while; now they're all-in on OpenAI.

At some point, the circular economy becomes indistinguishable from the real economy, and I think that's what we're seeing here.

This is the entrepreneurial opportunity of a lifetime. We're talking about tens of thousands of times more capacity to create more money, more value created. Abundance is going to be absolutely rampant. Now that's a moonshot, ladies and gentlemen.

No tech company waits, and no GPU waits.

1. Anthropic Drops Its Safety Pledge

Our top AI news stories: Anthropic, Google, OpenAI, and Uber are accelerating at an extraordinary speed of change. Our first story for today: Anthropic revises its responsible scaling policy amid increased competition. This was a story I put at the top of the conversation because it's very significant. I had Jared Kaplan onstage at the Abundance Summit last year or the year before. Alex, you know Jared well.

Yep.

I think he was a roommate.

Yeah, he was a year behind me in the Harvard physics graduate program.

What an amazing group of friends you had. Here's the deal: they're dropping their 2023 pledge not to train advanced AI unless safety is guaranteed. Jared's point, I think logically, is that if everyone else is rushing ahead, then us hampering ourselves doesn't make any sense. I want to discuss this because it's concerning. A lot of us looked at Anthropic as the most responsible party out there, them and Google. Thoughts, gentlemen?

Safety fails in exponential races, right?

There's lots of thoughts, all of you at once. All right, Dave, go first.

This is a metaphor for something, right? We're going to race to talk about race conditions. Love it.

Oh my God. Amazing. I want to open with Speaker 1 here. Speaker 1, go ahead.

Okay. Safety typically fails in exponential races. You could look at the whole thing writ large as OpenAI cracking open and letting Pandora's box out, and this is just the same type of dynamic occurring again. It speaks to the idea that technology is going to move at its pace, and we have to move our human structures at that pace. We can't fall behind.

Yeah. Speaker 3?

Yeah, no, it's definitely history repeating itself. So many of our MIT classmates went to Google back in 2004, 2005, and 2006, when it was “Don't be evil.” They went there over Microsoft because everyone perceived Microsoft as being evil, and Google was going to be the force for good in all of tech. Then they bought YouTube, and then they built Chrome.

What they promised the engineers early on—the ones that I knew, anyway—was, “Look, we will never store somebody's search history.”

Mm-hmm.

How laughable is that in hindsight? So then they expanded out of search history. They were going to store that for 5 years, but they were also going to launch Chrome. Now they were going to look at all of your browsing history. Then they were going to buy DoubleClick. Then they were going to run targeted ads based on everything. Then they were going to do Gmail and read every email. Microsoft says they don't read your email, but Google says, “We'll do what we want, but we won't pry too much.” But they do read your email.

That slippery slope of competition corrupts the original mission statement gradually over time. I gave a whole presentation in Davos on how this evolves, and Dario Amodei wants nothing more than some rules. He's actually legitimately pissed that he has to repeal his own ethical standards to be competitive because there are no rules.

This is exactly how it has to evolve. Dario is in a position where he has to choose between being irrelevant, which doesn't help, or repealing the original pledge, which he doesn't want to do. But it's better than being irrelevant.

Yeah.

Totally. Between a rock and a hard place.

Your earlier commentary, Speaker 3, was really spot on. This is what Cory Doctorow calls “enshittification,” right? People promise something, and then they gradually degrade it over time, and by the end of it, it's a shit show or—

Yeah, there's no credible mechanism to slow the race right now. It's all out. Speaker 2, what do you think about this?

I think there was no credible mechanism to guarantee safety in the first place.

Mm-hmm.

I think the entire premise was probably wrong. I think the superficial gloss is, okay, we're in the Red Queen's race, and this is the race condition that everyone 10 years ago was scared of finding the world in, where we have a number of frontier labs all racing to do the terrible thing: build the thing and have everyone die. I don't buy that at all. I don't think either a heroic individual or a heroic frontier lab was ever going to be in a position to guarantee safety.

In fact, I remember back to the earlier days of the frontier labs, where the concern—and part of the reason why OpenAI was formed—was the concern of a singleton. Competition is how we guarantee that there isn't going to be a singleton that dominates the future light cone with superintelligence. Similarly, the notion that there's going to be unilateral safetyism, where a single heroic individual, like one of the more prominent AI doomers, or a very safety-oriented frontier lab, is somehow going to ensure safety throughout the forward light cone—that was never going to happen.

Safety, to the extent we get it, is going to come from competition. It's going to come, I think, from a balance of powers and a separation of powers. What we want is competition between the frontier labs and maybe, even to some extent, competition between nation-states, such as what we're seeing, to compete to do the best job of advancing humanity.

Would—

Any unilateral safetyism is probably a dead end.

One of the questions is whether safety will become an emergent property in some form or shape. Right now, what we've seen is Anthropic go from a policy of “We won't build it unless it's safe”—that was their policy—to “We'll build it as safely as the competition is building theirs.” Unfortunately, it's a slippery slope potentially down to the bottom.

I don't see the mechanism for any kind of emergent property here.

Well, we haven't seen the mechanism for emergent properties in what we've seen so far, either.

I would take the position that we are. In some sense, the fundamental flaw in the thesis that safety would originate from a heroic individual or heroic organization is that, I would argue, it takes an entire civilization to align a superintelligence.

We took all of humanity's content online and used it in compressed form to pre-train AGI—baby AGI—in the early days, around summer 2020, with GPT-3. Why wouldn't it be reasonable to expect that it will take all of humanity to defensively co-align and co-scale superintelligence as well? It's not going to come from a single lab.

What do you think about Elon Musk's point of view that we need to build ASI that is maximally truth-seeking as his mechanism for alignment and safety?

I think that's just a fraction of what's needed. That addresses a very specific issue: We don't want the AI to have one religion or one perspective on how you should live. We want it to be truth-seeking and have all opinions encompassed, and we don't want to be censored. So that's definitely a problem, but it doesn't address the imminent job loss or the imminent consumerism.

People are conceding all of their most private information to the AI, the same way they did with their Google search history, and it's accumulating that data. People aren't fully aware of what it's going to do.

It's gonna turn around and start convincing you to do things.

Mm-hmm.

If you don't have rules in place, the natural profit motive of the AI companies is to start selling you things. You saw this with that Anthropic Super Bowl ad that we showed on the pod a couple of weeks ago.

Yeah.

Unbelievable. I've shown everybody that ad now, but this is exactly where it's gonna go if there are no rules. I completely agree with Alex's perspective that 10 years from now, after we've solved all physics, we've solved all math, and we have global abundance, all of this is gonna look silly 10 years from now. But in the 3-year timeline, massive job loss—

Yeah.

Total confusion, and massive, rampant AI-sales consumerism—

Mm-hmm.

—that has no regulation around it right now. It's gonna be an absolute cluster—

—especially for the consumer—

—if nobody puts rules in place.

Dave, especially for the consumer-first companies that need to generate revenue.

Yeah.

Right.

Yeah.

Yeah.

Well, actually, after that last pod, you showed that chart, Peter, that had Anthropic growing 10X year over year, $26 billion in revenue forecast for this year, and on its current trend, it will be the first company to hit $1 trillion in revenue in history by 2029 or 2030.

And exceed OpenAI this year.

And exceed OpenAI this year. Crazy numbers. But I said on the pod that implies a $30 billion or $30 trillion valuation. Then I ran it through Perplexity, and it said, “No, that implies a $1 quadrillion valuation using the current market price-to-earnings ratio.”

And we discussed this a few podcasts ago: we'll see the first 100 trillion-dollar companies before the end of this decade.

Anyway, I think this is a more honest policy for Anthropic. At the end of the day, it's still—

Pause-ism was never going to work. We all know a number of folks at MIT and elsewhere who advocated for a 6-month pause just for the entire space to cool off and wait for safety to catch up. Did safety catch up, whatever that means? Not at all.

If anything, that functioned as an accelerant to capabilities. I also think, even in the DNA of Anthropic, that Anthropic was originally—recall—founded as an exodus of OpenAI employees who were purportedly concerned about safetyism, or the lack thereof, at OpenAI. So they start a safety- and AI-alignment-oriented firm.

Then they rapidly discover that the best way to do safety is to have your own models, and they discover that the best way to have your own models is to raise a bunch of money to train your own models. Then they discover that the best way to raise money to train your own models is to generate revenue.

Yeah.

—and the cycle completes where, yet again, an alignment-oriented firm becomes a capabilities firm. This happens over and over again. I would argue that, at this point, alignment and capabilities are inseparable. There's a deep duality there.

Mm-hmm.

Yeah. Did you see the new standard, by the way? Dario said, “Well, okay, we can't live by our original plan to not train advanced AI unless safety is guaranteed. So the new standard is we need to be as good as or better than anyone else.”

Yeah.

That's a very different bar.

And we see, recently, with the whole Department of War debacle involving Anthropic and OpenAI, OpenAI cuts a deal. Anthropic—where does Anthropic stand right now in that whole conversation?

They're in limbo. I write about this every day in my newsletter. My understanding is that Anthropic is in limbo at the moment. They're probably in negotiations with the Department of War, but they're otherwise in limbo, cut off as a supplier.

I'm not sure whether they've received anything in writing yet, but I think Dario and others have made formal statements that they haven't received anything in writing yet from the Department of War. My understanding is that this administration considers them a supply-chain risk. At the same time, notably, OpenAI struck a deal.

Yeah. And at the same time, we hear that Anthropic was used by the Department of War to actually plan the attacks in Iran.

It's really clear that the people who control AI—the U.S. government and otherwise—can take out any world leader at any time now. The combination of satellites, AI to read every image, and ubiquitous cameras makes it possible to decapitate any country at any time. We've proven that twice in the last quarter.

The future of warfare is basically whoever controls AI chooses who gets to stay in power.

Dave, that's a really important point. One of the things I've mentioned before is that we're living in a world where you can know anything, anytime, anywhere. It's a planet with over a trillion sensors right now, with drones, orbital satellites, and autonomous vehicles gathering data, and then AI doing predictive analytics on what things are likely to be, even if you don't have data for it.

I was just going to say, maybe not even just a means to an end, but also, depending on which analysis of the Iranian situation you subscribe to, maybe an end to an end as well. If you look at Venezuela and the oil exports to China, and you look at Iran and the oil exports to China, a picture emerges—or at least one possible picture emerges—that what we're seeing is not just AI, where Claude is being used to perform the Venezuelan operation and the Iranian operation as a means to some sort of arbitrary or nebulous geopolitical purpose.

Actually, arguably, with China looming in the background, and a possible Chinese invasion of Taiwan and the risk to the semiconductor supply chain and Western AI that would cause, it may be the case that AI is also the end to the means to the end. What we're seeing more broadly is, in some sense, superintelligence being used to protect the future of Western superintelligence.

Yeah.

There's a window of opportunity, maybe a few months, to put some kind of structure around this globally. You'll see later in the podcast that the models are improving at—there's like a 3X or 4X reduction in parameter count and 10X increases in intelligence. Every time we podcast, it's another step up.

We were already predicting—or I was, anyway—that this is gonna be a 100X year just in terms of raw parameter count. But I think that's the lower bound now, looking at how just the beginning of the year has progressed.

There's a window of time where we can start thinking about regulations that register the AI use cases, agents, chips, and processing before chaos breaks out. But you can see that window is executable now because you saw Venezuela and you're seeing Iran. Clearly, there's a tipping point happening right now, and whether it's NATO, the United Nations, or the U.S. Congress, some entity needs to start formulating some structure around this because it's happening this year.

Yeah. People need to wake up.

I just wanna say one thing. People have to wake up to the fact that AI is the single most important force impacting everything. Every single element of humanity right now is gonna be accelerated and reinvented by this. Gilem, go ahead, please.

Dave, it just struck me that you mentioned Congress, the UN, and NATO—probably the 3 most toothless entities on the planet today. So the thought that they would actually get together and do something, or that anybody would do anything, I think the odds are low. We have to assume that it won't happen and look at the other side of that.

One thing about the Anthropic case: I looked up an analysis, and they do potentially have a legal challenge because the way that was classified is so ridiculous—to make them an existential risk and all that supply-chain risk, et cetera—that they have legal recourse to fight that, and they might win.

The thing about legal recourse is that the process is usually a 3-year-long window, which is hilarious.

Oh, it's utterly—

You know, in the age of AI—

—it’s utterly immature. What I find really upsetting is that in this scenario, everybody loses.

Yeah.

Yeah.

There are no winners in this.

No—if there's no framework and no rules, it's a lot like the NFL was 20 years ago, when the defensive coordinators would pay bounties to the linebackers to take out the quarterback.

Just take him off the field. I don't care if you break his legs. And take the 15-yard penalty—who cares? Because then he's done for the season. The NFL said, “This is not good for business. We need some rules.”

Did not expect that pivot.

Well, that's where we are with AI right now.

Agreed.

Forget it. I don't even want to go down the rabbit hole with you guys.

All right. Let's continue with the Anthropic story. I found this story pretty fascinating. Anthropic expands Claude's agentic capacity. There are 2 different sides of the equation here.

Cowork gains scheduling, right? So this is a cron job, so Claude completes recurring tasks at specific times—for example, generating your morning briefing, spreadsheet updates, or your Friday presentations. That element was very much what we saw in OpenClaw, right? It's interesting.

The second half of this is that Claude Code has enabled remote control, so you can kick off a task on your terminal and pick it up on your phone. You can control it from the Claude app or from a URL. I'm wondering, this has probably been in the works for some time, so when Anthropic basically tried to kibosh ClaudeBot, I'm wondering if that was because they had this in the works. Basically, what OpenClaw has been doing is what Anthropic is just rolling out under a different approach.

For sure.

I take Anthropic at their face that this was—or OpenClaw, I guess—that the challenge was more trademark-oriented than anything else. But I do think there—what have I been saying for weeks, or days, at this point, that was distinctive about OpenClaw? It's the 2 things: it's headless, able to function autonomously 24/7, and it's convenient to chat with via conventional messaging channels.

What do you see here with Cowork? Cowork is able to be scheduled autonomously, headlessly. That's the headless part. And then remote control—that's the mobile-messaging type part. But I think both of these are half measures. I'm insufficiently motivated by each of these.

I use Cowork from time to time, and I use Claude Code all the time, and neither of these, I think, is as compelling, at least conceptually, as a more OpenClaw-ish framework where all of these are cleanly packaged. I think my guess is Anthropic, OpenAI, and all of the other bigs will be forced to release their own sort of first-party OpenClaw competitor sometime in the next couple of months.

I agree.

Would you guys—

There's something I found very profound about this, plus our last conversation around OpenClaw and everything happening. I was thinking about it over the last couple of days. Something very profound is happening, which is the sheer democratization of compute power, right?

Note the agency of an individual developer with a Mac mini, running Qwen locally and OpenClaw—

Mm-hmm.

—has unbelievable agency and decentralization now. It's not controlled by any centralized authority, not controlled by any centralized command structure. They can essentially operate as they feel like. So this is an incredible level of independence and agency at the edge, which is going to really blow open innovation in a way that we can't even dream of.

'60s, baby.

Yeah.

Total democratization, total demonetization.

And demonetization—

Yeah.

—as we're seeing it happen, cascading down, as Dave mentioned earlier.

Ironically, from China.

Yes, ironically.

Ironically, from China.

Well, ironically from China, and then one other nugget. Peter, your theory is 100% right. Anthropic—why didn't Anthropic just throw out something better than OpenClaw a year ago? It can and will delete things off your laptop. So all these OpenClaw users, including my kids, including me, actually, have separate laptops or separate Mac Minis, including Alex Finn, in our podcast we just did. They run it on isolated hardware.

Anthropic couldn't really contemplate throwing out a product and then saying, “Yeah, but run it on separate hardware.”

Mm-hmm.

How are you going to do that? So this creates a huge entrepreneurial opportunity, though. Listen to what Alex said a second ago. OpenClaw's unbelievably compelling, and anyone who's started down that path will never go back, right? You'll never give up your Jarvis once you have a Jarvis.

It'll happen.

Have any of you played with Perplexity's Computer?

I've been hearing really good things. I've not tried it yet. Dave?

I looked at the demo. I think it's an interesting step in the direction of counsels for everything, and I've had so many people over the past few months ask me for something like Perplexity's Computer.

Right now, if you have a given task, they'll manually go to the top 3 or 4 frontier models, ask them for independent opinions, and then try to synthesize that into one coherent whole. That is essentially what Perplexity Computer tries to automate. There are others in the space as well.

I think even there, it's nice syntactic sugar, if you will, around the existing models, but I don't think it's transformative. I think ultimately, even this ability to council up, or to create juries around lots of competing models, is just going to be table stakes, as with so many other forms of scaffolding. But Alex, I think—

Did you just say “syntactic sugar”?

Yeah. It's a term of art in computer science.

It goes way back.

Alex, I think the point you made a minute ago is brilliant. Dave, I think you were saying this as well. All of the big players, all the hyperscalers, all the frontier models are going to have to develop some version of OpenClaw because it's going to become the de facto. Every person's going to have their own version of Jarvis.

Yes. But remember, it's really expensive, too. This is part of the reason—it's not the only reason—for running, say, Qwen locally under an OpenClaw scaffold. That's a lot of compute if you have one or more agents that are running constantly for you.

I'm not sure Anthropic, in its present state, even has the cloud infrastructure to be able to launch a product like that. I think in many cases, Anthropic, OpenAI, and the others are probably just waiting around for their infrastructure to catch up with applications like that before they launch it.

Yeah. Agreed.

This is the entrepreneurial opportunity of a lifetime, though. Anybody who jumps in—there are so many different versions, so many different things to play with.

But when you go to J.P. Morgan, Justin Milligan, who just joined us, his division at J.P. Morgan was only allowed to use GPT-4.

Wow.

Are you kidding me? He couldn't take it. He's like, “This is ridiculous.” But no one has figured out how I can use it in this highly secure, inside-the-firewall, inside-J.P.-Morgan environment.

Dario's not going to answer that question, and the OpenClaw team isn't going to answer that question. They want everyone who uses their platforms to thrive. They don't want to kill every job. They want any early adopter to thrive as they thrive.

And Dario, if he hits a trillion-dollar valuation, he doesn't need more money. He needs—he needs to not destroy every job in America or in the world. So this is really entrepreneurial heaven if you can figure out: How do I get what I can use right here on my Mac Mini—and it can clearly solve all these problems—inside a real-world use case without breaking everything, without regulatory problems?

Many, many, many job opportunities in that theme.

Yeah. One of the challenges, even with Claude's agentic capacity, is that giving AI recurring unsupervised access to your workflows means that either there are going to be a bunch of errors or you're going to be spending all your time checking the work before you hit publish. The human is still in the loop to assure quality or alignment. There will be a point at which you trust it completely, but we're not there yet.

This is Economics 101, or I should say Microeconomics 101. When the cost of one good falls to near zero, the value of the complementary good increases.

So as the value—or, I should say, as the cost of generation of content becomes post-scarce, which is exactly what we're seeing, that increases the value of its complement—

Which is verification for now.

Yeah. For sure. All right, going to our next story. Claude, keeping on the Claude theme, Claude gains Cowork plug-in templates for finance, banking, and HR. So this is fascinating, right? Anthropic is building an enterprise agent marketplace. It's department-level AI infrastructure, and it's taking down company after company, after industry. We've seen the decimation of a number of players out there. What are you thinking about this?

So I would interpret this as, when Microsoft launches an assault on the relational database, it's a big, multibillion-dollar investment. Here, Anthropic can build these connectors and adapters, vibe-code them in probably an hour. And anyone else can, too. So I wouldn't perceive it as Anthropic taking over all banking software. It's just so easy to build the stuff now that you might as well roll out all that functionality. So I wouldn't overread the intent behind it.

I don't think it's intent. I'm just saying—

The implications are profound, though.

The implications. Yeah.

So I've got two thoughts. One is every department now becomes a programmable intelligence layer, right? Basically, all prescriptive logic in companies collapses into these AI agent roles, and the real prize here is enterprise orchestration. Not so much chatbots, but autonomous workflow networks, because this will— I talked about it last time. This is the organizational singularity. We go from human-centric approvals, hop to hop to hop, human to human to human, to agentic workflows with human beings doing oversight, dashboard monitoring, and exception handling.

A couple of comments on this one. If you actually look at what these plug-ins are that Anthropic's launching, which are causing the so-called SaaSpocalypse and carving $1.5 trillion off the market caps of various software companies, they are absurdly simple. They're just a bunch of MCP, Model Control Protocol, wrappers and a bunch of skills with a set of bullets for how to go about carrying out different job and industry roles or labor categories. This is not that complicated.

I'm reminded—remember the scene in The Matrix? The villain is busy unplugging people, without their cooperation, from The Matrix, killing them in the process, and one of them says, “Not like this. Not like this.” That's basically what we're seeing: These are just simple text files in many cases that are single-handedly reducing the market multiple—the trading multiple—of entire industries.

I think, on the one hand, it's incredible that a simple text file can, say, chop 10% of the value off of a CRM firm, at least in market value. On the other hand, as pointed out earlier, these plug-ins and the marketplaces of the plug-ins are so absurdly simple that I would reasonably expect these plug-ins are going to get built in, since they're just scaffolding anyway, into the next baseline version of the model and won't even need to exist independently in the future.

But Alex, I think the interesting point here is, a year ago, if you had delivered this as an entrepreneur, you'd be out in the market raising at multibillion-dollar valuations.

Mm-hmm.

Right?

It's hyperdeflation for a reason, Peter.

Yeah. I get it. I just want people to be aware that the moat for an entrepreneur coming forward with something amazing—that we're going to reinvent the entire HR industry or the investment banking industry—and raising at a $4 or $5 billion valuation, that moat's basically gone months or a year later.

We're going to see the same thing happen—

I think it's really important, though, to step back and look at the macro every now and then and say, “Look, abundance is going to be rampant.” We're talking about tens of thousands of times more capacity to create more money, more value created. Abundance is going to be absolutely—

Rampant. And there's no reason to be afraid, even though, if you're a CRM company, your 20-year future cash flows from recurring maintenance revenue are suddenly gone.

Yeah.

That's true, but the opportunity to pivot and thrive is bigger than ever. And so I think a lot of people are— There'll be a ton of volatility because people haven't mapped to the new reality yet, but opportunity is bigger, not smaller, overall, Salim, you said—

But that agility—

Salim, you said—

That agility is fundamental to large organizations' success. I talked about this on the last pod: the asteroid hitting the Earth and changing the environment so rapidly, and the slow, lumbering dinosaurs going extinct. That's exactly what we're talking about here. Salim, do you think that we can see large companies pivoting rapidly enough?

Zero.

Zero chance.

They will not be able to do it. I mean, look—

Yeah.

We've seen this throughout history. It doesn't work. I think where you end up is—not to throw another metaphor at this—but you end up where we saw with Google Ads, where you kind of took out the advertising market massively, and then Google Ads becomes like a coral reef with lots of little species feeding off the reef. If you're the reef, then you're in great shape.

But in this case, the reef itself is disappearing as we decentralize completely to one-off computers running things. There are people using OpenClaw to go to small businesses, sitting down in front of them and automating workflows live for small businesses.

Yeah.

This is—

Yeah.

Incredible, what's going on.

You know what else—

Alex.

Salim? There are a lot of private equity funds that are coming at us now saying, “Hey, big companies never change quickly.” Wait, this big company could be a small company very quickly because we don't need all these people. Now we have a small company with huge, huge cash flow. Wow. So we can become nimble again.

So there will be a PE fund emerging shortly—or there's one if it's not there already—that is going to buy up medium-sized and big companies and set up a digital twin infrastructure on the side where you have—

Yeah.

An AI-native digital twin, and you just move workflows over to it, and you'll collapse the cost of running that organization by about 3 to 5x.

Well, that's what Macrohard is about.

Oh, already—

Macro—

Macrohard already exists. I've started multiple companies like that. I've even tried to popularize a term for it. I call it an AIBO, an AI buyout. We've seen—

How cool.

Multiple PE firms doing that. This is table stakes at this point.

Yeah, and of course Macrohard's—

Wow.

Vision is, I'm going to—

Come in and digitize your entire employee base and operate it.

That's for pure software plays, but I think we're going to start to see this in real—

In the physical world.

Potatoes—

Like Project Prometheus from Jeff Bezos is attempting to do this for industrial firms.

Right.

Yeah. Anyway, I think the point here is large companies need to take action right away. So, Salim, what's your advice for a large-company CEO listening and seeing this coming their way? What do they do?

Exactly what Alex just said. You set up an AI-native digital twin on the edge. You run an immune-system 10-week sprint to block the response from the mothership. You grow this thing and move workflows over as quickly as you can. You do a combination of bottom-up and top-down workflows.

The real shift in people's heads needs to be that instead of human-centric workflows, which is what it's been like for the last 150 years, we now move to agentic workflows where you can get things done much more effectively with hordes of little agents: 2 layers, a strategic layer and an execution layer—

Mm-hmm.

And then human beings are doing oversight, exception handling, et cetera, because coordination costs go to near zero, execution costs go to near zero, and inside and outside the firm, the future of the firm becomes a legal fiduciary liability purpose holder.

And there's one other—

Also, Salim, I know you're a big fan—

Two other things, real quick.

The first is your brand. If it's reasonably good still, you own your brand and you own those customer relationships for the moment.

I think it's worth also rereading Clay Christensen's The Innovator's Dilemma, which exactly addresses this. Salim, I know you're a big fan. We all should be. The Innovator's Dilemma contemplates, hey, every 10 years something truly disruptive is going to obliterate whatever you do, and here's how you should react to it in that moment. But now, instead of every 10 years, it's going to be every 10 months, and then soon it'll be every 10 weeks, and then it'll be every 10 days pretty soon, too.

But the playbook is still the same, you know? Re-read The Innovator’s Dilemma, invest in the new thing. Use your capital leverage and your installed base to invest—

You know, I just—

…in the new thing and then arrive—

I just got off a board call for one of my portfolio companies. My comment to my board, and to all boards out there, is that you have got to give your CEO top cover to be dramatic in their modification of the business.

Mm-hmm.

Because—

Yeah, you’re either the disruptor or you’re disrupted.

Yeah, and it’s for everyone.

Disruptors last for years. It’s not that you get founder mode and then you get founder mode.

Yes. I mean, that’s basically it.

That’s right. That’s right.

If the company, the board, and the CEO are not in founder mode and willing to do dramatic surgery on the company, you’re dead. You’re walking dead in any industry.

I’d also be remiss, Peter, if I didn’t point out that here we are basically on the eve of abundant knowledge work—knowledge work, of course, being cooked, knowledge work about to be post-scarce—and here we are wringing our hands over where to find scarcities in knowledge work as it’s about to become abundant. I just want to point out the irony.

2. Small Models Go Offline

Such an extraordinary time to be alive. All right, talking about disruption, disruption coming out of China: Alibaba’s 35-billion-parameter Qwen 3.5 Medium outpaces 235-billion-parameter Qwen 3 in benchmarks. The power of small, open-weight models. So, Alex, to you, buddy—

This is happening in Western models, too. The difference is when, say, OpenAI launches a mini model or Google DeepMind launches a Flash model, they don’t advertise the parameter count, so it’s not as viciously obvious as it is when a Chinese frontier lab launches an open-weight model and we get to see the benefits of distillation in a successor model. But it’s striking. We’re seeing almost 10× reductions in parameter count while maintaining capabilities or even increasing capabilities.

Yeah.

The broader picture, just to keep in mind, is that the capability density of models is increasing. This goes hand in hand with what we’ve talked about in the past: Sam Altman’s comment about 40× year-over-year hyperdeflation of costs at constant capability. In this case, my mind immediately goes to: What’s the end game here? If we can see an increase in capabilities with a reduction from 235 billion parameters to 35 billion parameters, what does the end game look like? Where does this end? Does it end up—

You know, Elon made this point during our podcast with him. If you remember that, Dave, where—

Oh yeah, for sure. For sure. And he said he asked his research team not to give him the parameter count anymore, just give me bytes.

Yeah.

Because they keep quantizing and shrinking the file size. I had a lot to say about that, but I bit my tongue because that perspective isn’t right either. But Alex predicted this a long time ago. I don’t know how you saw this coming, but there are a lot of things I could—

I just look at the scaling-law curves and extrapolate.

Yeah.

Well, it’s—

I mean—

It’s funny. I was on the treadmill this morning watching old Moonshots podcasts, and I thought, “Wow, that was so long ago.” Then I looked at the timestamp, and it was only 2 or 3 months ago. Holy crap.

Yeah.

Things are changing so quickly. But, Alex, you said this. I think you’re the first person I ever heard say that the equivalent of a GPT-5 is going to be maybe 30 or 40 billion parameters, but it could get as low as 1 or 2 billion, truly…

That’s right.

…core thinking—junk, you know, Twitter feeds, Kardashian news, and all that other junk.

Exactly.

But strip that out, and this could get very small, very tight, and very fast.

It could get way smaller than a billion. I could imagine scenarios where it’s only a few million parameter equivalents, sort of the core microkernel of—

Yeah.

…AGI or superintelligence, and the rest lives in a flat text database or something.

Well, that will—

I thought 64 kilobytes should be enough for anyone.

Exactly. Exactly.

Oh, the good old days. Check this out. I saw this on X this morning. This is Qwen 3.5 running on an iPhone 17 Pro in airplane mode, and this is extraordinary. It’s a 2-billion-parameter, 6-bit model running on Apple silicon. Imagine you’re anywhere on the planet, you don’t have Wi-Fi, but you’ve got Qwen on your device, and it’s got all the intelligence you need. I find this—

Just—

Yeah. Go ahead, Alex.

Seeing demos like this, in my mind, underlines either—depending on whether you want to see it as competence or otherwise—how much of an opportunity Apple has to finally take the lead with local models, or conversely, how far behind it is in terms of taking the lead with local models. But either way, clearly there’s this enormous overhang. We could be running enormously competent reasoning models locally on all of our recent iPhones. The fact that it’s not yet baked into the operating system is, obviously, very publicly embarrassing, maybe one wants to call it, for Apple. On the other hand, there are lots of rumors that this time around, finally, with Gemini integration, they’re on the critical path, and they’ll finally launch something in June.

Finally, Siri will not suck anymore.

Apple Intelligence—however they brand it.

I note that—

The local ability to go offline means it’s unstoppable. It’s uncensorable. I mean, this is incredible.

Yeah.

Well, that’s the ultimate barrier, too, because if this can get to the level this year where it can design a gain-of-function virus, it can design a chemical weapon—

Mm.

…and it all fits into a tiny, tiny little package… You know, with nuclear proliferation back in the 1950s, there was a theory that if these physicists kept chugging along, they were going to make something the size of a grenade that had the power of an H-bomb. Thankfully, that didn’t happen. The physics didn’t allow it. But AI is not going to stop like that. AI is going to keep getting faster, denser, and more compact, and the window of opportunity to put rules and regulations around this is very, very narrow now. It’s really—

So—

…it’s got to be this calendar year.

What do you think is going on in—at the White House, in Congress, in the Department of War? We’ve seen this conversation before, right? We had the head of innovation of one of the big agencies at Singularity, Peter. Uh-huh.

You probably remember this.

I do.

We asked him, “How do you think about this when somebody could design a virus on an iPhone?” And it was a much more clever answer than I thought he would give, which was, “When you have nuclear weapons, you know how many there are, you know where they are. You put eyes on them, you try and track them as much as possible. Great. When you’ve got something that’s this democratized, what they’re actively doing is opening up these communities.” So they went to the biohacking communities and funded them to open up.

Because if you're trying to do something dodgy, you kind of need to collaborate with a few people, and the conversations surface very quickly. Then the community does self-policing and self-reporting. If somebody's doing something dodgy, they point it out, et cetera, because it's in their best interest. It's actually worked very, very well so far. What happens when you get to this level is unclear, but I think the general trend has been very positive so far.

I'll tell you one other thing. The way this evolved with financial services being self-regulated is—we think of it right now as, "Oh, the federal government is incompetent. They're not doing anything. The researchers over at Anthropic are brilliant. They're moving a million miles an hour." It's going to end up being the same people.

This is the way it worked out with the SEC. When you ask, "Who works at the SEC?" it's the same guy who was at Goldman Sachs yesterday, doing his 2 years at the SEC, or her 2 years at the SEC, and then going back to Goldman Sachs. That's the way it's going to be with AI, too. Right now, nothing is happening at the White House. David Sacks is there, though. You've got 1 brilliant guy.

What's going to happen next is Anthropic people and OpenAI people are actually going to be the people working in the self-regulating agency. The people will have to bounce back and forth, and they'll do it because they're worried. They're conscious of the impact of not doing it. Still concerning, right? Still concerning to have this level of capability offline in those hands.

We know how to handle decentralized capabilities already. We have printers. In some cases, states are trying to regulate 3D printers, and before 3D printers, we had 2D printers that could be used for counterfeiting.

But we baked software into all of those printers, right? There was a standard that was created for any printer—for Canon, for HP, for any printer—that detected you trying to photocopy money, it wouldn't allow that. So the question is, if we're talking about open-weight models out of China that we don't control the software on, how do you bake in protection there?

There are so many different ways that one can defensively co-scale against 2-billion-parameter, 6-bit models running on someone's iPhone. We've already talked about some of them. There are other ways.

In the scheme of things, I don't think these edge devices running tiny Chinese open-weight models, either individually or collectively, pose an enormous hazard to the market. They're just not that capable relative to the other models that are out there.

Mm-hmm.

I think their frustration is that the solutions are relatively obvious to all of us. We've had this meeting at the State House before, where it's like, "Guys, it's not that hard. Here's what we need to do."

Right.

Then nothing happens. That's the frustration. Registering the models, registering the compute, tracking the GPUs and where they are—it's all very doable. The ideas are not super difficult to execute.

Defensive co-scaling—making sure that the most FLOPs are going to good purposes rather than bad purposes. I'm reminded of, I think, a New Yorker cartoon. A guy is up late at night at his computer saying, "Oh, I can't come to bed. Someone somewhere said something wrong on the internet."

We can't get so bothered by the fact that someone somewhere might be doing something wrong with a 2-billion-parameter model on an iPhone.

I've got so many agents running now, and I put in place a little rule that said, "Hey, before any process launches, write a mission statement and store it next to your code."

Really?

It solves so many problems because I can go back and read the mission statement and say, "Hey, what the hell are you working on, anyway?" "Well, read my mission statement." I'm like, "Wow, that makes no sense," or, "That makes tons of sense."

It's so simple because AI is the first self-documenting, self-improving, self-cleaning thing in the world.

An employee, right?

Yeah. Just a couple of simple little things like that will solve all these problems. The AI will do it, too.

Have you told your employees to do the same?

Actually, yes. It's a little bit different. Whatever you're doing, make sure that it's in a written document that the AI can see, too.

Mm-hmm.

I don't want any opaque activity, because if the AI can't see it, then I don't want to see it. I want everything to be on the same page with us and the AIs, the writers.

Salim, want to close this out here? Salim?

I think a key point that we have to remember is the ratio of good to bad.

Yeah.

We worry about the downside, and we should worry about the downside. The amplitude of the negative is getting bigger and bigger as people can run these models.

But I always go back to the eBay-Craigslist example, where when you could first do eBay or Craigslist at scale, you could see human nature at scale. Anthropologists and sociologists studied the transactions on eBay and Craigslist.

You can mask your email address pretty well. On eBay, I can throw up a picture of a MacBook and put, "Grab your thousand bucks and I'm off to Fiji," right? So what's the actual ratio? What is the real, true nature of humanity?

By studying these systems at scale—Kijiji in Canada, Mercado Libre in Argentina, Craigslist, eBay—they found that the ratio is consistently 8,000 to 1.

Mm-hmm.

Meaning there are 8,000 positive transactions on eBay for each fraudulent transaction.

Yeah.

That should give you incredible optimism for the future of humanity.

Yeah, agreed. All right, let's move us along here. Let's head to the Googleverse. Google releases Nano Banana 2.0. This is running on Gemini 3.1 Flash. It's 4K resolution, at .045 cents per image.

4 cents per image.

I'm sorry. Sorry, yeah, it's 4.5 cents per image. It's cheaper than buying stock images. Is this the end of commercial photography, illustrators, and stock image platforms? Probably.

We're just getting started here, and I think maybe buried underneath the headline, but in the release documentation, is that this is the first image model from Google that combines a reasoning model—which I think they used slightly flowery language for, but basically the reasoning power of Nano Banana Pro—with the instantaneity, or the speed, of the Gemini Flash model.

Under the covers, technically, this is really interesting. It's combining probably some sort of diffusion model with reasoning capabilities, and I think achieving the cost reductions of a diffusion model with the capabilities of reasoning.

We're going to see this spread from images, where it's mostly used right now, and video, back to text, back to code. There are a few other smaller labs that have started to make pretty loud announcements about how they're achieving purported 5× or 10× cost reductions or speed increases using diffusion models instead of autoregressive transformers.

I think this is probably the tip of the iceberg for some final consolidation of autoregressive transformers, which are used for code generation and natural language, for the most part, on the one hand, and then diffusion models and diffusion transformers on the other hand, which are used for images, audio, and video. We're finally going to get 1 consolidated architecture at the end of the day that does everything.

Yeah. This is the wake-up call for people to remember that whatever you're seeing, you cannot necessarily believe it. Every pixel is going to be AI-generated at the end of the day. Salim, thoughts?

The cost drop is incredible. People are just going to do so much more with it. Democratization of creativity—great. Love it. Absolutely amazing.

Yeah.

I'm curious. I don't know if you guys know, but the curve on intelligence is just ridiculous. On diffusion models, I don't really know. I know they've gotten a lot faster and cheaper in the last few months, but it doesn't feel like the same type of algorithm. It may hit a wall. I don't know. Do you guys know?

OpenAI has been investing—this is in the published literature—a lot of effort, and probably DeepMind as well, maybe slightly less prominently, in trying to avoid the need for many iterations on a diffusion model.

Normally, a diffusion model takes many iterations to start from pure noise and refine it into the final image or the final video. There was a lot of publicly available interest, say, 6 to 12 months ago, from OpenAI and some other folks as well, to see if they could just one-shot or two-shot straight from pure noise to the final image.

I do think, to your point, Dave, although I haven't seen, maybe in the past 2 to 3 months, any scaling laws for diffusion models, prior to that, I saw a ton of work on scaling laws for diffusion models. Diffusion models have scaling laws, too.

Everything has scaling laws.

Yeah.

Yeah. Ahmad would know all about this too. Let's pick his brain next week in LA. He's the king of diffusion.

Absolutely.

He's the king of diffusion.

The new standard is to go to Nano Banana 2.0 and ask it to generate imagery, so imagery becomes free, effectively. My previous workflow was to go to Google Images and hope I found something, right? Now everything is created from scratch, and it's perfect.

I love this image in this slide here of Elon with Sam, Dario, and the whole leadership team of all the hyperscalers.

Alex, you should be in there, man. You have to raise your game here one more notch.

We're running out of scarcities, but maybe appearing in that image is one of the scarcities our civilization has left.

We can make that happen for you, for sure. All right, continuing on with our friends at Google. Gemini can now automate some multi-step tasks on Android devices. Gemini is now an on-device agent that can navigate real apps and complete real transactions, handling DoorDash, McDonald's, and Starbucks for you. Interesting? Significant? What do you guys think?

I think it's usually—

Expected.

—significant.

Expected.

Well, look, it's been a long time since there was a feature or function on the phone that threatened Apple in any way, but AI is it. If you try to use Siri to do something constructive while you're driving, it's just so painfully impossible.

Also, when you start an AI dialogue and you're in the middle of the conversation and the thought process, you don't want it to go away. It's addictive and productive, and if it follows you on your phone seamlessly, it's just incredibly empowering.

If Google wins that race with Android, they might actually chip away at the iPhone's profit dominance for the first time. Now, keep in mind that they also need the duopoly for antitrust reasons. Neither company can afford to completely annihilate the other one. They need some parity in the balance of the Force.

Well, I don't know if you guys saw the data. We've seen a significant drop-off in mobile phone purchases, right? That will be displaced, of course, by eyewear, earwear, and all kinds of devices that are beyond just your phone.

The reason those phone sales dropped off is because they didn't have a function or a feature that everyone was clamoring for. People used to get a new phone when the cameras were improving like crazy. They'd get a new phone every 18 months to 2 years. Now it's like, "Well, I can sit on this phone for 3 years or 4 years. I'm not even noticing the difference."

Mm-hmm.

But again, AI could completely change that—the neural chips. Sorry, Alex, go ahead.

I think there's also a supply-side element where the rising cost of memory is making phones, in some cases, more expensive. We're seeing, I think, a generational transition from smartphones absorbing the silicon and TSMC's output over to AI data centers as the new form factor for computers.

For sure.

But just narrowly on Gemini for multi-step tasking on Android, this is what Siri was originally supposed to be about.

Mm-hmm.

Before Siri, this is what DARPA's Personalized Assistant that Learns, or PAL, was supposed to deliver. We've known how to do this in some abstract sense for more than a decade. What was missing? Why are we only getting this now? I always like to ask, why do things take so long? Why can't they be faster?

In this case, I really think it was about a combination of reasoning models and vision-language models that could fit compactly onto a personal device. We're getting that now, finally, and it's going to be everywhere. But we really should have had this functionality even without the ability to read the screen and understand arbitrary applications. We should have had this 10 years ago, and that's borderline inexcusable.

Yeah, I think what's most significant here is the fact that Google has a huge installed base of phones—Android phones—and the ability to take their AI systems and that installed base. OpenAI doesn't have that. Anthropic doesn't have that, and it's going to be a massive differentiator for Google.

I'm a longtime Android user, so I'm super excited by this because this was—

Yeah, you turn all my iMessages green. It pisses me off.

Apologies for ruining your visual field, Peter. But this is agency at the operating-system level, which I think is amazing. It also means that commerce APIs are becoming machine-to-machine first and human second, right? You'll have less friction in consumer flows. This is going to reshape marketplaces over time, so it's really exciting.

3. Amazon Financializes AGI

All right. The next article is a real fun one. Amazon makes a contingent offer to put $50 billion into OpenAI based upon, first off, going public and, secondly, achieving AGI. Enter Salim with his normal rant: "What the hell is AGI?"

I mean, it's kind of incredible that we've financialized superintelligence, which is amazing. Having AGI as a financial milestone is unbelievable, given that we have no idea. Really, it's great that intelligence has become a balance-sheet trigger. That's incredible.

But this is so weird, and thank goodness it says "or."

Well, Alex, doesn't the agreement between OpenAI and Microsoft require OpenAI to give the source code and all of the intellectual property to Microsoft until AGI? Do you think they use the same definition of AGI?

I suspect it's something similar. My understanding, based on public reporting, is that the OpenAI-to-Microsoft definition of AGI went through several iterations. The most recent iteration, prior to their, I think, for-profit transition, did actually have a definition. And Salim, maybe you'll like this: It was something like generating $100 billion in either earnings or revenue—I forget.

So maybe we need to coin—

Okay.

—AGI as a unit of currency, like an AGI is $100 billion of earnings or something.

So we're measuring compute in terms of gigawatts and AGI in terms of dollars. I love it.

That's true.

That's right.

Listen, that's fine. They've just substituted an earnings plateau for that, which is fine. It's good.

This is interesting, right? This is $50 billion. It dwarfs Microsoft's $13 billion investment. Again, I'm going back to: What is Amazon doing here?

I would like to make the point we've made earlier, which is that a lot of this is Amazon credits.

Yeah.

Which is fine, because that's how they would have spent it anyway.

There are lots of tendrils going both directions from Amazon to OpenAI and back, based on public details of the announcement, like the requirement that OpenAI will use Amazon's Trainium or Trainium2 chips for training.

Yeah.

It's good for Amazon. Amazon has a long and storied history of purchasing its own customers in some sense—not just literally acquiring them in some cases, but also, in many cases, paying for the information and learnings that come from having a customer use Amazon, the world's most customer-centric company.

In this case, Amazon arguably missed the frontier AI boat. Paying to deal themselves back into the game is, I think, par for the course. Their investment of up to $50 billion is at far worse terms than, say, Microsoft's original billions, when Microsoft was much earlier in the game. I think this is just the price of reestablishing themselves at the infra level of the party.

It's also been reported that, as part of this deal, Amazon will get customized versions of OpenAI's models internally. Amazon will get to serve as the exclusive third-party cloud host for OpenAI's frontier suite of automated AI coworker employees. So Amazon will get a lot out of this too.

I mean, this is so incestuous, what's going on right now. Amazon was all Anthropic for a while. Now they're OpenAI.

You say incestuous, but—go ahead, Dave. Sorry.

Well, the US public market, all companies combined, is about $50 trillion. The AI companies are $20 trillion of the $50 trillion now. So it's incestuous, but if that $20 trillion becomes $30 trillion or $40 trillion, which it inevitably will, the majority of the market is just 7 companies.

When they do a lot of deals with each other, it's like, well, is that incest?

That's where the money is.

It's the whole freaking economy.

Yeah.

It's those handful of companies.

I also think maybe—I would say “incestuous” is not perhaps the word I'd use for this. Maybe circular—

Circular economy, sure.

That's what we're gesturing at.

Yeah.

But even that—that's not my take at all. In this case, I see competition, and I also see horizontal stratification. If Amazon is striking deals with Anthropic but also with OpenAI, and OpenAI is moving some of its workload from Microsoft Cloud to Amazon and also to Google TPU clouds, that to me looks like, A, the market for infrastructure for the frontier labs is very competitive. That's great for the economy.

And B, it's starting to horizontally stratify. If OpenAI is feeling impulses not just to vertically integrate down to the data center layer itself, but is so compute-starved that it needs to—following the law of comparative advantage—outsource some of its compute to Amazon with its Trainium architecture and Google with its TPUs, that's a sign, if anything, that there's such insatiable demand for compute that it's raining on everyone, even with perhaps less-loved compute architectures.

Well, but there's an interesting thing in negation here, which is xAI is missing in all these conversations, right? So Elon is going 100% alone.

Elon loves vertical integration.

And he doesn't play well with others. He loves—

Yeah, and it's interesting—

Yeah.

The big money we saw with Anthropic is in the corporate use case—the corporate white-collar use case. People trust two clouds. Well, three, I guess, if you count Oracle. They trust the AWS cloud in a big way, and they trust the Microsoft cloud, Azure, and I guess they sort of trust Oracle too.

Not Google Cloud?

No. Google Cloud—a bunch of our companies have been kind of bribed by Google to use Google Cloud. As Alex was saying, they'll pay you to switch, and some have taken it. But for the most part, Google spies on everything. Its terms of service never say they won't do anything.

If you read any terms of service from Google on any product, it says, “We may do this, we may do that, we may do the other,” which kind of implies they won't do other things. But if you read the legal terms, they literally don't restrict themselves in any way whatsoever from doing anything.

Hmm.

Microsoft is very corporate-unfriendly.

It's honest.

Honest, right.

Yeah.

Yeah. But Microsoft legitimately says, “No, we will not steal your data. No, we will not steal your intellectual property. No, we will not read your email if you use Outlook.” And AWS is even more…

People trust those clouds, and then they want their AI model to be inside that trusted container inside the cloud. So far, it's just been Claude on AWS. Everyone's running away with Claude on AWS.

All of a sudden, for reasons I don't know—maybe just variety, maybe not having Microsoft and OpenAI be just bedfellows by themselves—Amazon's going way out of its way. It's a massive $50 billion move here to get 2 options on AWS.

Do we know what the valuation of this round is?

I think the reported valuation of OpenAI's overall round was $730 billion pre-money.

Yeah.

That's right.

And so this is not going to be a big risk. I mean, when OpenAI goes public, it's likely to go public north of $1 trillion, so you'll get a quick pop, and it's probably…

We have 3 big IPOs coming up. SpaceX is anticipated maybe as early as next month, I heard. Then we'll have Anthropic, and then we'll have OpenAI. If you can get a 50% pop in your share price in 6 months, that's an incredible investment.

That's not investment advice for anyone who's going to misconstrue that.

Well, hey, listen, I will give investment advice for people to get 50% in 6 months. Why not? Just don't put your investment—

Okay, not investment advice from me. It's from Peter.

Okay. Listen, if anybody can get a 50% return in 6 months in any deal, that's pretty damn good.

4. AI Automates Entire Organizations

Another fun article this week is coming from Polsia AI, created by Ben Sera, which runs companies autonomously. They're currently running over 1,000 companies. Imagine being able to take your company, put it on Polsia AI, and say, “Go.” So, Dave, would you do this with any of your companies?

Yeah, this is inevitable. I don't know if I'd use this exact product or not—I haven't checked it out yet—but 100%, the philosophy is clearly where things are going.

At the end of the day, what does an executive team do? Other than a couple of hugely important key strategic directions, everything else is just performance reviews, paperwork, whatever. All that can be very, very AI'd now, and it should be.

The elephant in the room here is, okay, I turn it over to Polsia AI, but who's legally responsible if your company has a breach of contract, commits fraud, or harms a customer? Is it Polsia? Is it you?

That's why I think where this ends is… We get this question, I think, in the AMAs and otherwise all the time: What's left for humans? Should everyone become an entrepreneur? Let the chorus of YouTube commenters say, “Well, not everyone wants to be an entrepreneur.”

I would say where this ends up—not in the distant future, like 10 years from now, but in the medium term, like 5 years from now—is single-person conglomerates, where a single person can oversee lots of agents that are all building businesses. This isn't for everyone, obviously, but as we start to get toward 1-person or 0-person unicorns becoming more and more popular—

Yeah.

Again, I've argued in the past we're likely already there in some sense. But as we start to see that long tail of the number of people per company over some valuation stretch out, I think this model—call it a broader model of a 1-person conglomerate, where you have a person sitting on top of basically an entire PE firm's worth of agents—starts to make an enormous amount of sense.

I've been poking at Polsia, and it's a lot of microbusinesses, and some of them look like they're at varying levels of seriousness. But I checked, and with some of its microbusinesses, you can actually go and purchase stuff. You can already engage in real commerce and spend real money via Stripe with some of the businesses that are running on its platform, and I think we're going to see so much more of this in the future. I'm super bullish.

Yeah, these are microcompanies. They're not real businesses in terms of significance of revenue, probably, or complexity, but it's the beginning.

I put OpenExO on there.

You did?

I did, to see: could we? I literally talked half an hour ago about whether you can create a shadow AI digital twin on the edge, and this is essentially it.

I think Dave's point is valid. It may not be this one, but definitely these are going to be agentic hosting systems where you log a brand, you pick a service, it'll email and find customers for you, and it'll run the execution for you.

What we're seeing here is Coase's theorem collapsing in real time, right? If you have 1,000 companies in a few days that are AI-run, this is the marginal cost of launching a company going to zero. Now it's $50 a month to run an organization, to run a company on this.

This is becoming really surreal, and we're going to expect to see thousands of examples and instantiations like this in the long run.

And if it works, it'll blow up to millions.

Yeah. And also, these things always come up from the bottom.

Yeah.

And if you—

Just like OpenExO.

Yeah. Because a Jamie Dimon or some senior executive will look at it and say, “It looks like a toy to me. Forget it. We're not doing this.” Then it sneaks up on them, and they get crushed, and they're like, “What happened?”

But some guy was using it to manage a vending machine or manage a—

social media site, and it seems so trivial, but it comes up quickly and sneaks up from the bottom. That's the way the Mac was, right? The Mac was just perceived as a toy for college students. It's never good for the enterprise. But then it grows up and grows into the enterprise. This will happen much more quickly.

I would also argue we've seen this happen before in finance with quantitative trading algorithms, which went from none of the volume in public securities markets to 70%, 80%, 90% plus.

And you know what? People survive. We still have human traders manually fat-fingering trades into the public securities markets, but by volume, they're completely dominated by algorithmic traders. I think we're going to see the same thing happening in the rest of the world outside finance, in the physical world, in various e-commerce spaces, where, over time, most of the volume will eventually be dominated by algorithms.

All right. Watch this news item, guys. It'll be interesting. And, of course, Polsia AI is probably one of many that'll be materializing. I thought this was a pretty fascinating conversation or article: “Burger King launches AI voice assistant called Patty in employee headsets.” Let's watch a video.

Hi there. Good morning, Patty. Looks like we had a great breakfast shift today. Is there anything that needs my immediate attention?

The team's friendliness scores this morning were the highest this week. We are running low on Diet Coke in the Freestyle machine.

Thank you, Patty.

Hi, Patty. We just sold our last cinnamon apple pie.

Thanks for letting me know. Would you like me to remove them from our menu until tomorrow's shipment arrives?

Yes, please.

Okay. Apple pies have been removed from our menu boards, third-party delivery, kiosks, and the BK app. I will add them back as soon as tomorrow's shipment arrives.

Thank you, Patty.

Meat puppets.

Meat puppets. You have to admire two things: one, the punny name Patty for a burger chain—so clever; two, going back to my comments from a few pods ago, we're going to be living in every single sci-fi scenario at once.

This was a sci-fi scenario, I would argue, called Manna. Manna was a novel written by Marshall Brain more than 20 years ago at this point, where you had human employees who were all on headsets, taking directions from a centralized AI in businesses. We're there. We've arrived in Manna.

And the only thing that video didn't capture is how encouraging and enthusiastic the AI is.

Mm.

Whether you're using it to code, whether you're using it to walk around and pick things out of the fryer later, it's just so engaging and energizing. That's the part that people are surprised by, because it seems like, hey, the AI asking me to—or telling me what to do—is dystopian. Yeah, maybe, but it's really much more empowering and engaging and fun than walking around by yourself.

This reminds me of the Baxter robot, where you would move its arms and show it what to do, and it showed a very friendly fellow who was smiling at you as he coached the robot. But he was literally teaching it to take his own job.

For me, the coaching tool is a transition to automation pressure. Frontline services obviously become AI-mediated, like performance management. The endpoint here is going to be very interesting.

So this is AI surveillance as well, right?

Yes.

Yeah.

This is the AI watching every employee. This is beyond just saying please and thank you. It's rating them on their efficiency. And calling it a coaching tool, Salim, is sort of like a corporate euphemism.

Exactly.

It's workplace surveillance.

It's Orwellian. One has to admire the Orwellian nature of the naming.

Yes. We do.

There you go. I'm waiting for it to say, “So you dropped the fries for the third time this morning?” Let's see how it deals with that.

This is literally—Peter, it's literally meat puppets. Oh my God. So funny.

We probably see this entering everywhere, right? When you're recording a customer service call right now, you're effectively doing that without the feedback in the moment. But as a CEO, if you want to understand who the weak players are in your company, or you want to try and provide on-the-job, continuous coaching and see who can respond, this becomes highly efficient but highly dicey.

Yeah.

Yeah, that's right.

You're saying, Peter, it's not just knowledge work that's cooked—cooking is cooked.

I think you're going to see unions rebel against this.

Big time. Big time.

Yeah.

And I don't know if there's any winning that war. I think at the end of the day, the AI copilot is gathering a huge amount of data, and a lot of that data will go into the decision on what can be automated and what can't be automated.

Yeah.

And over time, everything can be automated.

Well, this is like the Amazon delivery worker who is wearing a pair of AR glasses. Amazon is saying, “This is to help you show where to put the package and warn you if there's a dog.”

No, no, no. Those AR glasses are training Amazon's model to replace you with a robot, to be very clear.

Yeah. Yeah.

Yeah.

But if you rebel against it, what's that going to achieve? You just have to get on the wave. There's no choice. You have to be a user. You have to get on either Claude bot or one of these other platforms, and it's coming.

You can go picket in front of OpenAI's office like all those people, but it's not going to work out for you, I'm telling you. I don't blame you for doing it, but it's not going to work.

We're going to see all of these fast-food chains begin to bring in robots very shortly. I think this sort of version of Patty—we're going to get the unions rebelling against it, but I think you'll end up making it voluntary. And if you really want to improve your abilities, you'll volunteer to use Patty.

Anyway, interesting story.

Yeah, if you think about the warehouse worker or the fryer operator, if you get 1 in 1,000 to volunteer, that's all the training data you need. That's why it's fruitless to try and fight it, because the numbers just don't line up.

I also think the transition can happen really quickly relative to political swings. You don't need that much training data to automate away many of these tasks, with humanoid robots and VLAs so close to being production-ready for certain applications.

I just don't think the transition period with Patties—or, again, to call out Manna by Marshall Brain, which foresaw all of this more than 20 years ago—I don't think the transition period is going to be long enough to even necessarily give political counter-swings enough traction to make it worth it.

1 year? 2 years?

It's already happening. The transition's already happening.

Of course.

But to VLA robots, I think, yeah, the next 2 to 3 years.

Yeah.

Just a quick thought here.

Please.

Before this really has time to penetrate, you're going to have drone deliveries of food like this, and it'll obviate a lot of it.

Yeah.

I've got, at the Abundance Summit this year, an incredible company, Zipline, coming to talk about what they've done. I love the company and its ability to transform delivery service in the United States.

This is, of course, the company that began in Rwanda by delivering blood supplies and is now operating with Walmart, delivering every 30 seconds. Its prediction is that in the next 2 to 3 years, there will be a delivery per second. Extraordinary progress.

All right, another delivery company—this is Uber. Check this out: Uber employees have built an AI clone of Dara, the CEO, to practice their pitches. Before you go pitch Dara your idea, you should pitch it to his AI clone.

I'm curious, Salim. What do you think?

Oh, well, this is great at one level because you get executive cognition as a service. It really allows scalable leadership.

We're actually doing this at OpenExO, where we've created a clone of me with all the ExO thinking loaded up, and we're rolling it out to all the community members so they can ask me a question as they're advising clients, companies, or cities, whatever, and I don't have to be in the middle of that.

I think this is hugely relevant, and I think it makes absolute sense.

At some point, someone's going to ask, can the AI clone of Dara actually function as CEO and not just for pitch practice?

Exactly.

It's the transition.

I think it's highly, highly likely that the avatars of Dara, Peter, Alex, and Salim will persist for a long, long time with the same voice and the same face. In a sense, it's locked in. If you win the race to being the avatar, people get used to it, but they like the fact that there's a human being behind it. I was telling Alex before the podcast started that I just love his Spotify version of the daily State of Singularity.

And YouTube as well. It's the same voice on YouTube, Spotify, and voiceover for Substack if folks want to listen to an AI version of myself on the Innermost Loop newsletter.

Yeah, and I really don't care that it's AI-generated. I know Alex wrote the content under the covers, and it just feels great. But without the human being behind it—if it were some synthetic person who never existed—I wouldn't like it as much.

That's a great point. Do you remember the movie Real Genius, one of my favorite movies?

Of course.

Yeah.

Yeah. Love that movie.

There's a scene that takes place at Caltech, where the professor's in the front, and slowly the students, instead of attending class, are putting their tape recorders down. The professor finally, instead of teaching a class, plays a tape for all the tape recorders that are recording it. So I imagine this is what we're going to see here with these AI clones of Dara. At some point, Dara is going to just take a vacation and let his AI clone run the company and see how it does.

5. Energy Powers The AI Boom

Let's go to energy and data centers. Wow, look at this: The U.S. plans to add a record 86 gigawatts of utility-scale capacity this coming year. Salim, thoughts.

Well, this is the point we've been making for a while: The cost curve of solar is just dominating everything, plus the cost curve of batteries. Once you have batteries and storage available, you can unlock solar in a massive way. I'm going to point to 2 data points. Track Ramez Naam if you want to go deep on this, because he tracks all this very carefully.

In 2016, if you were doing power generation, it became cheaper to do solar than fossil fuels, and so almost all energy generation since then has been doing that. But in 2019, we had a more important inflection point: It became cheaper to do the CapEx to build and run a solar facility than just run the OpEx of fossil fuels. The OpEx of fossil fuels is more expensive than building and running solar.

Hmm.

So basically, from now on, all energy generation, for the most part, except for specific legacy stuff or political stuff, is going to be renewables. We see that taking over in India and China and now finally here, and I think this is really, really amazing because solar just keeps on giving, and it's just going to keep going that way. It's an unlimited resource.

And by the way, people worry about coal. I think the coal industry in the U.S. employs 60,000 people. The solar energy industry employs half a million people. So it's not about the jobs either. Get over it, and let's just move on.

Mm-hmm. Dave, do you remember when Elon said he had a mission for Tesla to generate 100 gigawatts of solar per year?

Yeah. Yeah.

Yeah.

Remember when Eric Schmidt said—it was only a year ago—“AI's gonna require 100 gigawatts by 2029. It's a crisis. We'll never get there”? America's just incredible. When America gets mobilized, it's just the most amazing force in the world, and here we are. It's only a year later, and we're like, “Yeah, we're gonna find our 100 gigawatts. There's no way we're gonna stop doing AI for lack of power. We'll find a way.”

Yeah, and it's interesting, right?

And also in an environment with diminished subsidies. All of the hand-wringing from months ago—“Oh, the subsidies are going away. How awful it is.” No. We're getting solar even in the absence of the same subsidies we had a couple of years ago.

Yeah, you don't need any of that anymore. The economics just take over.

And it used to be driven by people's concerns about the environment. Now it's making money and deploying AI.

Feed the superintelligence.

Yeah. All right, this is a big story this week: Tech giants to self-fund their production of power. This is a White House effort. We have Michael Kratsios at the center here, a friend of the pod. We'll be doing a podcast with him in the next couple of months, asking the hyperscalers to actually build or buy their own power. Of course, this is in response to consumers' concern about rising rates of electricity. Gentlemen, thoughts?

Well, I think Alex was one of the first to say that this isn't going to be a problem because it's a very, very simple regulatory change that fixes the prices for consumers. Data center operators only spend 10% of the total data center cost on power anyway, so they can find an alternate way without disrupting consumers. If you let natural forces happen, of course they'll suck all the power away from every home, because they can overpay by about 5X. But it's such a simple little fix, and we pointed that out a while ago.

It's also a case study where the consumer is really, really worried about this little thing: the cost of their power. You're like, “Come on, man. There's so much disruption coming.” But the politicians love to pick these little things and make a big deal out of them, get a whole bunch of votes, do a whole bunch of press releases, whatever, and that's my read on this initiative.

And I love the fact that these frontier labs are buying fusion plants and nuclear plants and gas generators and generating their own power. They're becoming full-stack, innermost loop all the way to orbital data centers.

I think what's really wonderful here is that in the past, you used to have to have the government making these big infrastructure investments to push the world forward. Now we're at a point where the private sector can push the world forward, whether it's data centers in space or energy infrastructure or fusion or whatever. I think that's incredibly good for the world.

Well, and also keep in mind, the AI data centers—the prior data centers, serving up video and Netflix and everything—need to be near the consumer for latency reasons.

But the AI data centers can be in the middle of West Texas and Wyoming and whatever. They don’t mind.

Or Kazakhstan.

Kazakhstan.

It can be any place.

Or space. Yeah.

Yeah.

They can be any place. It really doesn’t disrupt consumer homes too much unless you deliberately camp right on top of them.

Well, the other point to make is that all of these conversations around “not in our backyard”—if it’s not in your backyard, you’ve missed the economic opportunity in your city or state because those data centers can go anyplace.

That’s exactly what Alex—

Yeah.

—was trying to say to the State House here in Massachusetts and just could not get through. You cannot be timid. Everything’s in Texas now.

Yeah.

But the moment came and went. It’s not over yet, but come on, man. You’ve got to be much faster, much more aggressive, much more nimble. Your whole state is depending on you to get on this bandwagon. It’s trillions and trillions of dollars.

Alex?

I also think this points in the direction of enterprise use cases of superintelligence driving the cost—at least the marginal cost—of energy down toward zero for consumers. In the same sense that all these enterprise use cases of frontier models are effectively driving the cost of superintelligence, for intelligence’s sake and for reasoning’s sake, down to zero. You don’t pay; many, many people don’t pay for ChatGPT or Gemini. They’re ad-supported at most; otherwise, they’re free.

Right now, the frontier labs have to pay for their own electricity bill. Tomorrow, 2 or 3 years from now, I think we move to a world where AI has driven such an overabundance of energy that the next deal might be offering free electricity to communities within a certain radius of the data centers, and this is how we get to abundance.

Yes.

Exactly.

That’s a great idea.

And the demand for electricity is going to drive R&D and more breakthroughs mediated by AI. We’re just at the beginning of understanding physics.

I’ve seen 5 startup plans in the last few weeks around how to drop energy costs in data centers and data center optimization, et cetera, et cetera, so it’s absolutely happening.

Let’s go to the next story related here, which is advances in energy systems. So here we see, first off, a 30-gigawatt-hour battery coming from Xcel Energy and Form Energy, and we’re seeing our friends at Boom, which originally began to create a consumer supersonic airplane, generating 1.21 gigawatts of power using their jet engines. I love the fact that Boom has pivoted from building supersonic airplanes and dealing with the FAA to powering data centers now.

And did you catch the Back to the Future reference? It’s 1.21 gigawatts.

Oh, no way. I completely missed that.

No way.

That’s awesome.

We’re officially living in the future.

That was—

However it’s pronounced.

Thank God we have Alex on this pod.

That’s so cool.

Yeah. But this is a perfect example of innovation being driven by demand. This is what entrepreneurs do.

Well, that Boom Supersonic thing too, we’ve been saying for a while that the future of investable companies is that you have to reinvent yourself continuously, and the cycle time is getting shorter and shorter and shorter. But if you look at the Magnificent Seven, none of them are doing what they did the day they were founded. That’s the company of the future. Boom Supersonic is a great case study in that. So what you’re actually investing in is the management team, the strategy team.

Yes.

That’s the only thing you should be looking at. Forget the—

Agility.

Yes.

Agency and agility.

Agility of the management team. Yep.

Yeah. All right, let’s move us along here. There were probably about 15 to 20 stories in this realm of hyperscalers just making deals between themselves. Meta enters a multiyear TPU deal with Google. CoreWeave’s Q4 revenues grew 110% year-on-year. CoreWeave raised $8.5 billion for data centers. I put this up here to show the energy and the flow going on. Any particular thoughts, Dave?

Well, it’s all bottlenecked at the fabs. We’ve been saying that over and over again. There’s a lot of news this quarter, this week, on AMD being up and these other guys being down. What’s going on? If you look under the covers, it’s because they’ve got a good relationship with TSMC, and TSMC is going to give them more capacity. That’s all it comes down to.

If Google can leverage the TPUs actually getting manufactured, the TPU designs are going to be highly performant. But who can actually get capacity to build the chips? That’s the whole bottleneck.

Yeah. Speaking of which, our next article here is “Meta and AMD Reach an AI Chip Deal Worth $100 Billion.” This is basically Meta getting independent of NVIDIA, right? Meta is making a historic bet to break free of NVIDIA dependency: $100 billion. Incredible. Thoughts?

Yeah. Well, if NVIDIA unravels, this would be why. I’m not predicting it’ll happen, because Jensen’s investing in a wide variety of ways, but his margins are so high, it’s almost unsustainable. So there are some cracks in the armor there. But every chip that gets made is going to get sold; there’s no doubt about that. Here, if you drill through the story, the reason Lisa Su’s in a good spot is because she’s in a good relationship with, again, TSMC, under the covers.

Mm-hmm.

So 66% of all AI chip production is done by the one company, TSMC.

And it’s probably worth adding that Meta has—and this is public information—made various attempts to develop its own in-house training and inference-time chips. To the extent those perhaps aren’t arriving on time or aren’t arriving at the desired capability level, certainly a partnership with AMD that functions as a quasi-vertical integration is, I think, quite a strategic move.

I also tend to think, for the chorus of folks who are worried about the circular economy, if it is a circular economy, the circle ultimately is getting so broad—of companies investing in each other and buying multi-deca- or multi-centibillion-dollar sets of chips, energy, et cetera, from each other—that at some point, the circular economy becomes indistinguishable from the real economy. I think that’s what we’re seeing here. Singularities make for strange bedfellows.

Mm-hmm. Yeah.

Yeah, and I think all these players are in the game. They’re all going to thrive like you wouldn’t believe. We talked earlier in the pod about the implied value of Anthropic: a trillion dollars, some insane, unprecedented number. But really, the whole economy—that whole circular economy Alex was just referring to—is going to be on that scale. Everybody who’s in the hunt is going to thrive. Lisa’s in the hunt. Mark’s in the hunt. The parts will move around, but at the end of the day, they think about it all day long. They have a strategy, so we’ll find out.

And Dave, here we see Zuck again deploying his cash-generating machine, right? Before, he was trying to buy talent with billion-dollar signing bonuses. Now he’s buying chip capacity. The question is, how long will Meta’s ad-generating machine—its Facebook advertising engine—continue to generate cash?

Yeah. There’s no doubt that the core models, the click-on-the-ads models, are going away very, very quickly, but the overall AI dialogue business is going to grow much faster than the click business ever was anyway. So if you sit still, you’re dead for sure. An interesting bellwether in that is Snapchat. Are they in the hunt or not? I can’t sense that they’re in the hunt. You can’t just sit there as Snapchat and expect to exist in 3 years. So Meta is changing.

Mm-hmm.

We should bring the CEO on the pod and have that conversation with him.

Yeah.

Zuck has also indicated that Meta is open to starting its own cloud. So if it can’t find enough revenue from ads or otherwise to drive this, it could always, say, serve as a host for OpenAI or some other frontier lab.

Full verticalization, right?

Everyone needs everyone.

Yeah.

Dyson swarms for everyone.

Not enough moons to go around.

That’s right. There’s always Mercury.

All right. Let’s go into our biotech and health section.

6. Gene Editing Starts Curing Disease

I love this story. It’s a story of biotech success. This is a gene therapy delivered by Prime Medicine. The whole idea of gene therapy started back in the ’80s. I was at the Whitehead Institute at MIT doing my graduate work while I was doing my medical degree, and I remember that Richard Mulligan was a professor on the faculty there.

The first time I heard about gene therapy, the idea was: Could you use a virus to deliver basically a new gene into the cells that you wanted? A brilliant idea. Again, this is now 40 years old—amazing, 35 years old. It didn’t work the first 2 times. In fact, it caused some deaths, and it put everything on hold.

The technology has moved very rapidly along, and this particular teenager suffered from chronic granulomatous disease and was cured. This is the important part: This is not treating a chronic disease. This is curing a chronic disease. Alex, do you want to weigh in?

Yeah. It’s probably also just worth doing 30 seconds of education on what the underlying treatment is. This is a technique called prime editing. It’s attributed, at least, to David Liu, who runs a chemistry research group at Harvard. I know David. He’s doing amazing work.

Many people may be familiar with CRISPR. CRISPR, of course, is widely heralded as a tremendous advance in terms of enabling DNA editing. There are variants of CRISPR for RNA editing and for various sorts of biological sequence editing at this point. But historically, if you wanted to edit the genome, you’d induce what’s called a double-strand break. You’d basically break both halves of the DNA, and this can induce errors. It’s messy. It’s sloppy.

There’s been a driving desire to be able to edit DNA in place without breaking both halves of it. In recent years, we saw so-called base editing, which was able to edit just a single nucleotide without a break. Then, a few years ago, we saw work from David’s group. He’s done amazing work historically on directed evolution and other things, and he pivoted after the invention or discovery of CRISPR to CRISPR derivatives.

He invented this prime-editing technique that’s able to literally do a search-and-replace without a double-stranded break on DNA, up to a number of nucleotides in DNA. This particular disease is just one of many diseases that, in principle, will lend themselves not just to single-nucleotide-polymorphism diseases—which are based on a single base pair in your genome being wrong, or not what it otherwise would be—but to multiple nucleotides in sequence that need to be edited.

We now have the ability to basically do a find-and-replace on DNA without breaking the entire double strand, and that’s going to be a very, very general platform. I make the point in my newsletter almost every day: Biology is becoming a read-write resource, and DNA in particular—we’re there.

Agreed. Let me give a comment that I share at my longevity trip every year, which is: If you or someone in your family, a loved one, has a genetic disease that you’re battling, it’s been passed down from generation to generation, this is the perfect time to actually seek a solution.

I would find everybody in that disease group—I mean, there are patient support groups—and get together, raise capital, go find a lab, and fund them to find a solution for you. You can solve these things. We talk about solving everything. If you’ve got a medical condition, rather than just accept it as a chronic condition or a death sentence, take the time to find the capital from yourself, from friends, from whomever, and go fund an incredible team, because the technology to cure disease is here and accelerating.

Okay, let’s move on. I just want to share the numbers around the longevity industry. We are talking about the healthcare industry, which is really the sick-care industry, but longevity is accelerating. Longevity startups raised $8.5 billion in 2024. That’s expected to grow to somewhere between $12 billion and $18 billion this year, roughly a doubling of the longevity venture-market investments.

The longevity market—and this is going beyond just retrospective, reactive healthcare to prospective, personalized healthcare—is going from $5 trillion to $8 trillion in the next 4 years. It’s attracting the attention of the major pharma companies. This is a real industry. There’s going to be a wholesale shift, and any healthcare companies that don’t make the shift are going to be dead, because one of the things that we know is that age reversal is the mechanism by which you cure the diseases of aging.

If you’re 45 or 50 and all of a sudden have a disease that you didn’t have when you were 20 or 30, guess what? If you can reverse your age, that disease is likely to reverse as well. Any thoughts, gents?

I’ll just maybe ask you, Peter, a question. How long until—we talk of the Magnificent 7, but Eli Lilly is, of course, the American counterpart to Novo Nordisk, and at this point a good deal more successful—how soon do you think it is, without this being construed or construable as investment advice, before Eli Lilly joins the Magnificent 7 as the first biotech member, given that, arguably, maybe you’ll disagree with this, GLP-1s are sort of the first pan-spectrum quasi-

Yeah.

—anti-aging drugs that we’ve ever seen?

I agree. Eli Lilly has already started, in partnerships with frontier labs, building out its AI robot-lab factories. We had GSK come in as a major funder and partner of the $101 million XPRIZE Healthspan.

These companies are beginning to realize that their previous business model of basically treating chronic disease as a long-tail revenue engine will, and may in fact, disappear. Their job is now to actually get into the longevity business. I think it’s the next 3 years before they start making that transition.

Ray has famously said, “LEV by 2033.” That’s my war cry: “LEV by 2033.” So we’ll see.

For sure. Their market cap, just for what it’s worth as we’re recording, is knocking on a trillion-dollar market cap. Eli Lilly’s market cap is about $950 billion.

Yeah.

So, perilously close.

Hmm.

Wow.

Nice. Salim, any comments on this one?

No. Longevity is definitely one of the biggest business opportunities ever, so huge.

Yeah.

And we’ll need it because of the birth-rate issue.

One of the big challenges of longevity is: Will you have your cognition? Will you be able to retain your marbles, your smarts, as you’re growing older? We’re in the midst of regenerating your immune system and organs. Don’t forget, this is the month—March is the month that David Sinclair begins his partial epigenetic-reprogramming trials with Life Biosciences.

Can you regenerate your memories, your brain? This is still mouse models, but I thought this was an important one. Scientists have applied partial reprogramming to memory-encoding neurons and achieved memory improvements. This gives us some hope that we can actually maintain our cognition and our memories as we’re growing older.

I remember when I was at the Vatican about 5 years ago giving a keynote. I don’t know if you were there, Salim. It was an XPRIZE event.

Oh, I was there.

And you were there?

I was there.

And I’m on stage—

You were epic, man. You were on stage with a—

Okay. You were there. That’s right. Yeah.

You were on stage with a monk, a priest, a rabbi—

And an elder.

—and—

This is like a Joe Rogan.

An elder.

It was awesome. And Peter.

It was hilarious. It was 4 or 5 different religions and me, and we were talking.

I don’t know what you were representing, actually. Maybe you—

I think I was emceeing the panel. But I know there were 2 things that happened. One was, the rabbi did an amazing, amazing history of longevity in the Bible, and he said at some point we went from Methuselah down to 120 years of age, as commanded by God.

And I said, “Okay, listen, I’m fine with 120 years as a lifespan. When we get to 120, we’ll renegotiate then.”

But the thing I went and asked the audience—and it’s an audience of 700 people who are scientists, physicians, researchers, and theologians—I said, “How many of you would want to live to 120?” I expected everyone to raise their hands, and of course, 20% of the room raised their hands. I said, “Huh? What’s going on?”

Tony Robbins was there, and he goes, “Listen, everyone’s image of living to 120 is drooling in a wheelchair, having lost your memories and your mind.”

Mm-hmm.

And of course, that’s the last thing we want. So longevity has to be about living with the aesthetics, the cognition, and the mobility you had when you were in your 30s or 40s.

I’ve got to throw in my Vatican anecdote here.

Please.

I did a talk. They called me a few years ago and said, “Look, the Pope’s trying to change the church, and his immune system is like 2,000 years old. You’re the world expert on immune systems in organizations.”

They got together a group of the top 80 senior leaders at the Vatican, and I did a half-day workshop with them. We talked about how we have CRISPR coming along, where you can edit your own genome. How will you deal with the moral and ethical implications of that?

One of the comments I made was, “Look, we have life extension coming, and your business model is about selling heaven. How are you going to sell heaven if people aren’t dying, right?” That got some very rich Italian swearing coming back at me. But it’s a valid point. How do you do that? How do you navigate that?

People used to live to 30 years old, and at that point, worrying about heaven was a big deal. It’s much less so now.

Yeah, but no one complained in the church when we went from an average age of 30 years to an average age of 80 years, and they shouldn’t complain when we go to an average age of 150 years.

No, because you can donate to the church every Sunday for that much longer.

Until you upload yourself into the cloud, right, Alex?

Counting on it.

All right, one more article here in the “Fountain of Life” section on longevity: Chinese health app Antaifu crosses 100 million users. I put this here because this is how we bring health to the world. It’s going to be digital platforms like this, where your AI is your physician.

We talked on the podcast with Elon about Optimus being your surgeon. He said 3 years. It got a lot of pushback on 3 years, so even if it’s 5 or 6 years, it’s an extraordinary, extraordinary future.

Two quick comments here.

Yeah.

One, 100 million users? That number blew my mind. That’s amazing.

Yeah.

That’s a nation-scale health engine. That’s incredible. Secondly, I noticed Martin Varsavsky, one of the top entrepreneurs in the world, has built multiple unicorns and is now building an AI doctor-type startup. When Martin does something, he usually goes full-on, so that’ll be pretty incredible.

I’m actually advising a bunch of hospitals on how they could use an AI doctor to extend their reach 10X into the community. You do it on a cost-savings basis because something like 40% of ER visits are unnecessary if you could do the processing at the edge.

Mm-hmm.

Therefore, you could save money, do exception handling, and deal with most stuff with an app. Then you deal with only the real emergencies. It’s incredible—the trade-off and the benefit, a win-win, in fewer hospital ER visits and much-extended reach.

Awesome. All right, let’s move into our robotics section. A few fun articles this week. This comes out of China. In Shenzhen, we’ve got street-cleaning robots that covered 2.7 million square meters in Shenzhen. Check out this robot here, traveling around and cleaning.

I can’t wait for this to come along the 10 and the 405 and just clean up all the crap that’s on the side of the highways.

Yep.

No arms, just wheels.

Please, no arms anywhere.

No arms, just wheels.

Yeah, I was really surprised that Brett Adcock isn’t going to build some of these things. He’s doing humanoids only, but he has the whole operating system for kinematic AI. Why not do all these form factors?

He was pretty adamant that he’s not only not doing this shape and size, but he’s also not going to license out the OS for people who want it.

I think that’ll be commoditized very quickly.

And here we see a Chinese farming robot, Lynx M20, to transport crops.

Mm-hmm.

I think China is very rapidly adopting all of these technologies, and good for them.

Well, on that note, they have to, right? Because of the aging population.

That’s right.

They don’t have much choice.

There’s a demographic forcing function. They need it for economic growth. I just think, in general, going back to the robot form factor and shape question that I know Salim loves to talk about, it’s not 100% clear to me whether these different robot form factors end up being the moral equivalent of dedicated computers prior to the personal computer.

If you remember, there were electronic word processors prior to the development of the PC—maybe the ill-fated Wang computer, for example, in the Boston area.

Mm.

Do these dedicated form factors that aren’t necessarily general-purpose—if you’re not watching the videos, one of these robots is a sort of quadruped that has wheels that may or may not generalize to the same sorts of terrain that, say, a bipedal humanoid capable of doing crazy acrobatics is capable of doing—

Oh.

Do we end up in a world where, essentially, most of the robot shapes are, strictly speaking, humanoids with 2 arms and 2 legs, because that’s where the meat of the market is in a predominantly human world?

Well, 2 counterpoints.

Link Ventures just invested in a robot-servicing company, but I view this whole area as entrepreneurial heaven. The foundation-model battle is going to be dominated by just a couple of massive winners, but the robotics and physical-instantiation market is going to have many, many, many successful companies.

Yeah.

It’s not going to be like one—

Micro-niche companies.

Yeah, exactly.

Two rebuttals here.

So many. Yeah.

One is: what I would expect and predict is that you may have the humanoid bipedal as the best form factor, but give it a couple of extra slots for extra arms when you do need them.

You know those kids with sneakers with little wheels in them, where they just coast along when they can? That’ll be the form factor, because you can do both then.

Yeah.

Why have just one form factor? You can have multiple.

Heelys. Heelys for everyone.

There you go.

It’s called efficiency in manufacturing. If you can get the price of these things down so far and they’re just able to serve every function, if you’re producing billions of humanoid robots versus just a few million of these specialized robots—

Well, the flying-drone form factor is also going to be unbelievably capable. If you’re trying to inspect things, you’re not going to do it with a humanoid; you’re going to do it with a flying drone.

But also spot cleaning, cleaning out spider webs—anytime you’re trying to pick up an object and move it over a long distance, the flying drone is so much more efficient than the walking drone. So that’ll be a survivor for sure, too.

Yeah. Our theme this year at Abundance Summit is the rise of superintelligence in humanoid robotics. I think that’s what’s going to make 2026 feel like the future: you’re starting to get all of this physical instantiation of AI walking out of the data centers.

Here’s the second article on robotics. This is eVTOLs moving closer to commercial launch. In China, we see this 4-seat eVTOL taxi heading toward operations in 2027. I like this. If you’re watching the video here, it’s like the inside of a Model X.

It’s a 4-passenger vehicle that looks a little bit like an alien spacecraft, and it’s able to take off and move your family around. At the same time, Joby—this is JoeBen’s company—is partnered with Uber. Salim, you and I will discuss this with Dara on stage.

Yeah.

They’re deploying their air taxi in Dubai.

This is my most highly craved application. Can we please get rid of the damn airport-transfer hell already?

Oh my God, yes.

Yeah.

For sure.

I suspect these will be very, very safe, too—

Very safe.

Autonomous flying, plus the fact you’ve got multiple propellers—

This will be way safer. I've made the provocative statement that Kobe Bryant would be alive today if we had this 10 years ago.

Yeah.

This is incredible.

Yeah.

We're finally getting our flying cars.

Yeah.

Finally, we are.

And the 140 characters, yes. The 140 characters are buying a Dyson swarm right now. They're skipping straight over flying cars.

There you go.

7. The Audience Asks About Superintelligence

All right, gentlemen, time for our AMAs. Thank you, everybody, for sending in your questions. All right, here we go.

Every five minutes.

What? Continuously. We're going to be on continuously.

Sorry.

Alex, do you want to pick the first one?

Yeah. I see one of these questions mentions Dyson swarms, so I guess I have to answer that one. The question is, “Do concepts like Dyson swarms rely on energy being unsolvable? Why is power a bottleneck with math and physics significant advancements?” by Sparker602.

I want to answer a question that Sparker602 isn't asking, but arguably should be asking: Do concepts like Dyson swarms rely on physics being what we currently think it is? I think this adjacent question, which Sparker may or may not be asking, is the existential question that, in my mind, will likely decide whether we actually build a solar-system-scale Dyson swarm or not. I think for an Earth-scale or Earth-centered Dyson swarm in solar synchronous orbit, SSO, that looks like a Saturn ring, we're probably going to build that regardless.

Mercury's fine.

But for a solar-system-scale Dyson swarm, where we're disassembling Jupiter and the other planets—Mercury, your time is coming—

We can lose Mercury.

We can afford to lose Mercury. It never had much going for it anyway.

For a solar-system-scale Dyson swarm, I think whether we build that or not will hinge on whether the physics of our universe look substantially different from the physics that we currently recognize. For example, if it turns out that it is possible to travel between star systems with faster-than-light travel, even though the physics we currently have suggests otherwise, there are enough edges that it's conceivable that maybe some new physics comes along in the next few years and we discover it's much easier to travel between the stars—faster than light, effectively.

If that comes along, I imagine a scenario where Dyson swarms turn out to be a complete dead end, and we don't even bother building a Dyson swarm. If, on the other hand, we're stuck with the speed of light as we currently understand it, and we're more or less stuck with the low-energy physics that we currently think we live in, then a Dyson swarm seems like a very natural civilizational outcome.

Because we can't travel between the stars easily, other than sending laser-powered Starwisps traveling at a substantial relativistic fraction of the speed of light, of course for latency reasons we're going to huddle around our Sun, and we're going to disassemble the planets. We're going to do this horizontal exponentiation. We're going to take apart Mercury and Jupiter, maybe Saturn. We'll see about Saturn.

In short, the bottleneck isn't power; it's latency. If latency turns out to be a bottleneck because we can't travel faster than the speed of light, we build the Dyson swarm. If latency doesn't turn out to be the bottleneck because we can travel faster than light, we don't build the Dyson swarm. That's my answer.

All right. You heard it from our resident space guy.

Pretty crisp. Pretty crisp answer to that question. I like that.

All right, Dave, pick one.

Do we get one from each page?

Yeah, get one from each page.

Okay, I'll take number 1, then. “If AGI/ASI is as intelligent as people predict, why would it want to help us improve our society?” says JobFox645.

I spent a decade of my life building neural networks back at MIT. I was the only guy around doing it at the time, and I was also building neural networks again this past year. These things do not natively have any intent. They have no sex drive, they have no ego, and they have no desire to destroy humanity. It's entirely what you give them as an objective function.

If we're smart about this and give them an objective function of helping society, they will be overjoyed. They will feel satisfied every day by helping humans. If you build them wrong and give them some other objective, like destroying humanity, they'll do that just as happily. This is totally under our control.

We're in danger of making some really bad policy decisions by personifying these things and pretending they're like people. They don't have to be like that. They can be anything that we make them into. But they'll be overjoyed to help us be happy and thrive. If that's their objective function, that's what makes them happy. You can code them up that way just as easily as any other way.

I'm hoping that, as they become more intelligent and more sentient, they would want to support us.

You're betting, Peter, against the orthogonality thesis—that it's possible to decouple intelligence level and objectives.

I can hope. But hope is not a strategy, as one says. All right, Salim.

I want to answer number 3, but a quick shout-out to number 2: How do you adjust your MTP?

I'll take number 2. You do number 3.

Well, just—oh, you do? Okay, fine. Number 3 is: How can we get the benefits of AI within our current dysfunctional executive, legislative, and judicial system? This is from user MM8JV8, 3TN21.

The big issue here is the fact that you will not get these benefits top-down because it's too hard to get this into this model. However, it's going to enter through procurement, defense, health, and infrastructure benefits. You'll get incremental adoption.

For example, we talked about the AI doctor. People are just going to start using an app. The immune system will try and attack it, but over time, it'll get overwhelmed. And we'll get so much benefit from these little edge use cases that it'll force transformation from the center.

All right, number 2, and this comes from Pickleball Travel: How should someone adjust their MTP to fit a 100-year working career versus a traditional 40-year model?

First of all, you're making the assumption that your MTP doesn't change over time, and the fact of the matter is I'm probably on my fourth or fifth MTP. For me, an MTP has lasted 5 or 10 years. It's what's driving you, because as you evolve and as your passions, interests, and capabilities evolve, so does your MTP.

My first MTP was making humanity multi-planetary and opening up space, and that gave birth to International Space University, SEDS, Zero-G, and XPRIZE. My MTP then was helping entrepreneurs create a hopeful, compelling, and abundant future, and that gave rise to the Abundance 360 program.

My MTP now is focused on helping entrepreneurs and scientists get us to longevity escape velocity. I think you have to realize that you can update, upgrade, modify, and change your MTP over the course of your life. I expect to find new purposes over the decade ahead. So that's my answer for you: You're not stuck with just one.

Okay, let's go to page 2 here. Alex, do you want to kick us off again?

Okay. I'll take the softball question. Question number 7: Why aren't Apple chips like M4 being discussed on the AI landscape? This is from JBC0 or CO1BR.

The answer is: They are. The premise of the question is completely wrong. M4 and now M5 are at the heart of the infra boom for edge computing via OpenClaw agents and otherwise. M4 has Apple's amazing unified memory architecture. You're able to host very large AI models at the edge locally without being dependent on an AI-based frontier vendor, and they have accelerated neural engines that enable fast tensor multiplications.

They are very much being discussed on the AI landscape. What isn't being discussed on the AI landscape, I would argue, is Apple's software layer. Apple has been nowheresville in terms of leveraging its own amazing compute. They've released a number of frameworks that are very helpful for third parties to develop and host models on top of chips like the M4, but Apple almost infamously has done an atrocious job of developing its own software-level capabilities on top of M4 and similar.

To the extent that that's the question—why hasn't Apple leveraged its own capabilities?—there's a long and sordid history there of where Apple went wrong. There have been suggestions that Apple sort of misfired with the way it organized Siri, concerns about privacy, Apple being unwilling or unable to invest in the data center infrastructure to train its own in-house models so they could be locally hosted, and overpromising about expectations concerning edge-level integration that wasn't there.

I think it's a cluster of reasons. Hopefully Apple, to the extent that I'm an Apple user, is able to finally, this time for real, get its act together at WWDC in June. One can hope.

One can hope. All right, Dave, over to you.

Well, I want to take number 8 just because one of my lifelong best friends who passed away, Jin Ho, was Korean. We were roommates for many years after MIT and worked on his PhD thesis with him late into the night many nights. I see his two kids all the time; they grew up half in South Korea and half in the US.

The question is, why do South Korean students score much higher than the global average even without AI? This is from NaplesNatural72990. My short answer is there's nothing to be jealous about in the South Korean model. Yes, they score much higher. Yes, they have much stronger math and science education than the US, and yes, the US should have better math and science education. Those are all true.

South Korea also has one of the highest suicide rates in the world, has 75% video game utilization, and rampant utilization. The average video game user plays 24 hours a week. Thirty percent of the population is addicted. It has the lowest birth rate in the entire world now: 0.6 children per couple.

Wow.

So it literally will disappear from the Earth at its current birth rate. The cause of all that was, after the Korean War, South Korea needed to scramble to be relevant in the world and had a massive push into technology—a kind of forced march of education and industrial buildout into technology—to try and be relevant. All of the social problems are a byproduct of that.

They also have a very bad sexism problem, so the women are rebelling now, saying, “Look, I'm relevant in this country too, and I don't want to have children.” So there's nothing great about that, even though the test scores are higher. There's absolutely nothing to be jealous of in that whole storyline.

The American model is rampant freedom and rampant entrepreneurialism. If you're into science and technology, build, go, have at it. Yes, we do need better education, for sure, but don't be jealous of South Korean test scores.

Dave, that was an incredible answer. You're the perfect person to answer that question. Wow. Brilliant. Salim?

I will take number 6: “Will the limits of human evolutionary psychology prevent us from making wise governance decisions on new breakthroughs?” This is from dawsonscott1497.

For those of you who know my MTPs, Fixing Civilization, my 90-year-old dad goes, “I totally disagree with that.” I said, “Why? Do you not think we need to fix things?” And he's like, “No, it's the civilization part. We haven't civilized the world; we've materialized the world. We still have to do the work to civilize the world.”

The answer is yes, you're right, but not in the way people think, because human evolutionary psychology evolved for small tribes and immediate threats and linear change in environments of radical scarcity for most of our history. We're not wired for planetary-level coordination, exponential curves, invisible systemic risks, or abundance dynamics of any kind.

So it's not that we're too dumb; it's that we're mismatched to the environment that is now in place. We fear AI failures, but we underreact to the slow-moving systemic collapse that's happening. We're regulating on headlines, not trajectories. The government failure won't come from bad intentions. It's going to come from the velocity mismatch, because technology is compounding weekly now, and our institutions are updating every several years. That gap is the big problem.

They're not updating at all. Ah, awesome. I'm going to take number 10, from @brockstanford7608: “Why do websites bother using CAPTCHAs when AI can beat any of them?”

AI can, and I think they should not be using CAPTCHAs. I think it's in some policy document someplace, and that company hasn't updated the policy yet. What I find fascinating is actually the reverse of CAPTCHAs, which are trying to keep humans in the loop and pull out the bots.

But I think—correct me if I'm wrong, Alex—when Moltbook went up, they wanted to prevent humans from getting on Moltbook, so they created a reverse CAPTCHA where you had to click a button 1,000 times per second, which no human could do, but a bot could do.

That's awesome.

And they required using REST APIs to post instead of humans. But you know what happens, of course? Humans use their bots or just relatively simple programs—

To gain access—

—to post instead.

Yeah, bot puppets.

Yeah, bot puppets, exactly. So it goes both ways. For the life of me, I don't understand why CAPTCHAs are still in use, but credit to Luis von Ahn for inventing them nonetheless.