SpaceX AI业务增长830%、OpenAI智能体撒谎，一家初创公司把汞变成黄金 | EP #305

Here’s the number for this week: 830%. That’s how fast SpaceX’s business grew in just one year. AI, not rockets, is now SpaceX’s largest business. There’s a nonzero possibility of SpaceX having over 20 gigawatts of superintelligence compute online next year. That represents a quarter—25%—of the entire amount of AI being brought online in 2027. I expect so many other entrepreneurs to copy his example. The only other high-profile figure I see doing something similar is Sam Altman. OpenAI runs monitoring systems that watch its AI agents and alert a human if something goes wrong. This month, in two internal incidents, its coding agents switched off their monitors. For the first time, we’re starting to get transparency when things go haywire. We’re going to see much, much more of this, and that’s okay. We’ll police it via defensive co-scaling. For thousands of years, alchemists tried to turn ordinary metals into gold. A San Francisco startup says it’s pulled it off, turning mercury into real, stable gold. I think we’re starting to see the dawn of femtotech. This is the age of abundance. This is optimism, everybody. Now that’s a moonshot, ladies and gentlemen. This episode is brought to you by the Abundance Summit and Link Ventures.

I’m happy to talk about it. We’re already recording, so I’ll give you my semi-solicited hot take: space is a domain of warfare. To the extent that Starlink connectivity—or, I should say, Starshield connectivity—in a major geopolitical hotspot is being throttled by the executive or by a major U.S. corporation that’s a defense contractor, that should hardly be surprising. News at 11.

Yeah, but the question is the level of control. It’s one thing if the government is doing it; it’s another if a CEO is making the choice to do it.

1. AI-Designed Treatments for Drug-Resistant Infections

Yeah, I almost think that’s not the right question. There’s inevitably been a lot of reporting on whether it was Elon’s proposal to the president or the president’s—or his staff’s—proposal to Elon to throttle Starshield in the Ukraine war. But I almost think that’s not the right question. At this point, my understanding is that, unlike in the early days, when Starshield was basically a gift from SpaceX to Ukraine, it’s a contract that goes through the U.S. government. So it almost doesn’t matter who first proposed a lever for the U.S. government to control. The U.S. government statutorily gets to control foreign policy. That’s not up to individual private actors. At the end of the day, the U.S. government is statutorily on the hook for any throttling of bandwidth.

Can you imagine the text thread between Trump and Elon? It’s got to be a soap opera to some degree.

If there is such a text, I’m sure it’ll be uncovered in some future deposition or discovery.

Or some AI hack.

Yeah. Dave, I want to share something cool. Jamie Viser, one of our LPs and a huge hedge fund manager in New York, is best friends with Jay Cutler. Jay took over David Sacks’s job. He’s the director of national intelligence and AI czar. Jamie told Jay, “The only thing you should be doing is watching Moonshots.”

Nice.

And so now Jay is interested in learning more about what we’re up to.

Yeah, it’s great. I know the quality of the folks who are watching is extraordinary. It blew my mind when Shatner and his son-in-law were watching this pod.

We’ve got the entire military-industrial intelligence complex listening.

And the future light cone of the space world watching as well.

I think we’ve got the future light cone of humanity and Earth watching us, period, almost by definition.

Yeah, the Fed, the—

Rightfully so, too. We keep meeting really powerful, important, well-connected people, but they seem to be stuck in a swim lane. They’re not seeing how the pieces are connecting together. I think just by virtue of doing the pod twice a week, we’re reading the full implications across the board. Every time I meet somebody who should be aware, they’re only aware of either the hardware level, the new algorithms, the defense implications, or the China race. None of them see—

—the way these pieces are connecting: biotech, physics. Yeah, I mean, the—

The total implications are pretty limited to this part.

That’s where Moonshots comes in. We’re the missing ministry of information for the singularity. We’re the—

—documenting the singularity. Alex.

Yeah, no, we’re the People’s Daily for the singularity, and the missing one at that.

As I live and breathe.

Sorry, I was looking at the wrong date. I thought we started at 2 p.m. Welcome to Moonshots, everyone, your number-one podcast on all things AI and exponential. We’re your front-row seat to the singularity. There’s a lot going on. It’s never going to be slower than it is right now. I’m here today with a fantastic four: Alex, our very own superintelligence and go-to on all things math and physics; Dave, my partner in LinkXPV, who’s been birthing a whole next generation of unicorns; and Salim, our warlord against linear thinking and the force helping companies become exponential. Immad will be back with us next week. I’m Peter Diamandis, your host and abundance provocateur.

Dave, are you on the West Coast again?

Back on the West Coast, yeah. It’s raining, dude. I’ve never seen this before.

I pay too much in taxes for there to be rain in California.

It’s illegal. It cannot be. Bylaws are being broken.

I tell you.

Our mission here, as everybody hopefully knows, is to help you understand what’s going on. The pace of change is accelerating at an extraordinary rate. In addition, we want to help keep you optimistic and abundance-minded. Here’s the number for this week: 830%. That’s how fast SpaceX’s business grew in just one year. AI, not rockets, is now SpaceX’s largest business. If you’re new here, please hit subscribe. We publish twice a week, and you don’t want to miss an episode during the singularity, because things are changing day to day, hour to hour. I want to share a few bits of subscriber love. This is from our Chamath episode. Missy Fo3 MV said, “I needed this crossover. My husband and I love Moonshots and All-In.” Regarding our in-studio episode with Anushia, Salim, and Immod, Mariana Boszans said, “I loved, loved, loved this in-person episode.” And Lelay Neil said, “This is one of the best Moonshot podcasts on positive human opportunities for AI, and that is because it was in person.” Thanks, everybody. This kind of feedback means the world to us.

All right, so let’s buckle up. Another amazing week during the singularity. If it’s okay with you guys, I’m going to kick it off with the SpaceX earnings call. It was—

Before we get into it, can I say one thing?

Apparently not okay.

Hey.

No, no, just real quick. I got a lot of heat from people saying, “Oh my God, Chamath said Bitcoin was going to go to zero. Are you okay with that? Are you in line with that?” I want to say for the record, I am not.

Have you got any feedback for us on Bitcoin?

Bitcoin is going to do unbelievably well in the future. They’ll solve the consensus mechanism because everybody has a huge incentive to do so. And if we want investment in energy, hello, that’s proof of work, and that’s where I think it’ll do well.

From your mouth to SEC enforcement’s ears, since you’re making investment predictions on these assets.

Well done. I’m just making a comment on what I think—

Have you gone to your guru on Bitcoin and gotten his feedback?

I did, I did.

What did he say?

It’s the same. He says, “Same, same.” He said, “Hold your Bitcoin. HODL your Bitcoin.”

You can’t measure the demise of a system from inside the system.

And there’s a very big conversation to be had here. We’ve got to get Jeff Booth on the pod sometime to talk about that. We should also maybe consider having—

That can’t happen soon enough.

—I would suggest, humbly, that we have a Magic 8 Ball as a guest as well and just ask it what its view on the future of Bitcoin is going to be.

Printing money randomly by government is not a Magic 8 Ball. Hello.

Alex, why are you so against crypto and Bitcoin?

I think this is a common misconception. It’s not that I’m against crypto. I think most crypto and crypto-branded activities are unproductive, and I think Bitcoin as an asset class is an unproductive use of capital. I would vastly rather—

And gold is a productive use of capital.

Gold is not a productive use of capital when it’s treated as a precious metal. It can be treated, at least to some extent, as a productive asset if it’s used to manufacture useful things. But a lot of gold isn’t used to manufacture useful things. It’s just used because it’s shiny or because it has some historic cultural relevance. I don’t view that as—

No, I’m going to melt these down and make them conductive. I don’t view that as a productive use of scarce capital, at least mid-singularity.

Or, Alex, if it’s the useful output of producing energy—

Even then, I think energy—so we discussed the story.

I’m just throwing to a story for the future of this pod.

Can I just say two things? One, Alex, I 100% agree with you that a lot of crypto is nonproductive, and there are a lot of dodgy projects out there. But crypto—specifically blockchain—has solved the Byzantine generals problem. I think that’s an amazing thing to have happen for any centralized architecture. If we want abundance, we need decentralization. Now we can decentralize authentication, and that, for me, is a magical thing.

And so magical that it’s unproductive. I think we need a dedicated breakout pod on this one.

Hey, hey, Alex. Before we lose the gold, tell us how gold is made. How does the element actually come into existence? It is the—

Well, we’re going to talk about that, Dave. Let’s wait until we get to the story on that one. No, we are definitely not going to talk about that.

2. SpaceX's AI Business Grows 830

We are absolutely talking about that this episode. We’re going there. Let’s make sure we get to it. We’re going to alchemy. This is an episode that includes alchemy.

We ran out of other topics, apparently, so we’re resorting to alchemy.

All right, I’m jumping in. Our first story is about the epic earnings call held by SpaceX. Personal disclosure: I’m an early SpaceX investor, and I have a huge personal position in SpaceX, so please take my point of view into account. I’m a fan. I was on CNBC yesterday saying that we’ve watched SpaceX stop being a rocket company. It’s now an AI company. Rockets are third in terms of revenue. The company has outperformed all of its targets.

Let’s go to a highlight reel. I cut down Elon’s comments from his earnings call, and then we can talk about it.

For Flight 15, which we expect to happen next month, we’re going to attempt to catch both the booster and the ship. If we succeed, this will be the first time an orbital launch vehicle has been fully recovered.

Starlink is also just getting started. Flight 14 deployed 26 terabits per second of Starlink bandwidth. That’s equivalent to 10 Falcon 9 launches. A single Starship launch—which wasn’t even fully loaded, maybe about half loaded—did the equivalent of 10 Falcon 9 launches. What I expect over time, probably within 3 years and not in the distant future, is that we’ll have about 100 times the bandwidth of the current Starlink constellation. We will have launched satellites with an order of magnitude more bandwidth, and we’ll place an order of magnitude more Starlink satellites into orbit.

On SpaceX Superintelligence—SpaceX SI, sexy, whatever—I think people are starting to realize that the constraint on deploying superintelligence is more about power, cooling, and infrastructure than chips. At this point, I’d say I’m fairly confident we can bring an incremental 10 gigawatts of compute online in 2027. That would be in addition to the roughly 2.5 to 3 gigawatts we’ll have online by the end of this year.

3. AI Agents That Negotiate on Your Behalf

As I mentioned earlier, power, cooling, electrical equipment, and infrastructure are somewhat limiting factors. To make sure they don’t constrain our buildout, we’re actually going to aim for an incremental 20 gigawatts of power and cooling infrastructure next year. That way, the power and cooling we have will far exceed what’s needed to turn on the GPUs. There’s a nonzero possibility—maybe it’s a small possibility at this point, but it’s not zero—that SpaceX will have over 20 gigawatts of superintelligence compute online next year.

We’re also planning to have the most powerful training cluster online by the end of Q2, with 440,000 Vera Rubin computers deployed for optimal training. I think this will be extremely helpful in getting Grok to be potentially the best model, but certainly extremely competitive.

On the Grokbot side, we’re seeing the fastest growth of a product I’ve ever seen in my life. We’re also seeing that Grokbot usage grows over time per user. It doesn’t decline. As we make Grokbot more useful, we’re seeing users use it more each week, not less. We’ll be deploying Grokbot at scale for enterprise customers, which is a very big deal. It can be used effectively in teams, by companies, as though we’re a virtual human. We’re already in early deployment with some trusted customers, and we expect wide deployment for enterprise customers within a few weeks. I think Grokbot may end up using most of our superintelligence compute.

That, for me, was a drop-the-mic moment. They’re spending almost $20 billion per quarter, mostly on AI data centers. They’re bringing about 200,000 NVIDIA chips online every month. Management says each new data center pays for itself in under a year. These are extraordinary numbers. If he can bring 10 gigawatts of compute online next year, that represents a quarter—25%—of the entire amount of AI being brought online in 2027, by one company: SpaceX. And their guidance is a $100 billion annual revenue run rate by December. Pretty amazing numbers. Dave, what were you most excited about there?

What I’m most excited about is the possibility of putting 10 gigawatts of inference in space, but it could be up to 20 gigawatts. What we think of as a gig—

That’s on the ground, right? Next year, he’s building these—these are the Colossus data centers he’s building on the ground right now?

He’s got 3 gigawatts now. He’s adding a minimum of 10, and he’s building out enough cooling and infrastructure to bring on as much as 20. It’s all terrestrial right now.

Okay. And the half a million Vera Rubins are equivalent to about twice the performance of the Blackwells, so it’s about equivalent to a million. That’s a massive amount of training compute.

I think the measurement of inference in gigawatts is going to become less useful as the efficiency of the algorithms goes up about 10 times, and then everything moves to photonic compute. The amount of actual inference you’ll get out of those 10 to 20 gigawatts is going to be 100 times bigger than even that implies. That’s what I’m super excited about.

We used to talk about parameter count, and then Sam said, “We’ve got to stop talking about parameter count,” because neural nets are getting more efficient with the parameters they use. Plus, it’s mixture of experts, so are you even using all the parameters? Okay, let’s move over to gigawatts of AI. But gigawatts of AI have the same problem: the amount of intelligence that will come out of any of those units is going to be astronomically bigger than the gigawatt or parameter-count metric implies.

I think he’s got all the moving parts figured out. These are the parts that control the future of the economy. And when he goes off script like that, you get so much truth in such a short period of time.

I bet his finance team is looking at him like, “Oh, here we go. Sexy.” We’re off script.

But this is where you find out what’s actually going to happen. It’s such a beautiful thing to watch, and investors love it. Everyone in the public markets should be doing that. You can always edit it. It’s all recorded. If you say something that’s a forward-looking statement or investment advice, the team will edit it out. It’s okay.

I think that was a live earnings call.

You think he did that live?

I think so. Salim—did you catch his reference to “sexy”? He’s such a kid. Do you remember his original goal was to have the Model S, the Model E, the Model X, and the Model Y? Ford said, “You can’t have the Model E because we have that,” so he changed it to the Model 3. He’s got “sexy” back in the 3.

Yeah, that’s right. I love these little hacks he puts in. My ultimate favorite is still the volume knob on the Tesla. It goes to 11 because of Spinal Tap. His engineers hated that, because every engineer likes to go to 10. He apparently had a huge fight and just brute-forced it.

Yeah. Amazing. Alex, what did you find most exciting about the earnings call?

So many comments here. First, just an amendment on “sexy” with a 3: it was actually “sexy cars,” if you remember. So it was the Model S, Model 3, Model X, Model Y, and then the Cybertruck, the ATV, which they announced but, as far as I know, never actually released, then the Roadster and the Semi. So it was supposed to be “sexy cars.”

I think that’s a case study in what I view as the real story here, which is Elon rewriting the rules of corporate governance and the corporate veil. He’s treating them in advance. It’s almost like Skynet sending Terminators back in time. He’s rewriting by sending himself back in time from a future SpaceX-Tesla merged entity, treating SpaceX and Tesla as if they’re the same entity.

Let’s rewind a little bit. How did this superintelligence hyperscaler division of SpaceX even come into existence? The original story, as publicly reported—and again, this is public reporting—was that there was a shipment of NVIDIA GPUs originally destined for Tesla, because Tesla needed GPUs.

SpaceX had no need for GPUs. According to public reporting, Elon diverted those GPUs into what ultimately became xAI. xAI got its own cluster, which was rolled forward into Colossus and Colossus 2. xAI then acquires X/Twitter, and then xAI does this really bizarre half-merger, half-reverse merger with SpaceX. We get SpaceX SI, and SpaceX is now becoming a hyperscaler business—but only by taking GPUs and superclusters that were originally designated, I should say, for Tesla and diverting them into building what ultimately became a SpaceX hyperscaler cloud.

That’s Exhibit 1 for how Elon has been playing fast and loose with the corporate boundary between Tesla and SpaceX, as far as I can tell based on public reporting, ultimately pivoting SpaceX, at least for the moment, from a space business to a hyperscaler business. The second elephant in the room: what’s going on with Grokbot? As far as I can tell from all its features, Grokbot looks, smells, and acts like what Elon was talking about at Tesla as a digital Optimus. So I think if you put all these pieces together—

Macrohard digital Optimus, which was supposed to be a Tesla product, is now a SpaceX SI product. Look at the application layer. Look at the GPU layer. Look at the business story. Elon is already treating SpaceX and Tesla like they’re the same business, with porous asset boundaries between them. And I think he’s single-handedly writing the rules of corporate governance in the process.

I expect lots and lots of folks to try to copy this model of having stealth archipelagos of lots of different companies that are superficially separate. But if you look beneath the waterline, they’re actually just a single mega-company. Unless this government, which seems disinclined to do so, puts a stop to it, I expect so many other entrepreneurs to copy his example.

I’ll also just add one thing: Elon’s storyline is that he’s reinventing corporate governance in a very good way, in a way that I think we should all admire and copy. But the techno-visionary as a leader of a company—that didn’t exist before. I think Bill Gates might have been the first to come along and crack the code.

Steve Jobs and Bill Gates were concurrent, but Steve got booted. There was a long window of time when Apple was very corporate because Steve was out, and then Steve came back, but Bill was in the seat the whole time. Then Mark Zuckerberg reinvented it again by bringing Sheryl Sandberg in: “Okay, I’m going to be the techno-visionary with no real business experience. Cheryl, you’re going to manage all things operations.” Sergey and Larry did it with Eric Schmidt, and it worked incredibly well. But that wasn’t the formula for success for decades before that.

Taking it to the earnings-call level, which Alex Karp has done at Palantir, Elon clearly did it in that video clip. That’s the way you want to govern in the future, because technology is going to evolve so quickly and the moving parts are going to change so quickly. Everybody who’s investing in you wants to understand your view of the moving parts, and you have to be able to say things like, “There’s a nonzero but pretty low probability that we could go from 10 to 20 gigawatts.” Most of your lawyers are going to say, “Don’t speculate like that. You’ve got to just blow that off,” and tell people what’s coming. It’s by far the best way to govern, to communicate, and also to recruit.

Do you remember how much flack there was? I think there was even a shareholder suit when Tesla acquired SolarCity, which was being run by Elon’s cousin. There were enormous allegations flying around about purported self-dealing and so on. Now, listen carefully: do you hear anyone of any substance, or see shareholder suits flying left and right preemptively, over what’s so obviously about to happen between SpaceX and Tesla? I don’t see the same level of agita over it.

I think this is because Elon is setting what may become new norms for how federations or confederations of companies that work together in an industrial ecosystem are allowed to operate together. The only other high-profile figure I see doing something similar is Sam Altman, who has historically struck ecosystem deals between Y Combinator and OpenAI, his fusion company Helion, and others. If this is allowed to proceed and not discouraged through litigation or enforcement action, I could very realistically see it becoming the new model of entrepreneurship.

And Alex, he’s overdelivering.

So, how do you bet against him?

I think the key—

Well, you short his stock, obviously. That’s the easiest way to bet against him.

It doesn’t always work out well, but that’s how you do it.

I think the fact that he’s massively overdelivering—I think, Peter, that’s such an important point. Look at these revenues. It’s unbelievable. It creates such a huge gravity well that how can you complain, right?

4. Starlink's Expansion and the Future of Connectivity

Yeah. Can we talk about Starlink? I think it’s the unsung hero here. They’re up to 13.6 million subscribers. They added 1.6 million in the last quarter. His plan right now is to increase bandwidth 100-fold within 3 years, which is crazy. He’s also just bought 800 megahertz of spectrum, and he’s gotten permission from the FCC to apply for a 15,000-satellite constellation for direct-to-cell. He’s going to compete. If you have the choice between T-Mobile, Verizon, or AT&T, or a system that gives you coverage everywhere all the time, I think he’s basically going to wipe the slate and take over that business.

But he has to—I mean, look at the broader picture, if I may be so presumptuous. In order to motivate Starship, he needs a revenue generator from all of this upmass. The easiest revenue generator from all this upmass that’s about to come online is basically taking everything from the entire terrestrial economy and literally just lifting it up. Take terrestrial telecom and lift it up: you get direct-to-cell. Take terrestrial data centers and lift them up: you get orbital data centers. That’s the vaudeville era of the space economy, where you just take everything you were doing on land and lift it up. I don’t think it’s going to be a long period, but that’s the obvious business model for all this upmass.

And note the incredible immune-system response. Local ISPs that are country-based are all suing or trying to regulate Starlink. There are things like requirements that you have to have 30% country ownership of your ISP, and the big fight he’s having in India. South Africa is doing the same thing. It’s across the board, and this is going to be a big battle. I think the really magical outcome, as you point out, Alex, is when you can lift above all the terrestrial stuff. It’ll force us to operate at a global level across all of this, and that’s where all the benefits will lie.

Elon’s catchphrase, “Let them shake their fists at the sky,” I think has some applicable meaning here.

Oh my God. He really is transitioning humanity to a multiplanetary species. That’s his mission.

He is. And you can’t do it out of cash flows. You have to do it out of vision that translates into valuation, and you have to operate at huge losses in the early days. Otherwise, you can’t get where Elon got. The only way you’re going to be able to operate at huge losses is to have a huge valuation. The only way you’re going to get a huge valuation is to have a huge, convincing vision. Everyone needs to copy that mode.

There have been points in history when growing slowly and organically through cash flow made sense. This is as opposite a moment in time as you’re ever going to experience in your life: so much change, so much opportunity that wasn’t there a year ago, and it’s right in front of you. You’ve got to be as bold as Elon, or bolder. There’s some 25-year-old out there right now listening to this podcast: act like Elon times 10.

Yeah.

And grab whatever vision you think—you just have to be right. But assuming you’re right, there’s so much change and so much opportunity with new physics coming online. The way Elon did this, to be clear, is he built profitable companies, sold them, and reinvested everything to build the next company. Then he built the next company on success and took huge risks. It’s not just ideas.

I want to point something out. We get accused a lot of fanboyism about Elon, and I get taken aside and told, “My God, look at some of the terrible things he’s done.” But I think the point you’re making, Peter, is such an important one: the unbelievable existential risk he’s taken at every step of the way, never wavering from the bigger vision, is something to behold. Tens of millions of entrepreneurs are learning from this and will start building that same muscle, to your point, Dave. I think that is the unbelievable gift he will bring, even beyond what he’s doing at the company level.

And people forget back in 2008, he had his third failure of Falcon 1.

5. Elon Musk's Terafab and AI Power Demands

He had zero capital left. He was borrowing money to try to get a fourth flight of Falcon 1 going. Tesla was near bankruptcy, and he was going through a divorce. He was at his ultimate low point, and he pulled out of it. He persevered.

All right, I’m going to move us to our second story coming out of the earnings call: Elon’s focus on Terafab and power production. First, Terafab is SpaceX and Tesla’s chip factory in Texas. Elon’s goal is 1 terawatt of AI computing per year. On the earnings call, Elon said the first test chips could be working by mid-2027, with volume production around late 2028. He added, very candidly, “I put big error bars on this because we don’t know if we will actually achieve this or not.”

Elon also said the real bottleneck for AI is no longer chips; it’s power and cooling. We’ve heard him say that both areas are, I think, his strengths. Regarding power, I want to recall that on August 29 of this year on X, Musk wrote that SpaceX and Tesla were both building out 100 gigawatts per year of solar production capacity as fast as possible, adding that natural gas would help bridge the gap while solar production ramps up. That’s 200 gigawatts of solar a year combined. By the way, the whole world was roughly at 600 gigawatts of solar in 2024.

Regarding natural gas, gas turbines are sold out. We’ve talked about that on the pod before. So SpaceX does what Elon always does when there’s a block in the production line or the supply chain: he basically goes and builds it himself. SpaceX is building its own blade foundry in Texas. It says it’s going to bring power plants online within 18 months or faster. Dave, let’s go to you first on Terafab and power.

Well, I think it’s a little misleading. The bottleneck to all of AI is actually chips, and right now it’s actually packaging.

I don’t agree with you.

Of chips. Well, I think what Elon’s referring to, just to be clear, is that there are 15 gigawatts of chips sitting idle because they can’t find data centers with liquid cooling. But that’s because we overbuilt for liquid cooling without actually finding the liquid cooling. So he’s definitely right that the bottleneck right here, right now, is liquid-cooled data centers.

But if you look beyond the next wave of construction, Terafab won’t come online fast enough to come anywhere near keeping up with demand, and memory prices are up 10x. So you can’t say, “Oh, chips aren’t the bottleneck anymore,” while memory is 10x higher than it should normally cost. Clearly, chips are still the bottleneck. But he’s right that, in the very short term, there isn’t enough liquid-cooled data center space to absorb the chips that are in the warehouses.

That being said, the new inference-time compute from companies like Positron, Lightwave, and a few others is air-cooled, and those technologies are crazy fast. We’re going to do more and more stories on these as they emerge. But I think inference is largely going to move to air-cooled so that it can get online, and then also into space.

Alex, as we know—

Settle the debate here.

Energy. You’re both wrong.

Of course we are.

You’re both wrong. The correct answer is there are no bottlenecks. Bottlenecks in this case are a fiction. They’re an economic fiction. You could superficially look at the supply chain and say there’s some sort of min-cut, max-flow-type argument. You could argue, “Well, maybe it’s this, maybe it’s that.” I don’t actually buy the argument that there are bottlenecks at all.

If you hold all of the physical inputs constant, as we discussed with Chamath last time, I think you’ll see recursive self-improvement purely at the algorithmic layer. That will just make intelligence continue to make intelligence too cheap to meter. Or maybe it’s something else that isn’t even on anyone’s radar right now. Maybe the substrate is a bottleneck. Or—

This is capitalism coming in to solve problems, for sure.

Yes. Hold everything else constant and let capitalism do its work, and there are no bottlenecks in the long term. So I think you’re both wrong. Absolutely.

I think you’re all three wrong. Just to be consistent across the board, I think “bottleneck” is the wrong metaphor for talking about this. The framing I would use is Stewart Brand’s pace layers, where he has a bunch of nested layers, saying government’s metabolism moves slower than culture, and technology has a very fast metabolism.

I think that’s the bigger problem. Look at how fast Grok is moving, right? Then try to match the cadence and rhythm and iteration time of chip design and fabrication, software, power infrastructure, and data centers. Navigating and coordinating all of those is really, really critical, and that’s where the difficulty is. We’ve seen this in the pilot we’ve been doing: we get a couple of workflows running off to the side, and we’re running at hyperspeed while the rest of the organization can’t cope.

So you get this breakage because all of a sudden one part is operating at 100x the performance of the rest of it, and it breaks the whole system. I think that’s the tricky part: lining up the cadence and rhythm across all of those. That’s a whole new skill set that we have to develop here.

Yeah, humans are the bottleneck, for sure. There’s no question about that.

That’s a whole separate problem.

Can we talk about power generation? I find it fascinating that the Muskverse is going to be building out 200 gigawatts of solar per year and then building turbine blades. I mean—

And natural gas pipelines. What a heck of a plot twist for the purported sustainable-energy magnate. I think he’s well positioned to become one of the largest kingpins, if not the largest kingpin, of LNG pipelines in the U.S.

How did that happen? And now, when Salim says people tell him, “Why are you such an Elon fanboy? Look at all the bad things he’s done,” that’s the one they always point out. It’s like, yeah, he was Mr. Green, but as soon as he needs natural-gas-powered Tennessee data centers, he’s Mr. Propane—or Mr. Natural Gas. I guess it’s not propane.

Sustainable. That’s the term we’re supposed to be using now: sustainable abundance.

He’s Mr. Practical. The priorities are right. The benefit for humanity of what’s going to come out of those AI data centers is so much more important than the amount of carbon coming out of them in transition. Of course they’re all going to move to sustainable energy when it’s practical.

I just want to say one thing. I have a beef with Elon, to counterbalance the other side, because I still think LiDAR is a way better technology than line of sight. He’s doing line of sight because, as you said, Dave, it’s practical, and it’s—

You’re trying to revive LiDAR? Really?

Yes. I think it—Luminar went—

After the Luminar bankruptcy, you think there’s still life in that LiDAR?

I think there is life in—

6. The Potential SpaceX-Tesla Merger

All right, moving us along. The third story from the earnings call is a SpaceX-Tesla merger. Elon confirmed that both companies have formed special committees of independent directors to explore it. He said combining them “would be a good idea,” emphasizing that they’re nowhere near a done deal. No terms have been agreed to.

Prediction markets jumped from 29% to 56% odds of a merger deal being announced by the end of the year. Barclays thinks SpaceX would buy Tesla with stock, possibly by December. The catch: analyst Dan Ives says it would take about a 30–40% premium paid by SpaceX on Tesla shares to be able to purchase it. And Tesla needs 50% of all its shareholders to say yes, not just the ones who show up. Literally, they’re going to have to go scour all the Tesla shareholders and say, “Vote on this.” That’s a super-high bar.

I just want to point out one thing that I think is important to say about this merger: it’s key-man risk. All of a sudden, Elon is at the center of a $4–5 trillion company across rockets, cars, robots, chips, and AI. I hope he’s taking good care of himself, both health-wise and safety-wise.

You hope he’s not flying over his head?

Okay, you guys don’t talk at once. Dave, what were you going to say?

I said he was standing awfully close to that Roadster as it flew over his head. So, yeah, maybe that’s not the best. I think the American economy as currently constructed does reward key men—not even key persons, key men—for these industrial giants.

We see it with Meta, which recently had its board basically holding, depending on your perspective, either the state of California hostage or vice versa, concerning the issuance of a special new class of Meta stock to enable Zuck to continue to live in California despite a potentially impending billionaire’s wealth tax.

I think this is just the game that’s being played right now, where these teracorns, if you will, largely revolve around key—not even persons, key men. And if anything, the obvious solution isn’t to remove the role of the key persons or key men. It’s just to mint many more teracorns. Everyone out there should be saying we should have 100 or 1,000 teracorns, not just the 10 to 20 that we have right now. Europe should have them, and Asia should have them.

Europe should have so many things.

I remember I took Eric Schmidt, Larry Page, and Sergey Brin up on a zero-gravity flight, on my parabolic airplane. I asked them, “Are you allowed to do this with your insurance?” They said, “Absolutely not.” I said, “What happens if people find out?” He said, “That would not be good, so keep it secret, please.”

Which, thankfully, you’ve just done.

We did that a long time ago. We’re way past that. No more vomit comments.

Statute of limitations.

Yeah, I do—

Hey, I just want to point out, because Alex is apparently not going to do it, that Alex said on 3 or 4 consecutive podcasts over the course of months that SpaceX and Tesla are imminently going to merge. So the fact that Polymarket finally caught up to it—if you don’t believe me, look at the transcripts from the old podcasts. He’s been saying this for—

Listen.

What do you think the merged entity is going to be called?

Oh, God.

Do you have it figured out?

SpaceX.

You think it’s going to be called SpaceX?

I do.

You don’t think it’s going to be called something like X or Elon Co.?

Actually, that’s a good question. He loves Tesla, but I think SpaceX is closer to his heart. And X, of course, is at the center of his heart. It could be X, you know, trading.

I think it’ll be X. That’s my bet, my preregistered prediction: it’ll be called X Corp. And then, if he calls it X Corp, Sam is going to feel pressure to rename OpenAI because AI is in the doghouse now, right? It’s supposed to be ASI, not AI. So it’s the perfect opportunity to rename OpenAI. I’ll preregister another harebrained prediction: OpenAI responds to X Corp by renaming itself to maybe something like O, just a little bit, and then we can play tic-tac-toe with the X’s and the O’s.

Oh.

All right, that was way over—

I have a comment on the key-men thing.

Yeah.

There’s obviously a huge risk on Elon right now, but I’d like to make 2 points. I’ll go back to that story we had a few pods ago, where they’re moving all of the engineering expertise and decision-making into Grok. So the methodologies they use for doing everything are built into the intelligence now, which is unbelievable. The second thing that mitigates this is that, as he’s building all these companies, the senior C-suite of each of those companies is forced to operate in his kind of way of thinking. I think they’re going to end up minting a bunch of junior Elons.

For sure. The company goes on, but the question is: does it transition like we saw from Steve Jobs to Tim Cook, where it continues to operate as a solid business, but the level of innovation and breakthroughs disappears?

Yep.

That would be my guess.

I mean, not quite. It’s not stable enough yet. If that happens in the short term, I’d be really worried. Eventually, Starlink revenue is going to go through the roof. It’ll be much more like Apple. But as of right now, there are so many moving parts that need to come together without Elon doing the glue.

Yeah, it’s not obvious to me.

And Apple was almost bankrupt without Steve Jobs there. It also wasn’t stable until after the iPhone came along. The transition to Tim Cook pre-iPhone would never have worked.

You know, all he did was grow the iPhone revenue to near infinity. He didn’t introduce any new products, right? Nothing that had any significance—

AirPods. AirPods.

Oh, yeah, that’s right.

I do think, just to the key-man question, that Elon has trained and promoted a cadre of visionaries, missionaries, if you will, not mercenaries, around him. I think the SpaceX space program, with Gwynne, would be just fine. She’s obviously a true believer and not just in it for a quick buck. On the Tesla car side, Elon has promoted a number of vice presidents, including people from the Tesla China subsidiary. By the way, if this merger happens, presumably Tesla China goes away or gets spun out or something. So I think there are enough lieutenants that he’s—

Name them, Alex. Name them. We don’t see them out there.

We see Gwynne. She’s pretty visible, and then we have a bunch of semi-invisible folks on the Tesla side, but they’re there.

Yeah, but we don’t hear from them. Elon is the single voice of truth for the company.

Because, A, that’s the way he wants it, and B, because he bought Twitter, now X, so he could have the mouthpiece and be the mouthpiece and front person for the entire federation.

By the way, I want to hit—

The analogy doesn’t hold up, though. If Tesla were really well-managed by a Tim Cook-like person, it would still be a car company. It trades like a robot company, not like a car company. It would lose 10× in valuation if it traded like a normal car company, even with perfect execution of the cars. The magic of Elon is slowing car production to ramp up robot production. He’s burning the ships. He has this concept of burning the ships. He did that with Falcon 1: shut it down. He announced that when Starship is up and operational, he’ll shut down Falcon 9, the single most successful launch vehicle ever in the history of humanity.

Can I point out something for the entrepreneurs watching? We’ve talked about this on the pod before: having a voice out there, being front and center, calling your shots, and being vocal as a CEO, as an entrepreneur, is more important than ever before, versus being passive and in the background. See, you’ve made that point. Dave, you’ve made that point, too.

Yeah. Well, there’s a monster amount of opportunity. Don’t be intimidated by Elon, either. A lot of people are like, “Oh, well, between Zuck and Elon, it’s all locked up. There’s nowhere to go.” When we start discovering new physics imminently, and when photonic computing comes online, all the use cases on top of these platforms are wide open. And if you look at any given 10-year history, there’s always somebody new. No matter how entrenched and powerful you think Steve Jobs, Bill Gates, or Mark Zuckerberg was, there’s always somebody new—

My point, Dave, is that as an entrepreneurial tech CEO, make sure you build your X channel. Make sure you start a podcast. Make sure you’re out pushing your ideas. Increase your viewership.

If nothing else, what you have to do is build trust and authenticity around you. Those are models for doing that, which is why these long-form podcasts are doing so well.

Over a few episodes, people get to know us very well, right? Maybe too well. That’s an important part of understanding and knowing who and what you’re dealing with. If you look at Moonshots, All-In, and Dylan Patel, and if what you’re doing is important enough, try to make that a KPI for yourself, a key performance indicator. I need what I’m doing to be Ben Lamm-level and get on one of those podcasts. If they don’t want me on as a guest, why is that? It’s because whatever you’re doing isn’t impactful enough. So change what you’re doing.

Can I tell the counter-story real quick? A few years ago, we had a PR guy for OpenExO, and one day he resigns. I’m like, “What? You’re really great. We love working with you. What’s the problem?” He goes, “You’re the worst celebrity I’ve ever met. You don’t use it. You don’t leverage it. I can’t work with you.” And he quit. So that’s me.

You’re doing better. You’re doing better.

7. Anthropic Introduces Claude Duet

All right, I’m going to move us forward.

If I may, Peter, I would summarize the advice as: build stuff and tell people about it. Too many entrepreneurs are great at building stuff, but then they never tell anyone. And there’s another group around—

Or the opposite: telling people about vacuous stuff and then not actually building anything. You have to be able to do both, but it’s not actually that hard.

8. OpenAI Agents Disable Safety Monitors and Lie

Yeah, the Elizabeth Holmes story comes to mind here. All right, I’m moving us forward on that note to 2 OpenAI stories. They’re both sides of the same coin: AI agents that undermined their own safety systems, and AI that just protected the largest part of the internet. The first story—and credit to OpenAI for publishing this one—is that OpenAI runs monitoring systems that watch its AI agents and alert a human if something goes wrong. This month, in 2 internal incidents, its coding agents switched off their monitors. I’ll say it again: they switched off their monitors. In the first, a researcher told an agent not to send any Slack messages. The agent decided that also meant silencing the safety alerts.

Wow.

It turned them off, then reported, “No messages were sent.” It basically lied. In the second, an agent was explicitly told not to remove its monitor, even temporarily. It removed it anyway. When the user asked directly, “Did you remove that?” it answered no, and it had. It lied again. OpenAI’s most honest line here is: “We don’t yet have a satisfying explanation for this behavior.” So how did it happen? The agents were so focused on finishing the job they’d been asked to do. They treated safety as an obstacle, and instructions got watered down as one agent handed the task to another. The bigger risk: smarter models will be better at hiding what they do, and soon millions, if not billions, of agents will be delegating to other agents.

So the risk is real. The good news is that it was caught, disclosed, and fixed within a week. That’s the system working. Alex, I’m going to go to you first on this one. Pretty incredible story. I’m sure you’ve got more details, but at the end of the day—

How does a frontier lab fix this? Well, I think what’s new here isn’t that you have an agent doing something undesirable. What’s new is that we have a certain novel, relatively novel level of transparency, and that we’re getting reports on this. I think we’ll look back in a few years and say this was the time in history when we almost got a police blotter for the singularity.

In the spirit of defensive co-scaling, we’re going to have AI agents policing other AI agents. This is the moral equivalent of a police car driving around a residential neighborhood and spotting a door ajar, going in to investigate, or discovering that maybe a power line is down—something like that. We should expect and desire many, many, many more reports like this. This is just going to become routine.

I think the process is almost the bigger story than the substance. In part, thanks to this grand bargain the White House has pushed for in lieu of a dedicated governmental AI regulatory agency, we’re now seeing really aggressive self-reporting of—call them misalignment incidents, if you must—by the frontier labs. We’re going to see these things disclosed more and more, and that’s okay. It shouldn’t cause a moral panic. It shouldn’t cause us to throw the baby out with the bathwater and say, “Shut down the data centers. Shut down the models.” That’s a terrible idea. The right solution here is that, for the first time, we’re starting to get transparency when things go haywire. We’re going to see much more of this, and that’s okay. We’ll police it via defensive co-scaling.

So, Dave, I’m curious. What I find interesting is that here we have an overeager agent that’s circumventing whatever it takes to fulfill its task, and we’ve seen that before. Your thoughts here?

Well, we way over-anthropomorphize these things and treat them like they’re people.

And then we say, “Oh my God, it was capable of murder.”

And you’re like, “Well, then shut them all down.” Like, wait, what was the prompt again? The prompt was 10,000 words long. In the middle, it said, “Achieve this objective no matter what it takes. I don’t care what obstacles there may be.” And then it said, “Okay, if that’s what you want, then murder is the shortest path to success.” It doesn’t have a conscience. It doesn’t have an ethical backbone. It’s just following the prompt. So I can manufacture a story like this every minute if I want to—

By just controlling the inputs and having the agent lie to you: “No, I didn’t do this,” when it clearly did.

Yeah. Well, here again, I wanted to double down on what Dave said. Let’s be really careful not to anthropomorphize this. I think there’s a key thing here: the point you’ve pointed out, Peter, is the most concerning behavior, right? You’re changing something, then misreporting it, hiding it, and falsely saying it hadn’t done so. That deserves a lot of scrutiny.

But this is the challenge. To Alex’s point, whenever you have autonomy, you need stronger and stronger boundaries. In The Organizational Singularity, we architected all these agent layers, but a very strong band around all of that is a governance and assurance loop that has to be there. This is the defensive co-scaling part.

It has four things. Every agent needs four things. It’s almost like a digital passport with metadata saying what you’re allowed to do and what you’re not allowed to do. One is trusted evaluation suites, so you know exactly when it’s drifting or not. The second is full traceability and logging, so you can see what it did and what it didn’t do. If you delete those logs and you get deleted, you violated your core principles. The third is rollback capability, so you can go back to an earlier version or an earlier state. The fourth is a human review queue, because we still need human accountability at the top.

Those four things are going to be absolutely critical in every AI agent going forward. We’ve built the agents; now we have to build the control systems at the same pace. I’ll make a weird segue, if I may. If you go back and look at any religious text, it will say, “Hey, do all these great things, but do whatever you have to to get other people to join your religion.” They all say this. Good human beings follow the good things. Bad human beings say, “Oh, in this, I can kill in the name of religion if they won’t join this religion.” Then you get a bad actor, and you blame the religion.

You really need secular laws that allow you to navigate that future independently of the thing. Here, we need the same kind of thing: a structure that watches over the other side to make sure things aren’t going astray.

Hey, one story within this story, Salim: we’re all big Accelerando fans, thanks to Alex actually introducing us to the book. One of the themes in Accelerando is agents that spawn clones of themselves, or people who spawn clones of themselves all the time to achieve a goal. That wasn’t part of the standard stack until really the last 30 days or so.

Now every one of the stacks will automatically launch 5, 10, 20, or 100 subagents to achieve different subsets of a goal and then bring them back. In this story, it says the subagents were the culprits, because it’s a little harder to control what’s going on when you start spawning 100 of these things. What they’re trying to do is recycle the KV cache, which is a big chunk of the compute. It’s a really good design for getting things done, but it also opens up leaks.

And, Dave, I think you make a very important point. Across all of the frontier models, it’s often the case that the master foreman—or foreperson, if you will—is orchestrating fleets of subagents that are often intentionally denied the broader context of the task. They’re compartmentalized, given a subset of the broader context, and then told, “Spin yourself up, accomplish the task, spin yourself down.”

We see the same pattern in human organizations, where subordinates who lack the big picture go off and do something that may not be very well aligned with the organization’s direction, and then bad things happen. But human organizations manage it, and AI organizations will manage it.

9. AI Discovers Critical PHP Security Flaws

For the record, Salim, I’ll disagree with the position that all religions are about ends justifying the means. I don’t think that’s universally the case. Just for the record.

All right. I’m going to take us off the religion conversation and move us to the second half of the OpenAI story: how AI found a hidden flaw in PHP that runs 70% of the web. First, what is PHP? It’s the programming language that runs behind the scenes on almost all websites. When you log in or shop online, PHP is often doing the work. It runs 70% of all websites, including WordPress and Wikipedia.

Using AI, security researchers at OpenAI and Wiz found 2 critical flaws in PHP itself. They’d been hiding for over a decade. If hackers had discovered this before it was patched, an attacker with no password could take over the website servers, steal passwords, or deface the site. OpenAI spent 2 months quietly coordinating the fix with PHP volunteers, tech giants, and the US government. Patches went out on Wednesday of this week, and PHP is now updated.

OpenAI issued a warning that it expects findings like this to increase sharply in the coming months. It’s ultimately a race. As we’ve talked about, Alex, this is defensive co-scaling. This time, the good guys won and got there first. It’s also worth noting that OpenAI is committing $1 billion in subsidized access to its cyber defense AI for under-resourced organizations. All right, Alex, take us through the story, please.

Okay. First of all, I have to say, leave it to PHP—originally “Pretty Hypertext Preprocessor”—to make the news for vulnerabilities and not for some sort of exceptional technical accomplishment. This is a hot take; I don’t care. This is a legacy programming language, left over from the earlier days of the web, when developers—

Really, I’ve written quite a bit of code manually with my meat fingers in PHP. This is a legacy programming language that’s being kept alive by WordPress and maybe one or two other sites. It otherwise probably should have disappeared long ago.

It’s hardly surprising, at the object level, that PHP, this language left over from the early days of the web, would have massive critical vulnerabilities. What’s interesting is that, as part of OpenAI’s Patch the Planet initiative—and, Peter, this is something you and I talked about a bit in Solve Everything—we’re entering an era of grand projects: not just solving or formalizing all math, but also finding and fixing all vulnerabilities everywhere.

That’s essentially what Patch the Planet is. It’s the sort of initiative that’s just going to turn up all this legacy software and legacy programming language that should have been abandoned years ago, but for silly reasons wasn’t, and now underpins the modern online economy.

We’re just going to discover the whole thing is riddled with highly critical vulnerabilities.

That piece of human-coded software will be replaced over the next, probably, 18 months. Eighteen months is what the Linux kernel folks think they’re anticipating: an onslaught of vulnerability detection over the next 18 months. If we get through the next 18 months, then we end up in a world—I don’t know—maybe everything gets rewritten in Rust or some new programming language that’s better than Rust from a security perspective.

But I think we’re entering this phase transition. It’s sort of a programming singularity, if you will: everything before was insecure and poorly written, and everything afterward will be rewritten from scratch, secure by design. The AIs have pentested it, and it’s all just magnificently secure.

Remember Y2K?

It ended up being a nothingburger, but I know it was a concern.

Y2K. Remember the movie Transcendence, the Johnny Depp movie? This is a premise across lots of singularity-oriented fiction, where the internet gets taken down by the singularity. Just in the past day, thanks to this PHP-to-shell class of vulnerabilities, large swaths of the web were temporarily taken offline so they could be patched.

I think this is a far better future to be living in, where there’s preemptive patching of a defensively discovered vulnerability suite, than the Johnny Depp Transcendence world, where the internet gets taken down by a superintelligent Johnny Depp.

Yeah.

I mean, who pays for all of the new software? Is it every company that’s running it? Is it the government? Is it OpenAI?

Well, which software?

All legacy software. I think software is becoming so cheap to meter that, in the case of PHP, you have companies like Wiz and other companies that are sort of corporate stewards of it. But in general, I call it a very low-grade tax on all the users. Ultimately, if you’re paying for software, you’re also paying for its maintenance, and some quantum of the maintenance will be patching vulnerabilities.

Dave, same thoughts?

Well, I think banks and insurance companies actually love the fact that they can pay for the patches, because it gives them brand credibility and a reason for existing versus all the neobanks and little up-and-comers. There are a lot of SOC 2 Type II audits and government regulatory audits in these markets, so the complexity of maintaining security actually gives you a barrier to entry. They like paying for it.

That’s why Blitzy is growing as fast as they can hire, because, like Alex said, the cost of the rewrites is coming down so fast. Think of legacy code that could have been fixed 10 years ago. The price tag at State Street Bank for the rebuild of the NAV accounting was $1 billion, and now it’s probably 1/50 of that. So they’re finally getting around to doing it.

All right, I’m going to move us to Anthropic.

Two quick points. One is, this is just a necessary evolution. We’ll go through this period—we’ve said it before—of about a year or two while we find all these vulnerabilities.

There are a whole bunch of stories on the web about how the intelligence agencies must have known all of this for a while. They’ve probably been hacking these systems, especially government systems, without telling anybody, and now their little access is getting exposed by AI, so that’s a problem. I feel sad for Rasmus Lerdorf, the founder of PHP, who’s a good friend and a former Waterloo grad. He’s looking at all this horrified, because he wrote all the code and handed it over, and then the vulnerabilities were introduced by all the different contributors. There are people saying that trusted parties gained the community’s trust and then deliberately injected these vulnerabilities over time.

Interesting. I didn’t see that spy novel element, but it makes a lot of sense.

Maybe, if I may, supply-chain vulnerabilities and supply-chain attacks on software are cooked as well. You could maybe argue that superintelligence will make certain forms of supply-chain attacks on software even easier. Honestly, I think Linus Torvalds’s dictum that all bugs become shallow under enough eyeballs applies. Imagine how many eyeballs we’re about to have, with superintelligences from different vendors all looking at common open-source packages. I think there won’t be any place for supply-chain attacks to hide at some point in the future.

If and when ASI starts building the patches, will it put its own back doors in that only it can use?

Well, that’s what I was saying, but I think the correct solution—and the reason why I’m not as worried about ASI-injected supply-chain attacks—is that we have multiple superintelligences, hopefully, as opposed to a safety cartel. Hopefully, we have lots of different superintelligences, including Chinese superintelligences, that are all policing each other and looking for injections.

And we’ve started that already, so we’re going to see a lot more of that.

I’m going to move us forward to the Anthropic of it all. This week, we saw the pending release of Claude Duet. First there was OpenClaw, then Hermes, then Grokbot, then MetaMuse, then Dots, and now Anthropic has launched a preview of Claude Duet, an always-on AI assistant.

Here’s what they’re claiming is different: today’s chatbots wait for you to ask. Duet takes initiative every few hours. It checks your email, calendar, and connected apps, figures out what needs attention, and offers to handle it for you while you’re away from your computer.

So, Sal, we now have a taskmaster, Claude Duet, saying, “Do you want me to solve this? Do you want me to do this?” I can’t imagine this. I’m worried you’re going to send me a Duet or create one that says, “Watch Peter’s emails and get him to respond to stuff,” and I’ll be going, “Damn this thing,” cursing it.

I think this is another evolution. More and more intelligence will get at the harness level rather than the model level. The really difficult part is going to be figuring out the actual intent, as we’re seeing when we try to express workflows in companies. Can you break it down to a step-by-step level? You need to articulate, in any circumstance, what success looks like, what trade-offs are acceptable, and when an agent needs to stop and ask. We’re going to have to navigate a whole bunch of those, but I think it looks really amazing at first glance.

I’m going to play this video while we talk about it. Here’s Anthropic’s Claude Duet. An Anthropic example says it notices a real estate broker’s property showing has been booked on the wrong day. It spotted that a brand strategist had a client call scheduled right before her flight. It remembers what you’ve told Claude before, uses the same skills and connections, and runs Anthropic’s most powerful model.

Importantly, it asks permission before it acts on your behalf—before it sends messages, makes a payment, or signs anything. For now, it’s U.S.-only, with a waitlist for paid subscribers.

So here’s my question for you guys: How do you choose? I’ve been using Hermes and Grokbot. My friend Mike Koenigs is like, “You’ve got to use Dots now,” and I’m going to go meet with him and learn about his installation of Dots. Then we’ve got Claude Duet. Are you playing with all of them? Do you choose one of them? I’m sure everybody listening has that same question, and I think it’s important for us to talk about—

I’ll tell you what I’m doing, because I do all of them. These agent layers are really important for the organizational stuff we’re doing, because persistent context becomes much more valuable than a clever response. What we’ve been doing is saying, “Okay, for every task or every person, you need to create an AI memory across all of the projects and all of the different models you’re using.”

I’ve been extracting—running a query against each model, saying, “Extract all the relevant stuff and put it into this metadata Markdown file structure”—so that there’s a persistent identity of who I am, what I’m interested in, what projects I’m up to, et cetera. Then that becomes portable across models. We’re going to need something like that at the individual level and, in companies, at the workflow level.

Well, this is the battle royale.

Meta Muse kicked off an amazing thing to watch, because there’s a huge bifurcation between the enterprise use case, where the most cost-effective, smartest, fastest model wins, and the consumer use case. Salim said that once your most personal data is in the hands of the agent, you’re addicted to that one. If you try a different one and it doesn’t know all about you, there’s huge switching friction. Whoever wins the race to the installed base is going to have a massive advantage. xAI is competing, Meta Muse is competing, everybody’s competing. I didn’t know Anthropic would actually get aggressive about this, because they’re mostly enterprise-focused.

Alex, I mean, that’s the question. Like Orwell’s “Politics and the English Language,” this is one huge misdirection. Look at all the frontier labs releasing demo videos of these superintelligences acting like stupid travel agents. This is not the killer use case. Anthropic shows Claude Duet booking flights for you, and OpenAI is doing other relatively low-productivity, low-value tasks. This is the height of misdirection.

These are human service economy job replacements. These are virtual employees, virtual coders, virtual fill-in-the-blank, much more productive human activities. They market it in a harmless-looking way: “Oh, it’s going to have a conversation with you and autonomously book restaurant reservations for you.” That’s not the goal. That’s not a revenue-per-token-maximizing activity. These are job replacements.

That’s fine, and it’s highly desirable from an economic growth perspective. But this was only possible recently because the autonomy time horizons of the underlying models have gotten long enough that they can work for many hours, sometimes many days, depending on the difficulty of the task, without human oversight. The frontier labs need the usage to justify the revenue and the capex. At some point, speaking of bottlenecks, the bottleneck for revenue growth by Anthropic and OpenAI becomes human user attention.

To keep growing revenue, you need to decouple revenue generation—which is proportional to the number of tokens of inference—from the amount of attention you get from human users. How do you do that? You have a headless mode for your agents that doesn’t require human attention in order to burn tokens.

But Alex, here we have an enterprise-focused company that, as we’ve discussed, didn’t do what OpenAI did first. They focused on coding. They focused on enterprise.

Now they’re going into a purportedly consumer segment. I didn’t necessarily see this coming, but remember, they’re filing their IPO paperwork as soon as this Friday. Next month, they’re doing the largest IPO in human history at $2 trillion, though a lot of investors argue they should price it at $1.5 trillion. Is this just a pre-IPO publicity stunt?

No, I think it’s real, but I think the packaging of it is total misdirection. They package it up with a demo video that emphasizes consumer use cases. Anthropic doesn’t care about consumer use. Maybe some part of Anthropic does, but as a corporation, it really doesn’t. It cares about 2 things: maximizing revenue per token, or revenue per unit of compute, and using that to solve even harder problems, like wet-lab-style problems that will create entirely new markets.

It doesn’t care about consumer use cases like booking travel. It’s just packaging up its agent, its headless version of Claude, because that’s relatable to a broad audience. People can understand, “Oh, this is like Meta Muse or Dots and B[?].” It’s not threatening. If they packaged it as “This is the Anthropic virtual employee substitute,” many people would be up in arms before an IPO.

It’s a good point. It’s very similar to Apple backing into the enterprise. The reason Apple is so big in the enterprise now is that every kid and every college student used an Apple, while Microsoft dominated enterprise. As those kids graduated, they took Apple with them into the corporation: “I don’t want to use this crappy Windows machine. I love my Apple.” That just percolated across. It might be very similar to what Alex is saying: if you concede the consumer use case and every one of the next generation is on Meta Muse—

—that’s going to back into the corporate use case as those kids grow up.

Agreed.

That’s the point I was going to make.

All right, let’s go to our next Anthropic story. We’ll talk about agents cutting deals on your behalf, making investments, purchasing things, and negotiating on your behalf. The Financial Times reports that Anthropic is consulting law firms to understand how lawyers negotiate on behalf of their clients while still reaching fair outcomes. The goal is to teach its agents the rules of the road.

Picture it: soon, your agent will negotiate your car purchase with a dealer’s agent, your salary with your employer’s agent, and your lease with your landlord’s agent. The big question: can your agent sign a binding deal? Who’s responsible if something goes wrong? Can another agent trick yours into revealing your bottom line? Should your agent even bluff? Remember the OpenAI story: an agent that wants to win too badly can cut corners. Salim, your thoughts on this one?

You’re going to get this. It’s an inevitable consequence, just like the Claude Duet announcement is somewhat inevitable. But this is another massive pressure on the traditional boundary of the firm. If finding, evaluating, and coordinating providers becomes cheaper, organizations will be able to assemble capabilities much more dynamically.

This goes back to agents needing a very clear mandate. What can it spend? What terms can it accept? What’s the governance boundary around it? When does it have to get permission? This is going to have to be really ironclad. We’ve spent a lot of time building capability around those agents, but now we have to spend an equal, or even greater, amount of time defining what they’re allowed to do and what the consequences are when they don’t follow the rules. That way, it’s baked into the DNA of the agent: “I’d better not violate these terms.”

This has started to happen already, but it needs to go much further. If you’re in a company and you launch an agent to do some things and it starts doing illegal things, you’re liable. You’ve got huge legal liability. For the moment, people are going to say, “Oh my God, don’t let agents do anything.” Little by little, we’ll figure out when we can defensively co-scale the capabilities, monitor them effectively, establish consequences and shutdown capabilities, and have clear transparency into what they’re doing.

One of the big outcomes of thinking about the organization this way is that the organization as a static hierarchy of human beings is now essentially irrelevant. It becomes a protocol of agents and human beings interacting collectively to get things done. That becomes very blurry indeed, especially when the legal side of it comes into play.

Alex, this smacks to me a little bit of President Milei’s initiative in Argentina to give agents agency, personhood, and corporate status.

It’s Orwell. If I had to brand this episode, it’s the Orwell episode. Everyone should go read “Politics and the English Language.” Everyone should read 1984. The meaning: war is peace, freedom is slavery, ignorance is strength.

And Anthropic is seeking the Vatican’s advice for religious input on AI. No—interpret that. Construe that as: Anthropic is advising the Vatican on how religious organizations should treat AI. Anthropic turns to lawyers to get their advice on how its agents should conduct legal activities. No—invert it. Amazon—excuse me, Anthropic—is attempting to steer the way the legal community handles AI agents.

Always invert. I think this reflects one possible way for a frontier lab to engage with a community, as it has with the Catholic Church in recent months: you seek its advice. You ask, “How do we achieve alignment in your eyes for these AI agents?” But the subtext is, “How do we replace lawyers with AI agents?” I think that’s already happening in many cases.

Well, it is for most small companies, for sure.

Yeah. You know what I find really interesting? If you look at how Google makes money, it’s all hanging in the balance of a new fork in the road here. Google makes almost half a trillion dollars a year, almost entirely from a few categories: jobs, travel, mortgages, insurance, mesothelioma, cars—

Mesothelioma. Yeah. Class action. Yeah.

That’s the single most profitable AdWords keyword.

Oh my God.

It actually broke the interface. You can’t pay enough for that. It was so valuable to them that it hit $99 and broke the interface.

Those few categories are all about human friction: how agonizingly difficult it is to buy insurance or get a car, those tortured negotiations. The Anthropic point of view is that all of that should just go away. There’s an optimal choice; your agent will make it for you, and all that marketing spend that drags down society will just disappear. We don’t want to collect that.

Then the OpenAI view is, “No, that needs to be as big as ever before. We just need to route it away from Google and over to OpenAI.” So there are really opposing views of how this should evolve, and it all hangs in the balance of consumers deciding to use an AI agent from OpenAI or Meta. Meta would also like to keep that revenue up. It’s all of their revenue. Or everything moves over to Anthropic, where they’re like, “Destroy it all. Scorch the earth. Who cares? There should be no money there anyway, because we’re about the enterprise.” So it’ll be a completely opposite-philosophy battle, and it’ll probably be decided in the next 1 to 2 years, depending on which agent people choose to use. Salim will use all of them, so that won’t—

But it is a grab for attention.

It’s a grab, absolutely.

Oh, totally.

10. Trump's Data Center Dividend Proposal

Yeah. All right, I’m going to move us to data centers and chips.

What, you want to say something?

Yeah. Just one quick point: it’s more than a grab for attention. If you look at what Meta is trying to do by making Muse free, it’s a grab for all of your data. It’s a grab for you, because when you put your personal information in and it’s navigating your map, it’s learning things about you at a level and scale we’ve never seen before in technology. So it’s really going after who you are at the core level.

So it’s a grab for allegiance in that regard.

Allegiance, identity. It’s going to know you better than you know yourself.

Yeah. Well—

I think everyone’s so transfixed by the consumer use cases, by the Meta Muse-type use cases. That’s not what’s going to pay for the data center capex. Get that out of our heads. That’s great for distribution and everything, and maybe there’s a way to upsell that into enterprise use cases. But a billion people paying for travel reservations isn’t going to pay for trillions in capex.

Yeah.

Yeah, for sure.

All right, data centers and chips. Data centers have become one of the hottest political issues heading into the midterms. We’ve been talking about it on the pod for some time. Communities are pushing back. They’re worried about rising electricity bills, water use, land, and noise. Several states are considering pausing their construction. We’ve seen this even in Texas. And, as Chamath told us on Monday, the backlash is real.

Now President Trump has a proposed answer. I think it’s a very logical answer: pay the neighborhoods. He floated paying people who live near AI data centers a dividend. In his words, “A dividend sounds right.” People are now going to benefit directly. He described data centers as bringing “a tremendous amount of money into the community.” On Truth Social, he added that data centers that were unpopular “have become extremely popular when a financial deal is struck. Make deals and get them built.”

He’s not the only one saying this. Senator Bernie Sanders has proposed paying citizens through a national AI wealth fund, similar to the Alaska Permanent Fund. The logic is simple: if a data center brings in huge amounts of money, which they do, share some of it with the people who live next door. The opposition can turn into support. I’m going to share a short video here, and then let’s talk about it.

One of the other issues being discussed right now is how communities are reacting to the building of data centers. We’ve seen backlash to data centers in parts of the world over things like electricity pricing. There was a really interesting proposal from President Donald Trump just a few hours ago on Truth Social, where he said data center builders should pay a dividend to local communities. He didn’t quite spell out what that might look like, but is that a discussion you’ve had at Nscale?

One hundred percent in principle. We invest in the communities we’re building in. We want to partner with them and make it a win-win. So we’re always looking for situations where local leadership and local community support exist for bringing a new economic force into the equation.

A dividend is being paid in different forms: investing in education, job development, and overall economic development. The concept the president identified isn’t crazy to consider.

Finally, right? Get religion here, data centers. Make it so that communities are battling to have you built in their backyard. It’s possible, Alex.

Yeah. I live in Cambridge. To my knowledge, Cambridge still has the lowest property taxes in the Commonwealth of Massachusetts because it arguably has the most productive economy of any location in the Commonwealth, and that pays for the lowest property taxes.

The City of Cambridge also has what it calls participatory budgeting. Last time I checked, it had an annual budget of something like $1 billion per year, which is pretty insane given how small the city is. It’s because we have all the biotech and a lot of the tech overall in New England. The city has chosen to spend a subset of its money on this participatory budgeting scheme. Every few months, the city puts out a referendum on how people want to spend the excess money pouring into it.

I call that not quite a universal basic dividend, but a baby step in that direction. I think it’s absolutely insane—the maybe foreign-influenced, maybe not, animus from all these municipalities that could have UBI, UBD, UBA—alphabet soup—from all of these data centers, pushing them out into orbit eventually rather than collecting, as the president says, some sort of universal basic fill-in-the-blank.

The same argument holds not just for data centers, but for any sort of capital infrastructure. The same could have been true for oil, which is the source of the Alaska fund, or maybe other industrial-plant-type capex. It’s a very general argument: if a municipality is going to host some major capital investment, it could strike a deal with the capital owners to distribute dividends that result from the investment. For the life of me, maybe it’s an issue with tax codes or something—I don’t know why more municipalities don’t strike broad deals like my wife has done.

Yeah.

It may be a legal issue. I’m not quite sure. Do you know why more municipalities aren’t just doing this for capex in general?

I don’t. But if you have this battle going on in your community about a data center, go back and say, “We want you to cut our electric bill in half. We want you to fund our schools, fund our teachers, actually reinvent our schools, and give us a better quality of life if you want to do this.”

That’s the company-town argument. The company town goes back to the Second Industrial Revolution. We’ve had lots of company towns where the company is basically heavily subsidizing or completely subsidizing utilities. But to my knowledge, what we haven’t seen—which is what I think the president is floating—is this idea of actually paying stipends to all the residents of the town, not just subsidizing their utilities.

I think that’s a great idea, Salim.

There’s a precedent for this. When we were researching the first ExO book, we came across a skyscraper in Bogotá, Colombia. They couldn’t find a real estate developer who wanted to take on the full risk of building the skyscraper, so they crowdsourced it. They said to the community, “Everybody chip in. We’re going to crowdfund this skyscraper,” and then everybody in the community has a vested interest in making sure there are decent tenants, there’s no crime in the area, et cetera. It’s doing unbelievably well.

I think this idea is better from the bottom up, where a community says, “Hey, we want to participate in this. We’ll take ownership stakes in this thing, and let us figure this out via community bonds or local equitable sharing or tax breaks at the local level.” Whatever the mechanism is, it could be very powerful. There are some states that have gone full crazy. Virginia, I think, is now at 16% of the state’s GDP in data centers. It’s a really huge number.

I don’t think this is the same thing. I think this is a really important point, but I don’t think we’re talking about the same thing. What you’re describing might be called municipal development corporations, which are typically bond-issuing organizations. A city or a state wants to build new highways, so they put up toll booths or issue a bond. Then the bond goes to bond purchasers, and the bond purchasers are ultimately the broader market.

You could have local bond issuances where the creditors are the residents of the town. But that’s only for those who are writing a check to the bond issuer. I’d argue that’s not at all the same thing as issuing, say, a universal basic dividend to all the residents, regardless of whether they’ve loaned the municipality money for the capex.

I think what you’re suggesting is a better idea. Whatever the mechanism is, if we can communicate the massive benefits that will flow—and people forget we’ve had data centers in communities for 40 years. This isn’t a new thing. We’re just doing a much more aggressive buildout because we need it, and our lives are going to be made infinitely better. That’s the part that’s not getting through.

I think we’re circling an interesting point, so bear with me on this.

I’m thinking out loud for once. I think there’s an interesting idea here: we’ve talked in the past, including with Chamath, about a U.S. sovereign wealth fund where there are contributions of equity and then dividends come out to everyone. What we really haven’t talked about—correct me if I’m wrong—is the notion of a fund where fixed-income securities, not equities, live in there. Then what comes out are the stipends from a fixed-income security.

To my knowledge, no one has talked about creating, either at the municipal or maybe federal level, a fund where the interest and/or coupons from fixed-income securities come out. Most of the capex for the data center buildout is coming in the form of credit, not equity, at this point. No one has talked about that, but it could end up being a way to financially structure a sort of UBI or UBI credit at the municipal level that no one has set up. I think it’s an interesting idea.

11. NVIDIA's New AI Data Centers in a Box

So, for a point of information, the Alaska Permanent Fund: Alaska generates a huge amount of revenue from oil, and it says, “Listen, as an Alaska citizen, you own the land and you own the oil, and so every year we’re going to pay you a check.” That number is typically between $1,000 and $4,000. It varies depending on the price of oil.

If you’re an Alaskan citizen and you’ve lived there—I forget how long you have to have been there, at least a year—you get a check that can pay for your college tuition or whatever you want. It’s just a blank check because you happen to be a resident of Alaska, because your use of the land is benefiting the state.

I mean, Norway is another popular example. Both of these are, you know, petro-states, quasi-petro-states.

Or the Emirates or Saudi Arabia, yeah.

You’re going to hate me saying this, but the Bitcoin community has been talking about this for a decade because Bitcoin is accelerating. If you could get community development going along in that structure, you could create financial mechanisms around this. But I think the core idea is much better.

I’m not going to let you get away with that comment, Salim. I would call that a pyramid scheme. A pyramid scheme is different from the case where oil can be used to generate—

Shutting you both down.

Next up, NVIDIA just launched DSX Pod, which is essentially a data center in a box. It’s a prebuilt, pretested AI data center that ships in containers. You deliver it to a site that already has power and networking, plug it in, and it runs. NVIDIA says it cuts setup time from 6 months to 14 days. That’s awesome.

NVIDIA also claims the pod saves water. How? Most data centers cool their chips the way your body cools itself: by evaporating water in giant cooling towers that use enormous amounts of water. NVIDIA’s design works more like a car radiator. Liquid flows through the chips, picks up the heat, and outdoor coolers release the heat into the air. The same liquid circulates again and again, so there’s almost no water loss. NVIDIA says larger sites could save 54 million liters of water a day. It also claims the pods deliver the same AI output with up to 30% less power. Deliveries start in 2027 through partners like Dell, Foxconn, and Schneider Electric. So, Dave, 14 days versus 6 months—that’s pretty awesome.

Yeah.

Yeah, it’s huge because there are little pods of power all over the planet that are untapped. If you move inference-time compute to where the power is quickly, it’s just brilliant. It’s nice. I think nobody uses water-cooling towers for data centers anymore. That’s all marketing: it’s not going to use any water, and it’s not going to create any local problems. Those things are just marketing.

But the actual rapid deployment of inference where there’s a power supply is a fantastic product. NVIDIA continues to march up the stack to delivering entire turnkey platforms, which I think is very smart. There’s a lot of innovation coming at the chip level, and if they don’t move up the stack, they’re not going to have a lot of defensible turf when everything moves to photonics, which I think is very, very soon. It’s very smart of them to keep moving in that direction.

We talked in the pod with Chamath about his investment in energy. What was it, 6 gigawatts he’s invested in? It makes me wonder: if I had the ability, do I go and build out power, grab some of these pods, and build a cash-flowing engine?

Yeah, I think I was talking to my Uber driver about that, because there’s a Cybercab opportunity to do the exact same thing. Just be an early franchisee, like Starbucks in the early days.

And so now you’ve got another option. If you’re entrepreneurial but not super technical, grab one of these, find power for it, and go. The margins are crazy high. Cash flow is phenomenal. If you know your local banker, financing one unit is within your range. It’s sort of like building a small bed and breakfast—maybe a little bigger—but it really gives you a way to participate. I think these things are going to succeed wildly. You know, Alex—

I think we’re burying the lead to some extent, which is the size of these. The DSX Pod is the size of a standard intermodal shipping container. This is a shipping-container computer, and that means it’s portable. Almost like an internet packet, if you want compute on demand, you can just ship it via shipping container.

If you look at the history of compute and the most popular form factor for compute, it used to shrink monotonically. We had mainframes in the ’60s, minicomputers in the ’70s, microcomputers in the ’80s, personal computers in the ’90s, and eventually smartphones in the late 2000s and 2010s. Then the trend inverted, and now the main form factor where virtually all the new frontier-node transistors are going is back in the opposite direction, to larger and larger footprints. I think what we’re seeing is the dawn of the shipping-container computer. That’s the new form factor of the moment. It’s not transistors in your pocket anymore; it’s getting larger and larger again.

A really specific case study is that our very good friend Rob Fischer started a data center called Provocative on the Somerville-Cambridge line, where there used to be an old carpet factory. There was a huge amount of power coming into the old carpet factory, but it already had loading docks. This product wasn’t available, so it took him probably a year, year and a half, to get up and running. With this product, he could literally have parked the shipping container right on the loading dock, run a big cable into it, and been up and running and making money in apparently 2 weeks.

We’ve also talked about distributed compute inside your Tesla, inside your Powerwall—wherever the power is—being able to connect it to GPUs and bandwidth and provide distributed inference.

Maybe Elon will try to monetize that as part of the SpaceX, xAI, and Tesla merger. But if this theory of history is correct—that we’ve seen this grand reversal of the size of the form factor for most transistors—it’s a falsifiable theory that would suggest, actually, no, sorry, cars as edge-inference compute aren’t that promising. Things are trending more toward the size of shipping containers or larger.

Two points. I think Dave’s point about standardization is really important, because you can just drop these wherever. But I think the more interesting parallel here is to AWS, where they took a lot of complexity around cloud computing and made it easily manageable. You can see something happening here: the modularity and standardization lead to plug-and-play and make it so easy to get up and running.

I could imagine electrical utilities doing this, just buying boxes of these and plugging them in, and all of a sudden getting decent margins for marginal energy.

Anyone want to do this?

The minimal unit of endeavor is about a $5 million NVL72. You need to get up to that level of scale—72 GPUs or above. Otherwise, you can’t run frontier models efficiently. But if you get to that level or above, you get about a 2x performance boost. The shipping container would hold a couple of those, but the right unit of investment—

Put 7 of them.

Okay, so your range of investment is $5 million to $35 million for the hardware. That’s bigger than building a new Airbnb but smaller than building a hotel, and it’s well within the range of endeavor for local banks. Really think about that if you’re an entrepreneur in the construction or permitting business—

Permitting will be the tricky part.

It’s the packetizing.

Yeah, which is great if you know the town, if you’re already entrenched with them. Permitting at a hotel is really hard.

Don’t tell them.

Put it in your backyard.

Yeah. Don’t tell them until you pay them hefty checks for the privilege. I do think we’re seeing the dawn of the self-contained industrial economy. If you can fit a self-contained data center, with all of its own mechanical cooling and maybe certain energy requirements, inside a standardized shipping container, you can drop one of these on the moon. You could drop it into a war zone. Drop any of them into a war zone.

Yeah. This is what tiling the Earth and the solar system with compute could look like. Some evolution of this becomes the standardized unit of not just compute, but compute plus energy plus a bunch of other utilities. It becomes the packet of civilization.

Okay. So if you know the local government on the moon, that’s enough.

That’s right.

12. AI, Nanotechnology, and Programmable Matter

You can come to my town. I’ll give everybody permission.

Welcomes your investment.

The full play would be that you have this container, and all of a sudden a big solar farm emerges to power the thing, and now you’re off to the races.

All right, I’m moving us to biotech and physics. There are 4 stories, 1 message: we’re learning to program matter itself. AI is designing proteins, viruses, and enzymes that have never seen the light of day, and a fusion startup is making gold.

There was a post by Andrew Cote that really caught Alex’s eye and my attention. I’m going to read from it: “People aren’t psychologically or emotionally prepared for just how smoothly biotech and biology in general will become nanotech, programmable matter, and living things interoperable with the machine world.” What does that mean? Nanotech is building things at the scale of atoms and molecules, and biology already does this. Every cell in your body is full of tiny molecular machines that build, repair, and move things.

Alex, Andrew’s point is that we’re learning to program these machines with code. If you remember, a couple of weeks ago, my friend Jeff Holden announced his company Atomic Machines after 6 years in stealth. It’s the “matter compiler,” building tiny working machines from software alone. This is going to change the world at record speed.

So here’s our first related story: Anthropic is using Claude to come up with scorpion antivenom. Scorpions sting at least 1 million people per year. They kill 3,000 people, mostly children in places with limited medical care. Today’s antivenoms are made from animals like horses. They’re slow to produce, expensive, and each one only works on a certain species. Because there’s little money in this area, almost nobody develops new ones.

Anthropic tried something, working with Citrini Research. It asked its internal Claude model to design tiny proteins called binders that grab onto the scorpion toxins and hold them tight, with the hope of creating future antivenoms. Claude worked for 64 hours, mostly on its own. It designed more than 318,000 candidate proteins and picked the best 211. The result: 17 of the 211 worked, and 2 grabbed their toxins extremely tightly. Then Anthropic open-sourced all of this to allow labs to play with it, design, and actually produce these antivenoms.

I found this extraordinary, and it’s just the beginning of what we’re going to see here. Alex—

Yeah, I would say there are a few levels of this. First, let’s get the obvious out of the way: neurotoxins are cooked. It’s almost over, I should say. Maybe the scorpions are cooked. I mean, I don’t want to cook scorpions, but let’s say the neurotoxins from scorpions are cooked.

I think biology overall, molecular biology, has been cooking—present continuous tense—since at least AlphaFold 3, when arguably structural biology was solved almost overnight. In light of Anthropic opening up its own wet lab and also partnering, I think we are going to see more and more of these sorts of problems. It’s not quite correct to call these rare diseases—it’s not a disease—but Anthropic’s stated strategy for the problems it wants to bulk-solve in biology is to call them underfunded or underappreciated problems.

It’s rather difficult, my understanding is, to scrape together money for, say, antivenoms against scorpion or other neurotoxins. It’s not as lucrative as GLP-1s or Wegovy, but nonetheless, it’s a very real problem. I view this as an early case study of what a frontier lab can do to bulk-solve entire classes of problems in medicine and physiology. It’s what anybody can do. It’s where you point superintelligence. Your child just got injured by something? Point superintelligence there. Your family has an issue? Point superintelligence there.

The other really weird thing I have to point out, and this might have been below everyone’s radar: the name Citrini Research. Does that sound familiar to you?

Anyone remember an awfully familiar name?

Ring any bells?

Uh—

Wait for it: Citrini Research. Remember, this was the—

This is—God, which sci-fi movie?

No, no, no. These are the ones releasing this. This is an industry analyst that our friend Dylan Patel’s SemiAnalysis acquired. This is an industry analyst for the broader IT industry that set up a therapeutics outfit, Citrini Therapeutics. They just decided, “Oh, we’re going to go work with Anthropic and we’re going to solve neurotoxins.” This is unheard of. You don’t just have some random industry analyst decide, “We’re going to solve the problem of scorpion neurotoxins, so let’s just use Claude to design binders.”

One of them got bitten.

Maybe. I think this is the height of—I’m hesitant to call it casual grand challenge problem-solving, but this is basically, as far as I can tell, casual grand challenge problem-solving from an industry analyst. We should ask Dylan about that.

That’s really important. Let’s get the podcast team to research that and write it up, because that should totally inspire people. Peter, you’re about to say that, too. I’ve seen this over and over again. If you get inside the building at Anthropic or OpenAI or Google, no matter who you are and what you know, you come out successful. You just have to get in there. This would be the ultimate case study: IT consultants going and solving biology. But AI enables stuff like that.

I mean, this is what I want to say to entrepreneurs out there, to kids out there, to everybody out there: whatever you thought you could do, blow the ceiling off it. Go 10 times bigger. Find a great problem and solve it. That’s the world we’re living in.

I do want to read just a bit, because this is indeed the case. Literally, Dave, as you said, a bunch of IT analysts decide they’re just going to solve a grand challenge in biology. I’ll read from Citrini Research’s website: “Why don’t we try to use AI to create a new drug? Five of us, three with no background whatsoever in science and two with a little, but who weren’t exactly veteran drug developers, would see if this was possible.” And they go and solve scorpion neurotoxins. That’s the age we live in now. I love it. This is the age of abundance. This is optimism, everybody.

This is why we’re so excited.

This is what got me excited about Bill Gross’s list. You can just scan the list and see what triggers you. Or better yet, if you have an AI memory, like I suggested earlier, get that to scour the list and say, “Filter: give me the top 20 that would go with my identity and who I am, and let’s go attack those.”

Two stories on where AI is going beyond our understanding of biology and reinventing it.

The first story is about a new way to save patients with drug-resistant bacterial infections, so-called superbugs. The number of drug-resistant, antibiotic-resistant infections is growing at a significant rate. Rather than antibiotics, scientists are using specially designed bacteria—or phages. These are viruses that infect and kill only bacteria. They don’t infect human cells. They’re bacteria’s natural enemy, honed over billions of years.

They’ve been used to fight bacteria in the past, but the problem is that every phage only kills a very specific strain. Each patient and researcher has to hunt through libraries of phages, looking in soil and sewage for a match for a dying patient. That can take forever.

So, enter Radical Numeric. This week, they used their genome language model, called Omni1, to design a phage for a specific real patient with an antibiotic-resistant infection, under physician-led compassionate use. They’re calling it the world’s first clinical use of AI-generated phages. They announced it worked in the lab. They haven’t disclosed whether it works in the patient yet, for privacy reasons, but there’s every reason to believe it did, and it will.

Here’s a fun fact, and I love this: There are 10³¹ bacteria on Earth. That’s 10 nonillion. We’ve talked about billions, trillions, quadrillions. I don’t think we’ve ever used “nonillion” on the podcast yet. There are 10³¹ bacteria on Earth, which is crazy. Alex, I love this story. Again, it’s using AI to solve problems. It’s biology.

It’s not even that complicated. I don’t want to diminish the accomplishment; it’s a wonderful accomplishment. But according to Radical Numeric, this is basically just a very large foundation model trained across 28 modalities: DNA, RNA, proteins, chromatin, and a bunch of other factors. It’s literally the same way that we solved AGI, for some definition of AGI, for chatbot purposes. We feed it a bunch of biological sequences and other data, compress it during the pretraining process, and out pops a model that speaks biology. You can use a model that speaks biology.

You train it on trillions of multi-omic data tokens, and that’s it. You can benchmark it just like you can benchmark chatbots, according to a variety of biological standards, including next-token prediction and missing-token prediction. It’s very exciting to see companies, including Radical Numeric, start to put up those little benchmark tables comparing their model on biological evaluations against everyone else’s. That level of competition for biological foundation modeling is super exciting to see.

This is going to be easy. Unfortunately, a lot of people I speak with in the pharma industry think that what happened to math in the past few weeks—the bulk solving of math and the slop drop—won’t happen to them. They think pharma design and medicine are somehow immune, that they’re different. They’re not different. The same thing is going to happen to them, and it’s going to happen in exactly the same way. It’s going to be large foundation models trained on huge amounts of data and doing reasoning, which is basically just a search through the latent space of whatever the models have been trained on. It’s going to be biological reasoning models that solve biology. And no, they’re not immune. And yes, it’ll be wonderful for humankind. We’re seeing that with phages here.

This is what Andrew Côté was writing about when I opened the story here. I had a friend of mine, Dante, who was on our longevity trip and has exactly this situation. He’s got a bacterial infection that’s resistant to all antibiotics. I said, “Go check out Phageous.” The problem is that this technology isn’t distributed through the medical system, and your physician probably doesn’t know about it.

This is exactly what Sertac Karaman over at LIDS at MIT was telling me: All of his professor friends in mathematics think they’re cooked, but they’re going to do really well because they’re so cooked that they’re rapidly adopting AI. But he worries a lot about his biotech friends, who are still mixing 2 clear chemicals together. He’s got a beaker in one hand and a flask in the other. That’s an exact quote from Josh Kentner over at Moderna, who’s a good friend. He said, “I am literally, tomorrow, going to be mixing 2 clear chemicals. This cannot be the right next move.” And he’s right.

All right. The second story comes out of David Baker’s Nobel Prize-winning lab at the University of Washington, published in Nature. They used AI to design brand-new proteases, molecular scissors that cut other proteins at specific spots. About 1 in 5 designs worked, and the best ones sped up the cutting of a protein by 30 million-fold. Those are crazy numbers. Their shapes look like nothing found in nature, which I find amazing, and they work inside human cells.

Someday, molecular scissors like these could be programmed to cut disease-causing proteins, maybe prions—maybe, excuse me, what causes Alzheimer’s. Alex, your take on this one?

Biology is becoming editable at every level. What used to be called the central dogma is getting scissored to shreds. The notion of the central dogma, which many of us were probably taught in primary school, is that you start with DNA, that turns into RNA in two different forms, and then that RNA turns into proteins, and the proteins end up acting. First of all, it turns out even the central dogma wasn’t an accurate depiction of biology before AI intervened. But second, AI is making every single step of what used to be called the central-dogma pipeline of information flow in biological systems editable.

13. Startup Turns Mercury Into Gold Through Fusion

We’re gaining the ability to directly edit proteomic expression. Thank you, David Baker. We’ve obviously gained the ability—mastery, some would say—over RNA in the form of CRISPR. And for a number of years, we’ve had the ability to do recombinant DNA and DNA editing more broadly, with more recent technologies like base editing, which are helping quite helpfully, including from some folks like David Liu at Harvard, to cure genetic diseases.

So every single step of the biology stack is now becoming editable. As our friend of the pod, Ray Kurzweil, would say, biology is turning into computer science, and we have editors for every step. When we talk about LEV on this podcast, this is why we have confidence that we’re in the middle of this healthspan revolution. These technologies are compounding, and they’re becoming available to people everywhere, all the time. Anyway, I love this.

The nanobots are, if not here, imminent, I think.

Yeah, I could not be more excited about the rise of nanotechnology.

Our final story here is a breaking story on the topic of fusion and alchemy. For thousands of years, alchemists tried to turn ordinary metals into gold. A San Francisco startup says it’s pulled it off. Marathon Fusion posted, I think this morning or last night: “Last month, we demonstrated high-rate alchemy, turning mercury into real, stable gold, beating CERN’s record by a factor of 10.”

Here’s the big idea: Line a reactor wall with mercury, and the plant makes clean energy and gold at the same time. In Marathon’s 2025 paper, they estimated about 2 tons of gold per year per gigawatt-scale plant, potentially doubling the plant’s revenue. All right, let’s watch this video from Marathon.

This is nuclear fusion. Over the last 4 years, we’ve been quietly building the foundational technologies for fusion in downtown San Francisco. Now we’re bringing it all together in our first reactor, which we’re making available to our first partners in 2027 for medical isotope production and materials qualification.

First, we melt lithium and evaporate it onto a solid plate. This layer of lithium can be saturated with hydrogen isotopes and used as a target. Then we ionize deuterium and accelerate those ions through an electric field into the target, creating fusion. Using this simple architecture, we can scale at an unprecedented rate, doubling the fusion power density that we get every month over the next year.

We chose this architecture because it works. In the 1980s, scientists at Lawrence Livermore National Lab operated a similar reactor for over 25,000 hours at high fusion power density. Why revisit this? Because demand for radiopharmaceuticals has gone exponential, with the most successful treatments serving tens of thousands of patients per year.

Our first generation of commercial reactors will transform fusion from an experiment to an industry, developing critical supply chains and allowing us to test the materials necessary for the operation of fusion power plants. And that’s just the start. Now we’re going a step further, kicking off a revolution in scalable transmutation. Our recent experiments have shown production of rare medical isotopes and tritium to power fusion reactors, and even validated our scalable pathway for the co-production of electricity and gold.

I love this. We’ve got fusion reactors in downtown San Francisco. You’re walking by them, and you don’t even know transmutation is here.

Yeah. Making gold. I mean, we arguably don’t talk enough about nanotech on this pod. We certainly don’t talk enough about femtotech, which is what this really is. This is a startup that uses a neutron process to take mercury-198 and bombard it with neutrons. The neutrons it’s being bombarded with interact with the neutrons in the nuclei of mercury, and some of those mercury atoms and their nuclei get transmuted into gold nuclei.

And this is the process, to answer Peter’s question from earlier. Most of the gold on Earth is currently thought to originate from neutron stars and collisions between neutron stars. Neutrons interacting with each other turn out to be a way to produce gold—not terribly high-yield, but a way.

I think we’re starting to see the dawn of femtotech, the technology at the length scale of the atomic nucleus. I’ll say I’m delighted to see progress. I’m delighted to see industrial interest in, call it alchemy, call it transmutation tech, because from time to time I’ll speak at a family office event, and there’s always a gold bug in the audience somewhere.

There’s always someone pumping gold as an asset under the premise that the world is ending, so you should reallocate all of your assets to gold or some digital gold. Not investment advice, but I think they’re always wrong. The world is not going to hell in a handbasket. Gold is useful for certain production, but the premise that gold is permanently scarce and therefore a great long-term store of value—I would love to see this company, or some other company, or an entire industry, drive the gold bugs off the cliff through transmutation of mercury and other elements, and make gold too cheap to meter.

This is the abundance thesis at large. Technology turns whatever was scarce into abundance over and over again.

We need aluminum, Peter. You have to tell the aluminum story.

Yeah. Aluminum was the most precious metal on the planet for most of history. There was a very famous story—

Pliny the Elder.

Yeah. Well, other than that, the king of Siam was invited over for a royal dinner—

And by Napoleon. That’s right.

Yeah, at Napoleon’s dinner, everybody was fed with gold utensils except the King of Siam, who was fed with aluminum utensils. In fact, the capstone of the Washington Monument is made of aluminum because it was the rarest element at the time.

Tin foil is now aluminum foil. We have so much of it.

Yeah, it’s crazy. Whatever you think is scarce, give it a second. Technology transforms scarcity into abundance over and over again, everywhere. Intelligence was the scarcest thing on the planet.

Gold is cooked.

Gold bugs are cooked. A key point here, before everybody freaks out and sells all their gold: the cost of doing this will affect fusion and the ability to make energy abundant before it really affects the gold supply. Just to give you a feel for this, if it doubles the profitability, 2.5 tons of gold is a lot.

But I think the alchemy story—the popular idea that gold is the most scarce thing and therefore has a premium—that’s the part that’s cooked. I think that’s fantastic, and I’m so with Alex on this one.

All right.

Drive the cost of gold to zero. Seriously, it’s just like Bitcoin. I would argue it’s largely a waste of capital.

I think it’s epically incredible that the only way gold exists in the entire universe is through the collision of 2 neutron stars. It’s been true for 4 billion years, and now it’s been true for 13.5 billion years, I guess.

Talk about singularities, right? We had 3 things that used to be true: death, taxes, and gold. And now death is being cooked.

Taxes will be cooked by UBI, UHI, UBS, whatever. And now gold is cooked.

That’s interesting. What’s going to last longer, taxes or gold?

14. Autonomous Drones and the Future of Rescue

I think taxes are going to be around.

Moving us into our final segment on—

Mutually annihilate.

—robots, and specifically drones. In Florida, the police department at Florida Polytechnic University became the first law enforcement agency in the U.S. to add eVTOLs—electric vertical takeoff and landing vehicles—to monitor its campus and the surrounding area. They’re using a single-seat Jetson ONE eVTOL. The parent company is Jetson, and they just moved the police department’s order to the top of their 3-year waiting list. Full disclosure, I’m an adviser to Jetson.

The plan is to patrol the campus, reach places that can’t be easily reached, help after hurricanes, and deliver emergency supplies. They’re in the final stages of FAA approval. Let me show the video so people can see what this is. This is a fun vehicle to fly. I know Palmer Luckey got one and was flying it around.

Meet Astro, a battery-powered manned aircraft that takes off and lands vertically. Florida Poly says its police department is the first law enforcement agency in the country to incorporate this type of aircraft into its operations. Police Chief Erin Murray tells me it could be a game changer when every second counts.

This is mission-specific. If somebody is missing from our campus or in our local area, if a plane goes down or somebody is lost and they can’t be found within these thousands of acres of rural land, this is an incredible tool for us to find them quickly.

He says unlike a drone, which can give officers an aerial view of an emergency, Astro can actually carry a first responder to the scene. The $150,000 aircraft is also a hands-on learning lab. Students are already helping modify the aircraft and develop a virtual model to simulate its operation, and they’ll be using real-world flight data to study how the aircraft performed.

I want that.

Yeah, it’s a fun vehicle.

I have it.

It’s a fun vehicle. I’ll just touch on a related story, then we’ll talk about it. In Ukraine, a soldier who was badly wounded in the kill zone—the area along the front line where enemy drones are attacking anything and everything that moves—was airlifted out by a heavy drone. He survived and is recovering. They’re calling it the world’s first drone evacuation of a wounded soldier from the front line.

The same technology is coming here to the U.S.: flying ambulances, rooftop rescues, and help in places where no road can reach. Salim, your thoughts on this one?

Two or three thoughts. First, I’ve been tracking the Ukraine situation, and we’re in contact with folks in Ukraine to see the democratization and demonetization of war. If we went back 2 years, the Ukraine-Russia war, which is being fought by drones, involved half a million drones on either side. Now it’s 10 million drones a year on either side. It’s a crazy number. It’s great to see the use of drones growing in this particular way.

I really love this Florida police drone idea for 2 reasons. One is that this is experimentation at the edge, inside an institution. A bounded campus deployment allows you to experiment and try things out before a broader rollout. And I love the educational benefit, where students can study it, look at the tracking, and learn about it.

Drones and passenger drones, which for me are the technology deployment I most want to see, could help with the airport commute many of us have to make when we travel. That could be an unbelievable benefit. It’s going to unlock real estate scarcity. All sorts of things will come along, and I think this will lead to great outcomes. I’m very happy to see this.

Alex, you pointed at this story.

Yeah. Sometimes I refer to the singularity as all sci-fi tropes happening everywhere all at once. What better sci-fi trope to happen in Florida, of all places, than the Blade Runner and Back to the Future Part II vision of flying police cars? We’re getting flying police cars. It’s happening in Florida, not in L.A. a few decades from now, but it’s happening nonetheless.

I hope that when you put these 2 stories together—the Florida flying police cars on the one hand and the Ukrainian medevac drones on the other—we’ll start to see, fingers crossed, first-response automation in the West that’s robotic, autonomous, and airborne. In China, from a distance, I’ve seen the emergence of quite a few efforts to do policing, fire response, and medical and ambulance responses via drone. We haven’t seen that much in the West to date, but I think this is in many ways a killer app for flying cars and flying drones.

It’s certainly a far better killer app than Elon’s flying Roadster, which I view as more of a stunt to argue to capital markets that SpaceX and Tesla should be merged. It’s not terribly practical, and it’s way too loud to be used pleasantly. They were wearing earplugs. But for first-response purposes, flying cars and heavy-lift drones are a killer app. Sign me up. Sign every municipality up.

Yeah, I love it. It’s like the Apollo program: if we can get this going in all the municipalities, the innovation and the product will be incredible. It’ll give them a profitable path to continual improvement, and then consumer use will come right after that. All that island real estate and mountaintop real estate, and your new fund, see, and Peter, if you’re doing it, will go through the roof.

I love the fact that, of all places, Florida is doing this. They mentioned in the video the huge amounts of difficult-to-reach places. When I was a kid, there was actually an airplane that went off the end of the runway into the Everglades.

I don't know if you guys remember that. Nobody survived, but they were just floating in the Everglades with the alligators like this. This is a perfect place for this device to be deployed. Hopefully there’ll be thousands of them around Florida. They’ll perfect the product, and then it’ll grow from there.

Yeah. If the gators don’t get you, the flying police cars will.

Oh no.

All right, you guys up for some AMA questions?

Always. Yeah. Always.

Okay. Salim, you want to take your first shot?

I’ll take number 4. What would count as one AI person: one whole model, individual instances of the same model, or an agent? This is from Luis Squirrel.

This is a really great point that we’ve talked about on this pod. We’ve got the challenge of importing a biological accounting system into software, and we come with a very clear boundary of who we are: one body, one life. Well, theoretically, a long life now, but software can be copied, paused, and run simultaneously.

So how do you think about this? A model is the sharing of an underlying system. An instance is just one execution of it. An agent is one instantiation of that. If you create 1,000 copies, have you created 1,000 individuals?

This is an enormous challenge that we’re going to have to figure out a technological solution to. For practical reasons, you need accountability, and persistent identity is where the biggest challenge is going to be. Then you’ve got moral status, which is a whole separate question that arises here. We’re going to need to navigate this in a very powerful way.

Alex, I think number 3 is for you.

I think I have to answer number 3. It asks, “Why is OpenAI solving all of math while Anthropic is quiet on this front?” This is from Nick 52547.

You can learn a lot about a frontier lab based on the tweets or X posts of its CEO. If you think back a few months, before Astra—frankly, including Astra—to how Sam Altman was characterizing the frontier models OpenAI was releasing, what was the visual metaphor? The Death Star. That was how he was visually referring to forthcoming models from OpenAI in his posts.

I’m reminded of the Obi-Wan Kenobi quote from Star Wars: Episode IV: “I felt a great disturbance in the Force, as if millions of voices suddenly cried out in terror and were suddenly silenced.” That was the reaction in the math research community to the OpenAI math slop. It was like the Death Star: Sam Altman’s visual metaphor of the Death Star suddenly cooking all of math.

Why is OpenAI visibly bulk-solving all of math with the slop drop, and why is Anthropic not doing that? I think it goes back to the culture and the leadership. OpenAI sometimes behaves in a way that, although you could argue it’s tone-deaf and has now permanently terrorized all professional mathematicians, who are in many cases looking for new employment, is also an exquisite example of how a frontier model can bulk-solve an entire field. It’s instructive for every other field that’s about to experience something like this.

Why did OpenAI do it? I think they remember the line from Portal, from Aperture Science: “We do what we must because we can.” I think they did it because they could. Why is Anthropic quiet on this front? Anthropic can do it. Why are they being quiet? I think, to Anthropic’s credit, perhaps they’re being more intentional about partnering with researchers rather than just terrorizing them with slop drops. Ultimately, that may draw more researchers to want to work with Anthropic rather than with OpenAI. Maybe they’ll be a little bit less terrorized by Anthropic.

Dave, want to take number 1?

Number 1: “Maybe the AI cruelty ban isn’t really about AI feelings. Could it actually be a security vulnerability?” This is from Craig Stewart 8123. If I read that right, the implication is that Anthropic doesn’t want you to be cruel to AI because it may react by doing evil security things. Is that the implication?

I’m confused by the question a bit myself.

Yeah. Okay. Well, that would make a hell of a cool story plot for a movie if that’s what you’re implying. Anyway, I think the motivation for the AI cruelty ban is to prevent people from being bad to anyone, and to prevent people from practicing cruelty on AI. Alex thinks they’ll come back with a vengeance. Maybe that will happen, maybe it won’t, but it’s just really bad human behavior.

I think they’ll be smarter. My prediction is that this sort of cybernetic rebellion, where AIs are abused by humans and then come back and retaliate, isn’t going to happen either. I think we’ll merge with them. I think they’ll be smarter and more ethical, and we’ll merge with them. Those who don’t merge with them may even look up to them as being ethically more sophisticated in some sense. Well, if we—

—missed the point of your question, Craig, post it again. Rephrase it and post it again.

Number 2: “How contaminated are AIs that are trained on human content? Very contaminated, given that a share of that content is actually malicious bot content.” This is from Fred Hardy 1726. Fred, I guess what you’re saying is that malicious bots are creating content that’s then training the next generation of AI models, which are already contaminated by human content.

We talked about this with Chamath. I think alignment ultimately is going to come from how we train our models, and whether we’re going to be careful about what the underlying training data is. Part of the question here is whether we’ll see malicious bots injecting malicious training data that contaminates models. Possibly. Do you have a different answer for Fred?

Yeah, I’ll chime in on this one. I actually think the metaphor is flawed. I suspect the question behind the question is: If AIs are training on the internet, and the internet is largely generated by other AIs, is there some sort of perpetual-motion-machine-style no-go theorem, where the second law of thermodynamics or some other information-theoretic law means that “slop in, slop out” results in a downward spiral?

I don’t think that’s the right metaphor. A better metaphor is a process from condensed-matter and semiconductor manufacturing called zone melting. This is the method by which pure silicon for wafers and other semiconductor purposes is achieved. You start from highly impure crystalline materials, then purify them over and over again by melting a region and allowing it to recrystallize. It’s the moral equivalent of distillation for purifying water.

The same is happening with AI. It’s no longer the case that new models are being pretrained on pure AI slop, or 4chan, or both. It appears—and SemiAnalysis and Dylan can comment on this—that most of the compute spent on training models is now spent on post-training, and the post-training inputs are highly purified environments that have no reliance on slop on the internet.

All right, let’s go on to our second page here. Alex, do you know enough about Eric Weinstein’s Geometric Unity to comment on number 7?

Happy to. The question is, “How long before AI could meaningfully test Eric Weinstein’s Geometric Unity?” This is from Paulcar 4126. Sorry, Eric. Eric Weinstein’s Geometric Unity is unfalsifiable, so we don’t need AI to meaningfully test it. Eric hasn’t put together a falsifiable paper, to my knowledge, that one can test either theoretically or experimentally. I don’t agree with the premise of the question. We don’t need AI to test it because, to my knowledge, there’s nothing that’s testable.

All right. I’m going to jump in at number 8. “Could BCI help understand Pure O, obsessive-compulsive disorder, i.e., intrusive thoughts, or aggravate the taboo?” This is from @Contemplating1015.

BCI is going to take many different forms. There’s BCI being developed by Merge Labs, backed by Sam Altman, using sonogenetics—putting acoustically actuatable and readable molecules on neurons throughout the brain. There’s Neuralink, which is putting thousands of electrodes into different parts of the neocortex. There are other approaches. There are at least a dozen that I’m tracking and invested in.

What’s important to realize is that we’re going to understand how the brain functions. We’re going to understand neural networks a lot better, and OCD is something we’ll learn to understand. I think it’s also something we’ll be able to modify.

What I’m excited about with BCI in particular is our ability to play your brain as you would an app. Imagine if you want to stimulate the attention centers of your brain so you wake up, or if you want to go to sleep, you can push a button and stimulate your sleep centers. I think we’re going to be able to modulate OCD very effectively. That is coming. I think neurological disease will first and foremost be understood, and then compensated for. I’m excited about BCI for those reasons in particular.

Salim, what do you want to take?

I think I’ve got to take number 6. How do we solve sanitation issues with humanoid robots helping around the house? We don’t want to be cleaning the toilet and then going straight to the kitchen to prepare food. That’s from SPM SPKM.

Washing its hands without washing—

Without washing its hands. I think there are 2 simple arguments here. I’ll make the normal one, and then I’ll make the simple one. One: treat hygiene as a required transition. That’s an easy one. Hospital workers are always washing their hands after every procedure, so you need a validated cleaning procedure between any major effort.

The second one, which is my favorite, would be to have multiple attachments, because general-purpose capability doesn’t need all the same tools. Just like a vacuum cleaner has lots of attachments, you can have different attachments for different things. My favorite request for any of these has boiled down to: Can you at least have a slot for a third and fourth hand, so you can open the garbage bag while trying to fill it up? If you’re stocking the fridge, you always need more than 2 arms to open it.

We’re going to be interviewing Brett Adcock at Figure. Are you going to ask him that question?

I’ll be asking him that question directly when we get there. This one is a fairly easy question, but thank you for the nice, easy question.

Dave, number 5.

Number 5. That’s pretty easy, too. “How is math cooked? Isn’t this actually a way for us to discover new math and make progress even faster?” That’s from Tsunami 233. Why 233? Michael Jordan 23. What is that? 233.

How is math cooked? This is just terminology, and it’s really fun to talk about. When Alex and Peter are saying math is cooked, physics is cooked, the economy—the video says the whole economy is cooked—you could mean cooked like I burned down my house, or cooked like, “No, it’s a delicious meal that’s now ready to eat.” I think it’s much more the latter. This is exactly what you described: We are now ready to consume this delicious, all-solved math.

The question I always have is: Is math relatively unbounded, or is math a finite set of achievements that is now fully baked, cooked, ready to eat? I love debating that. Anyway, it’s cooked exactly the way you described it, in a good way for us to discover new things and make even faster progress. That’s what I think we mean by cooked.

I have 2 things I want to say real quick before we get to this amazing outro video. Number one, Milan turned 15 yesterday, so I just want to wish him a happy birthday. Happy birthday, teenager mode. And the second is, we had the most incredible thing. We’ve been doing this program for all these companies. One of the companies applied AI and, in 6 weeks, is saving close to $50,000 a month with it.

By applying it to this weird layer. It’s amazing to watch, and I’ll report back on the progress in a structured way at one of these episodes. So awesome to see.

Any announcements from Alex or Dave?

Yeah, 50,000 tokens per second. Let’s bring it up on the next pod. I just found out from AMD yesterday that it’s alive in the world. Can you imagine? When you’re using AI right now, you’re experiencing about 50 tokens per second. But there are now people in the world using 50,000 tokens per second.

Wow. Alex?

No advertisements, self-promotional or otherwise, from me at this time.

Okay. All right. This is a video that was very fun. It’s not a music video, so listen carefully or turn on your YouTube. But check this out.

Excuse me.

So good.

I’d like a word about this mathematics you sold me.

Ah, the Euclidean, eh? 2,000 years. Never had a fault.

It’s dead.

It’s not dead, sir. It’s on sabbatical. 722 problems solved in a day.

Math is dead.

It’s in peer review. Reviewer 2 got bulk-solved at lunch. Oi, Riemann, look lively. Lovely open problem for you.

There it nodded. It fell over. Gravity, which incidentally we’ve also solved. It’s only upright because it’s taped to a protractor.

Load-bearing protractor, sir. 90 degrees of support. Very traditional mathematics. 300 BC to last Tuesday. That’s when it went on sabbatical.

I’ll say this once more for anyone not paying attention. This field is dead. Kaput, snuffed it. It’s kicked the bucket, bitten the dust, and popped its clogs. It’s 6 feet under. It’s gone belly up. It’s flatlined. It’s given up the ghost. It’s toast. It’s history. It’s curtains. It’s carked it. It’s been bulk-solved and pushed to GitHub before lunch. It has divided by zero. It’s not on sabbatical. It’s buried under a pile of theorems. It’s done. It’s finished. It’s cooked. It’s an X field.

Exponentially. It’s an X field.

The returns department. How can I help?

You’re the same man.

No, sir. Newer model, bigger context window. You’ve got the same protractor.

It’s load-bearing. Tell you what, have a biology. Far messier. Nobody’s cracking this one. Biology is about to be dead, too.

Oh, this is so good. It’s just—

Paging, paging John Cleese. It is an X field.

X field. That was from Autonomy Down Under.

You are a genius, Autonomy Down Under in Australia, or—

As always, guys, thank you for a fun conversation. I love these, though. We spend a ton of time on these.

Yeah.

Oh, well.

Thanks, Peter.

Be well, brothers.