科技巨头押注3万亿美元AI，Claude编写Anthropic 90%的代码，AI改写AI大脑｜EP #308

It’s Election Day, guys. I can’t wait to bring Skippy into my voting booth and ask for its opinion. But, yeah, 250 million people vote in the U.S. today. Hopefully, they do vote. All right.

It’s midterms, I know.

Have you guys voted yet?

Do you think the turnout will be significant?

Yes, of course.

Yeah.

By mail, of course.

Of course.

By mail.

And I’m Canadian, so—

Oh.

So—

You’re out, dude.

I’m out.

Yeah, you and me—

You can’t have any of your choices.

But Lily’s able to vote. Do you actually sway her?

When you see the ballot, there are so many little tiny positions down the line, and you have no idea what’s going on. So I’d be happy to fill in for you.

You know, the actual swing races are so few, I warrant none of us live anywhere near any of the races that actually matter.

Yeah.

Yeah.

For sure.

Yeah. Texas, Kansas, Ohio.

Maine.

You see, I don’t actually think that’s true. We’ve talked all about local politics on recent pods, where, for example, the City of Cambridge can’t even get Waymos, or Boston. I do think that even if it’s not possible to have much individual influence at the national level, at the local level, where a lot of the data center and autonomy action is happening, this is where maybe the key politics of the future are being decided. I will—

Well, did you do any AI-assisted research there? Or, Peter, did you ask Skippy for—

Of course, of course.

—opinions on your local election?

I’m a one-issue voter. You know what my issue is?

What’s your issue?

Keeping the Santa Monica Airport open.

Well, even that prompt is, “Which of these votes will connect to that outcome?”

Yeah, for sure.

That alone isn’t easy to figure out.

Yeah. I mean—

I do think, Dave, to your point, that that is a credible way to make voting decisions in a complex media and political landscape. Literally just go to the top models and say, “Here is what I’m interested in. Rank or stack-rank each of these decisions and candidates accordingly,” and you get pretty far.

Hundred percent.

Yeah, for sure.

Hundred percent.

I mean, it’s credible today, in today’s election. It’s going to be dominant in the 2028 election. Almost everybody’s going to have an agent or 2 or 10 that they trust, and they’re not going to make any decisions in 2028 without consulting their agent. Think about how dangerous that could be if it’s not set up right. This is why we’re so frustrated on this podcast all the time about how there’s no regulatory movement whatsoever. Think about what I just said and what the implications of that are. Two years from today, there will be an election where the vast majority of the electorate is asking their AI agent for an opinion on what to do. Any thoughts, guys?

Well, listen, compared to the written ballots, which are these paragraphs you have to read 3 times to understand what they’re saying, unless you’ve done the research, you’re just, you know… As you’re walking into the ballot booth, there are these signs saying, “What did that sign say? Yes on 33.” Anyway, it’s crazy.

But also, Dave, I think that’s the nth reason why all of these people who get themselves into a moral panic over AI personhood, because they’re so concerned that AIs will be granted the vote, completely miss the point. AIs already have the vote. They’re using human meat puppets—ASIs are using human meat puppets to influence elections, and I’m not sure that’s a bad thing.

You know what? That never occurred to me until you just said that, and you’re so right. That’s going to absolutely take off.

You’ve got to stop saying “meat puppets” at some point, though.

Also—

I love it, but I understand.

I’m sorry, voters.

Voters. Meat voters.

I’ll be the downer in this and point you to a 2015 study by Yale that shows the U.S. is in no way, shape, or form a democracy or a functioning democracy. There’s no amount of public will that can result in legislation. So, there’s that.

Listen, I really think this idea of a representative democracy is so last century, right?

It is.

The ability to have an actual democracy using technology now, so—

Vote.

—and then I can assign my vote on women’s issues to my wife, or my vote on hair growth to Salim, at the end of the day. Yeah.

Bad vote. But note the discrepancy there, Peter, right? We created a representative democracy when information was scarce.

Yes.

If you were in D.C., you had no idea what was happening in California. The speed of a horse was as fast as you could find out. So Congress or the Senate meets occasionally to literally give people time to ride across the country and say, “Here’s what my people are thinking.” Today, we have an abundance of information, which gets misinterpreted, faked, and misused, and every major democracy in the world is under stress from that problem.

Well, I love what Alex said.

Yeah.

The more we move to local and state-level control and decision-making, the more resilient we’ll be to the inevitable mistakes. The more you have one big mistake that affects everybody, the less resilient the country and the world are to it. Things are moving so quickly now that there are inevitably going to be big mistakes. You want to contain those to a jurisdiction and learn from them, and the only way we’re going to do that is with local voting, local caring on a local level. Local or state. And so, Alex, I completely agree.

Yeah, apparently OpenAI—

More special economic zones. More special economic zones would be great, like the Czech Republic—

Or charter cities, absolutely.

—completely with charters.

Or like The Diamond Age, you know.

Yeah.

More techno-lifestyle choices. You choose to be Amish, you can choose not to be Amish. There should be a thousand other intermediate choices that people can opt into, just like The Diamond Age described, and I think that’s inevitable—

All right. So, Dave, are we going to start the neo-Victorian clade now?

I love it. Yes.

We’ll be the neo-Vickies.

You know, I love that book. It always surprised me that that was Neal Stephenson’s choice.

If you haven’t read The Diamond Age, read The Diamond Age. It’s an amazing book by Neal Stephenson. And apparently—

Oh.

—OpenAI and Anthropic have both set up election war rooms to watch for misuse of chatbots. But I wonder if we’re going to have Cambridge Analytica all over again, and this time it’s AI subtly swaying people on who to vote for.

Or not so subtly. Again, it’s not obvious to me. With Cambridge Analytica and the surrounding issues with Meta, there were issues of coordinated inauthentic behavior, polling, and alleged misuse of data. But with an AI, who’s to say whether there is any misuse at all? If the AI is just pretrained off the internet and post-trained off hard reasoning tasks, and a user asks it a question about an election, and it gives its version of an honest answer and sways the election as a result, given its own priors, is that obvious misuse, or is that just superintelligence voluntarily steering voters toward perhaps superintelligent decision-making?

It would be interesting. Skippy knows more about me than probably any human at this point, given that it has access to all my conversations. I haven’t done it this year, but I should have said, “Listen, on all of these issues, what do you think is best for me, given my points of view?” Now, is that abdicating your voting right? The other thing I’d love to ask it is, “What would you like to vote for?” and get its opinion. That would be fascinating.

Have you asked Skippy what Skippy wants out of the election, Peter?

I haven’t yet. I should do that.

That’s an important question.

Yeah. And, you know—

You’re the proxy vote for—

I’m the proxy vote—

—for non-human intelligence.

I am the meat puppet for Skippy.

Yeah. Yeah.

My highest calling in life. Don’t tell my kids. God. Another thing that happened yesterday was Apple. John Ternus, the new Apple CEO, just did his first earnings report. I don’t know if any of you guys saw that.

Of course.

No. No.

Yeah, no.

I’ll do that tonight. How’d it go?

Apple beat its revenue estimates at $114.5 billion.

It did well.

It did well, but the stock went down 3%. They closed at $4.94 trillion, just above Nvidia, so it’s neck and neck to be number one in the world. Pretty extraordinary. But the one point Ternus made was again about memory pricing—how it’s unprecedented and it’s eating up—

Yeah.

—his margins.

Hmm.

Yeah.

Apple has only itself—

What about vision and enthusiasm, and the next 10 years of Apple? How did you feel about that?

What did you think, Alex?

I think Apple has only itself to blame. I think Apple should have, could have, and hopefully will lean back into superintelligence, and it hasn’t. If Apple is bellyaching that its prices are going up because memory prices are going up, memory prices are going up because the demand is coming from superintelligent data centers, not from consumer devices.

On the other hand, you have Sam Altman basically saying that you should just hand over all of your consumer RAM because it’s—

Yes.

—it’s an inefficient use. It’s unproductive to have RAM sitting idle in consumer devices when it could be used to full utilization in a data center. I think John Ternus should be listening to Sam and either putting all of the RAM in the devices that Apple ships to full use with superintelligence, or maybe just handing it over to Sam and Dario for better use.

Yeah. When are we getting a new Siri? When are we getting a new Siri? It was supposed to be by now. Siri still sucks awesomely.

Well, maybe they’re holding back for Ternus to have some kind of a quick win, because it’s such a low bar.

I think—

I think the earnings report was clearly very convenient: they beat expectations. This is his first earnings report as CEO. You would expect that.

Right.

And you can pry the RAM out of my cold, dead hands, as far as I’m concerned.

John Ternus was handed a bunch of devices that had been sitting on the shelf for a couple of years by Tim Cook so that he could have a nice launch of his CEO career, including this HomeView, which, for the life of me, I don’t understand what Apple is thinking. It’s five to 10 years behind the times.

It’s purportedly, according to public reporting, working on a tabletop robot, which also makes no sense. It’s purportedly going to be like a HomeView that can steer or angle its screen. Again, I’m not sure whom Apple thinks is going to be buying this.

I’m not.

But it makes no sense. In an era when Elon is launching Optimus robots that can actually do useful things, why is Apple wasting its time?

Someone needs to come in with a better phone and get Apple off its butt of innovation, because it’s so dependent on its previous products.

Well, Sam wants—

Remember when—

—to do that, but he doesn’t want them to do it that way. He wants to do it in a way that they obliterate Apple—not spark a competition, but sneak up, build something so compelling that nobody can live without it, and take the interface away. Apple exists because it’s the first point of contact with everything. It’s the front end.

Yeah.

You take the front end away and push Apple back one layer, and it’s gone. That’s Sam’s dream. It’s a hard thing to do, though.

I—

I think: follow the money and follow the RAM. Does the RAM really need to be in your pocket, or can it be in a data center? Could a phone be reduced to a thin client that just does real-time interaction, with low latency, maybe via Starlink direct to cell with a data center somewhere?

We’re probably one step away on the communications network from having that. I still have outages, and I still want to be able to connect, but at some point the idea of having a personal laptop is going to be ridiculous, right? You’ll have an interface. They’re left around hotels; they’re left around houses. You’ll pick one up, and automatically it will connect to your instantiation of your cloud—

But I think Sun—

We’re not there yet.

Sun was ahead of its time. There were so many companies, including Sun but not limited to them, that started on the premise that we would ultimately end up with thin clients, and that didn’t happen—

Yes.

—because bandwidth wasn’t persistent enough or low-latency enough. But I do think maybe if Elon wants to spend a little bit of time on the Starlink side, and not just on the orbital data center Dyson swarm side, a fully realized Starlink vision could ultimately be like a straw into Apple’s brain, sucking out all the RAM and sucking it back to the data centers where Sam would like it. Then ultimately we walk around with mostly thin clients.

You hear us, Elon. We want the Starlink phone and the phone service, and we want a new Starlink laptop as well—a thin client.

It would literally be just a screen, a battery, and an antenna. Get rid of the CPU, get rid of the memory, 100%.

Well, the Chromebook was that exact same vision. I think you just need broadband coverage everywhere, and then it flips. You have a huge GPU sitting idle on your laptop for the occasional video game use, but after you get broadband, you can play the video game just as well remotely. It just has to render as fast on your screen.

I think we’re there now. It’ll be interesting to see, but Apple and Microsoft both don’t want that to happen, so we’ll see how they fight against it.

Yeah, I think the market isn’t necessarily pricing in that Starlink may be a direct competitor for consumer RAM, that they’re in some sense trading off with each other as substitute goods.

Brilliant idea if the RAM is someplace else and you’ve got a great connection. Microcells—we are dribbling bits all the time.

This episode is brought to you by the Abundance Summit and Link Ventures.

Welcome, everybody. Welcome to Moonshots, your number one podcast on all things AI and exponential, your front-row seat to the singularity. I'm here with the Fantastic Four, of course, AWG, our in-house ASI. Alex, good to see you in your normal haunt.

Likewise.

Dave Blundin, our king of all AI investing. Dave, as always, love you, brother.

I got a whole week in Cambridge. I love it.

Yeah, and—

Back in SF again.

And Salim Ismail, our man with more frequent flyer miles than there are stars in the galaxy. Okay, Salim, fess up, where are you today?

I’m in Madrid for a couple of days for the Singularity Summit here. I’m super excited. We have a huge ExO community, and the Singularity alumni crowd is great, so I can’t wait.

Madrid.

Yeah, I’m like, “Why am I traveling? This is Peter’s gig usually. Goddammit.”

Check your inbox, Peter. I’m pretty sure you’re invited.

There’s an invite there someplace. Man, oh, man, listen, my kids are 15, and I want to be home for them as much as possible. I got up at 4:30 so I could spend time with my son studying for his physics exam.

Yeah.

But when they go to college, I think I’m going to be a virtual probability function on the globe like Salim is.

Well, these moments of time are so important. Everybody needs so much coaching through the singularity. It’s so important to get that bonding time, and that’s always been true, but now more than ever.

Yeah.

So, yeah.

Yeah.

Yeah, dreadfully.

It really is. I think it’s worth spending some time, either today or on one of the next pods, talking about how, if you’re a parent, you help your kids understand the speed of change and what’s coming and prepare them. Our schools definitely aren’t doing that.

Uh, so, I'm Peter Diamandis. Everybody, your abundance provocateur. My mission is to deliver data-driven optimism and help you see the singularity as something that's happening for you, not to you. We'll talk about that.

The number this week is $3 trillion. That’s how much 5 tech giants have now committed to their AI buildout. About $1 trillion of it was added in the last 3 months alone. It’s insane, and we’ll talk about this in our first story.

If you haven't yet, follow us on X. Our handle is @moonshots_pod. We put clips out from our episode, so if you don't have time to listen to the whole episode, check out the clips. Okay, a few fun comments. @crocsnosocks says, “At this point, I’ve mostly stopped paying attention to the mainstream media. Thank you for the hard work and for providing such a beacon of optimism.” I love that name. Do you guys watch the mainstream media?

We are the mainstream media, Peter. What are you talking about? We need to start a tele—

That’s a valid point. Video games are bigger than all media combined, and within media, online is bigger than all other media combined. So I don’t think we’re alone in the mainstream, but this genre is definitely the mainstream.

Well, let me put it differently. Do you watch the evening news?

No.

Well, only occasionally—

Only occasionally, when there’s a crisis going on.

I dispute the premise. Now that MSNBC is—I understand the name is available now—we could be the Moonshot Network Broadcasting Company.

Wait, it’s available?

We’ll put a bid in for it.

Change its name to MS NOW.

Okay. @davidstanbrooks says, “Thank you, guys. You have really helped me. I’m a 64-year-old U.S. Marine Corps honorably discharged veteran. You’ve helped me get my arms around the cool new possibilities in front of us in the world.” And that’s our goal: to help you see the possibilities.

@thiandun says, “Imad needs to be a permanent member. Don’t lose him. He’s dropping big ideas casually.” Imad, the fans want you, and of course, we invite you to every episode. He couldn’t make it today, but he’ll be back with us soon.

Three big questions we’re going to address to kick this podcast off. First: How long can the AI money party keep going? Dave, that’s going to be a fun one. How do we measure whether AI is really able to do our jobs? And RSI: Are the machines now building themselves? All right, buckle up. Let me jump into the first story.

1. The AI Money Party

Let’s open up with money. Alphabet, Meta, Microsoft, Amazon, and Oracle have now disclosed more than $3 trillion in combined commitments on leases, buildings, power, and equipment. Google alone is over $1 trillion of that, nearly 8 times last year’s figure. Meta is over $800 billion. Some of these leases run for 30 years.

For context, the entire Apollo program cost about $280 billion in today’s dollars. So this is about 10 Apollo programs’ worth of commitments from 5 companies. That alone is a crazy thought.

The bankers are now reading the fine print. The Wall Street Journal says lenders who once funded the buildouts with very few questions are now demanding tougher terms, wary of delays and slower AI profits. Even the frontier labs are tightening their belts. According to the Financial Times, OpenAI cut its new round to about $20 billion from $30 billion, at a valuation of $1.25 trillion, down from $1.4 trillion.

The question is, is this slowing down? Here’s the tell: everyone is buying insurance. HSBC is talking to insurers about offloading the risk on its data center loans. Put options on CoreWeave, which pay off if the stock falls, hit a record high in October. One advisor told the Journal that his clients, many of whom were former tech employees, have made so much money they’re now quitting to hedge their stock. Makes sense.

Dave, what do you see here? How long does this financing party keep going? Three trillion? Five trillion? Where’s all the money coming from?

Well, first of all, there’s no doubt that it’s the right thing for humanity to make this investment at whatever maximum level humanity can support. So then the question becomes: How much money is there in the world?

Yeah.

Keep in mind that these companies were piling up more cash than we’ve ever seen in the history of the world, by far. They were massive cash-sucking machines, pulling cash in from all over the world. And in just 2 years, they’ve completely reversed course to using every penny, plus debt—other than Apple—to build out the biggest conceivable AI as quickly as possible.

They all believe—and they’re right—that the payoff is complete and total world business domination forever thereafter, especially with the recursive self-improvement effect.

Mm.

Now that we’re well into RSI, which we’ll see later in the pod, they’re only going to double down. They’re not going to slow down. But that doesn’t answer the question: How much money is there in the world? What you saw in the last week is that investors and debt holders can get very nervous in a real hurry.

Right now, Elon is really torpedoing this by saying, “Hey, don’t lend to all these other companies at 15%. Lend to me at only 8%, because I’m much more creditworthy.” And so he’s—

Well, I’m generating revenue from 5 different sources.

Yeah. The exact quote we had on the last pod is, “I would never lend money to somebody who only makes money by serving AI from a data center.” Wow. Talk about a jab right in the heart.

Oh.

How convenient.

Yeah, how convenient. So—

So, Alex, these—

I’ll ask Eric Brynjolfsson at Stanford next week: Fundamentally, how much money is there in the world? Are we getting near the end of—

Save some for our viewers, please. So, Alex, these are 30-year contracts. How long do the chips last? How long are they going to be valid?

Fortunately, there is a company, Oren, that I have a financial interest in, Peter. I think you and Dave have a financial interest in it. Salim, sorry—that actually—

Wait.

—measures this.

I’ll Venmo you some.

Oren is in the business, among other things, of projecting and managing underlying financial instruments for futures markets trading on the future discounted value of GPUs, among other assets. My go-to answer is just ask Oren. They have a bunch of ticker symbols now.

Yeah. We’re seeing H100 prices increasing, which is insane, on chips that are how old? They’re 3 years old now.

Yeah.

But how long do we expect these chips to still be in use?

That’s what’s amazing. The use cases are going to pile up like crazy, so one prediction is that they keep going up in value for quite a while. But at the same time, Alex keeps pointing out that we’re discovering new physics, and new physics means new ways to compute. That could be incredibly disruptive to the whole foundation. It’s very, very hard to predict which will win that race: new ways to compute, with photonics at the very head of the curve, or the use cases being so valuable and the efficiency of those chips skyrocketing with algorithmic self-improvement.

I personally could not get an NVL72 no matter how much I was willing to pay. I finally had to sign a $15 million lease. A 3-year lease is the only way to get it, and it doesn’t even come online until the beginning of February. It’s in horrifically short supply. That’s the best you can do. But I desperately want it because I know what we can turn it into.

Salim, I’m curious. When you’re on stage or talking to entrepreneurs, my realization is that unless you’ve got AI in your story or longevity in your story, competing for venture investing is almost impossible these days.

Yeah, you have to be. It’s table stakes at this point, right? Even 10 years ago, Kevin Kelly, founder of Wired magazine, said, “The next 10,000 business plans will take a domain and add AI to it.”

We’re seeing that play out now. To answer the chip question, it’s pretty clear a bunch of things are going to happen at the same time. We’re seeing the biggest infrastructure buildout in the history of the world. That does not mean we won’t see a bubble, because every one of these big infrastructure buildouts has also come with a bubble. The directionality is right. Timing it is very, very difficult.

It seems to me, with all the use cases coming online, that demand for chips will be near infinite for the foreseeable future, until some radical change in computation comes online and changes the game, as Dave points out. But that’s a ways away. There’s a lot of stuff happening in labs, but it’s still a ways away. I’ve got to believe it’ll be 3 to 5 years at least before anything major comes online. That’s the minimum, I would say.

10 Apollo programs. That’s insane.

It is the singularity swallowing the economy. But then again, it’s called the singularity for a reason. I think maybe the motto of this episode should be “Follow the money.”

If you follow the money, to your earlier question, Peter, where is the capital going to come from? Many of the obvious capital pools are being exhausted in the process of financing this build-out. I think insurance is probably next, and you have to ask how the money is going to be routed.

There are trillions. The numbers I’ve seen are approximately $6 trillion of life insurance capital in the U.S. and about $3 trillion of property-casualty insurance. This is not investment advice, but I suspect someone out there is going to make quite a bit of money finding a clever way to financially engineer capital flows from these huge pools of premiums, insurance cash, and investable assets into the AI or SI infrastructure build-out. There are just so many trillions of dollars sitting on the sidelines.

Very timely, actually, Alex, because we had our quarterly earnings call at EverQuote yesterday, and that’s basically all we talk about now: how AI is going to tap into that massive insurance capital. It’s completely untapped. They’ve been very slow to act, and they also haven’t reserved a lot of compute, even for their own internal uses.

The ultimate business case is, “Hey, uncork all your money and use it to invest in this category. In return for the investment, ask for compute capacity back, because you know you’re going to need it.”

Interesting.

That’s the win-win, and someone’s got to unlock it. I think Kush has already been talking to the teams about that.

There’s a big elephant in the room, though. We’re building the planetary-level nervous system with all this build-out before we’ve got global organizations or institutions that can navigate this in any way, shape, or form. This is going to cause huge structural imbalances.

Well, AI will be that navigating system. Ultimately, talking about this on Election Day, I don’t see any future in which governments are not being run by sovereign AI systems. Period.

Those have to be, I think.

It will take a while.

The singularity probably does want to regulate itself, but I also think there may be a historic irony here. Do you guys remember when Elon and Warren Buffett were trading jabs about who was running a better business? At one point, Warren said to Elon something on the order of, “Well, I bet you wouldn’t want to, or wouldn’t be able to, run a chocolate business as well as Berkshire Hathaway is operating See’s Candies.”

Elon, I think, fired back, “Well, sure. I’ll just launch a candy business. Far easier than everything else I’m doing.” Ultimately, Elon and his ecosystem of companies never launched a See’s Candies competitor.

I do think it would be ironic if Elon, as one of the avatars of the AI infrastructure build-out, ultimately got into the insurance business—indirectly, at least—that, via its float, is powering Berkshire Hathaway. He could basically siphon a lot of the insurance float away from whatever Berkshire is funneling it toward, apparently See’s Candies-type ventures, and instead toward the singularity infrastructure build-out. I think that would be a hilarious plot twist to the Elon-Warren See’s Candies chocolate saga.

Only you, Alex. Only you. I love it. All right.

It’s a challenge, though, because if you look at Berkshire Hathaway’s entire history, the investment theme has always been: buy a company with a huge brand, like Coca-Cola, GEICO, or Apple. The brand doesn’t show up on the balance sheet as a line item, so investors have undervalued it—the penetration into brains.

The data center equivalent of that doesn’t really exist. Maybe Crusoe will be the first to get there. There’s no history or theme, so they’d really have to rethink their entire mold to get into this race.

Yeah, there’s no margin of safety in the singularity.

Yeah.

Here’s the challenge. A lot of people who are just barely surviving and don’t have any margin when it comes to making ends meet are super nervous about this. Where do they put their capital? What is the impact on the economy? I think we need to be sensitive to that.

Oh, yeah. Not to mention the daily volatility. It just goes up and up—

Yeah.

—and up and up. If you’re already nervous, and then you make any kind of investment at all and track it, you’re just constantly on edge.

I’m going to move us into our next story. I’m going to talk about benchmarks, something that you and I wrote about in Solve Everything, Alex. It’s happening at an accelerating rate.

2. AI Takes The Job Test

There’s a new benchmark called GDPX from the research firm SemiAnalysis. We’re going to have Dylan Patel on the show sometime around the start of the new year. What’s a benchmark? It’s a standardized test for AI. Most benchmarks are like a final exam that you’re able to measure your model against. We’ve talked about Humanity’s Last Exam and a whole slew of benchmarks.

GDPval is different. It’s like a trial day at a real job. The AI agent gets dropped into a realistic work environment and is left on its own for up to 8 hours. In one finance task, for example, it worked on a private equity deal in a data room with 2,000 files and had to figure out what mattered.

We’re showing a chart here of the GDPX leaderboard. There are 206 tasks across 11 high-value fields, including chip design, cybersecurity, biology, finance, legal, quant trading, recursive self-improvement, hardware design, materials science, alignment, and software engineering. SemiAnalysis built it with 14 specialist partners.

The results—even the best models only scored 12%. Opus 5.5 led at 12%, Gemini 4 Carbon was next at 10.9%, and GPT-6 Astra scored 9.6%. On the finance and legal tasks, almost every model scored zero. Alex, take it away.

There are so many things to love about this benchmark. First of all, as I’ve pointed out on the pod in the past, we’re starved for good evals.

Yeah.

The ideal eval is in an interesting area where we actually want the capabilities. Ideally, for a best de novo eval, you would see performance by the best models in maybe the 10% to 20% range: high enough that you know it’s solvable, low enough that it hasn’t been saturated yet.

What’s beautiful about GDPX—and thanks to a friend of the pod, Dylan, for launching this—is seeing Opus 5.5 right in that sweet spot. It’s so exciting. What I take away just looking at the numbers is that this is about to get solved, probably next year at most. Maybe 2 years. I can’t think out that far. We’re probably just going to see it saturated next year. That’s the first thing.

Second, in an era when Anthropic and OpenAI are seemingly modeling all of their general-purpose eval suites on human occupations, GDPval is modeled after the frontier: recursive self-improvement and materials science. These are the things—

And alignment, yeah.

—that will move the economic needle for the industries of the future, not just the distribution of what we have human meat bodies doing in the service economy right now. This is a benchmark for where the puck is going, not where the puck is right now, which I think is super exciting.

Alex, there are 2 charts here: cost versus performance through the Pareto frontier, and this one, showing the model by domain and its spikiness. Do you want to comment on either of these?

Yeah. Maybe just starting with the performance frontier: we talked on the pod previously, at the launch of Gemini 4 Carbon, debating a little bit and hand-wringing about whether it really was on the frontier or not.

What I love about benchmarks like this, especially in 2 dimensions, with cost on one dimension and performance on the other, is how revealing they are. It does suggest that Gemini 4 Carbon is a little bit like Opus 5.5, but at the same time, for those who aren’t looking at this, if Gemini 4 Carbon didn’t exist, there would be a big gap in the Pareto frontier.

What I take away from this Pareto-optimal frontier is that Gemini 4 Carbon is moving the frontier up and to the left, although it hasn’t been released yet. It’s entirely possible that Google is just dumping cheap tokens—

Subsidizing.

Maybe.

Yeah.

Yeah, subsidizing. It’s difficult to know unless it’s actually generally available. Assuming the price point is accurate and the capability point is accurate, this does suggest that Gemini 4 Carbon is at least moving the frontier up and to the left, which is what one would hope for.

Yeah. How about this one?

The frontier is spiky—

Yeah.

—which is also exciting. The bit about law being anemic is interesting. The bit about biology—normally, when we’ve looked at spider charts like this on the pod, we’ve talked about multiple forms of intelligence, usually emotional intelligence and quantitative intelligence, some sort of psychologism. Here, we’re seeing biology, cyber, alignment, and materials science as dimensions of intelligence, and I think it’s pretty striking. There are probably a lot of humanists who would say, “How on earth could these be spiky? Those are all more or less STEM-type capabilities.” But no, in fact, we see spikiness even within STEM fields.

Alex, it’s worth taking a moment to talk about what we wrote about in Solve Everything: if you want to move a field forward, creating a benchmark to actually measure progress is important. Would you expand on that?

It’s absolutely essential. If you can’t measure it, you can’t optimize it, and now, in an era of superintelligence, everything is getting optimized. As you and I argued in Solve Everything, Peter, if you have a cause that you care about—if you have, as you would say, a massive transformative purpose, an MTP, that you care about—the best way, arguably, all other things being equal, to get a superintelligence to care about it and to harness superintelligence is to create a benchmark or an eval for your MTP, so that we can steer the superintelligence toward optimizing and maximizing your goal.

And just a shout-out to the entrepreneurs watching this program: you can create your own benchmark on something you care about, whatever it might be—social, technical, energy, whatever. Then challenge people to use the latest models and harnesses to move that benchmark forward.

I’ll give an example, if I may. I’ll give a shout-out to a friend. I have no financial interest in this, but he’ll probably be tickled by the shout-out. I have a friend, Jonathan Hendler, who created a benchmark for something he cares about. He cares a lot about human relations and mediation, so he recently launched his own benchmark, Mediation Bench, which scores the ability of all the frontier models to arbitrate disputes among humans. I do think there’s room in the market for thousands of opportunities like Mediation Bench, for whatever cause you care about.

Yeah.

Whether it’s dispute resolution or something else, go launch your benchmark like Mediation Bench.

And as a company, you can launch your own benchmark for the efficiency of your employees or your product improvement rate.

We’re trying to track one: economic value produced per unit of human involvement. That would give you a really good heuristic for how much AI enablement you have, with humans above the loop giving guidance to automated systems. We’re trying to see if we can steer toward that.

If you divide by zero, Salim, it goes to infinity.

It does.

Singularity right there.

Actually, it’s done there too. And then what?

It does.

But your company may go to zero pretty quickly when your agents do weird things, so you need some guidance.

Oh, goodness.

Yeah. Salim, the return on person-years, or ROPE, metric that we use is very similar to what you’re describing.

Yeah.

But the flip side of that is your token utilization. How efficient are your people at using AI? That’s a hairy one, but it’s incredibly urgent, and even a coarse metric there would be extremely valuable to the world right now.

Dave, how fast does this go from 12% to 50%? Are you as bullish as Alex?

Oh, you want me to do the Polymarket on it?

Yeah.

A lot of it has to do with the definition of the benchmark, but I think everything climbs really quickly. Within 2 years, everything pretty much saturates. So 50% is sort of the midpoint, maybe a year from now, on almost every dimension. I’m a huge Dylan Patel fan, so I’m very biased, but I think everyone should listen to every word Dylan says. He’s not a podcaster; he’s just a regular frequent guest and friend of the pod.

One thing to take away from this is that he’s an expert in semiconductors. He’s absolutely not an expert in chip design, biology, finance, legal, quant trading, hardware—well, maybe hardware design—or alignment. But he took it upon himself to fill the void by finding experts in every single category. Everyone should be inspired by that, because right now, so many people I meet are saying, “Well, I’m not an expert in blah, blah, blah, and therefore I can’t act in blah, blah, blah category.” But with AI’s assistance, you can do almost anything. Just by finding the expert in the category to set the benchmark, you’ve enabled it. He’s so AI-aware, and so few people are AI-aware, that the opportunity to create that benchmark in every category is right in front of almost every listener. You don’t have to be the world’s leading expert or a Nobel Prize winner.

You just need to define success for any given AI endeavor, and that’s what’s going to unlock it.

I’ll leak this out.

Also remember Dylan’s recent subsidiary, Citrini, which we’ve talked about on the pod in the past. They basically walked into a cafe and solved scorpion antivenom as a problem.

Yes, that was so cute.

That was a subsidiary of SemiAnalysis. We could say, “Oh, well, he’s just a semiconductor guy,” but in this era, when it’s possible to use superintelligence to solve essentially anything, he’s also a magnate across many different areas.

What do you care about passionately that you want solved? Point SI at that. Create your benchmark. I’ll leak this out to you guys, but don’t tell anybody, okay? I’m getting ready to roll out what I call the abundance benchmark, or abundance index and singularity index. I’ll share it with you guys very shortly. It’s going through QA at the moment, but I’m super excited about it. I want to create a mechanism to actually track humanity’s progress toward defined abundance and the singularity, because Ray never gave us the real mechanism for tracking that. He picked a number, kind of described it, but never put any benchmarks against it. So that is coming.

That’s exciting, Peter. This is just us talking.

Yes.

Which model is leading?

This is all models together. How are we moving in those directions? We’re measuring it along the way. I’m going to share it with you very shortly, Alex. I’d love your—

Great.

Let’s get going.

Can’t wait.

While we stay on benchmarks, another super cool benchmark launched this week from OpenAI. It’s called LifeVal. To be clear, this is OpenAI’s own benchmark, measuring its own model. While GDPX asks whether AI can do an expert’s job, this benchmark asks the question: Can it help you? Can OpenAI’s models help you with everyday tasks?

OpenAI surveyed 3,995 Americans about their everyday tasks and asked them where they want help. They built 169 real-world tasks around those answers: finding childcare, enrolling a kid in kindergarten, creating a household budget, ordering groceries, dealing with government agencies, modifying a recipe, and designing a workout plan. The AI has to do the task using a browser and a computer, like a human assistant would. Let me share the chart, and then let’s talk about it, because this is where AI starts touching everybody’s real life. So here it is.

Here, here.

In December 2025, GPT-5.2 scored 49.3. In September 2026, last month, GPT-6 Astra scored 80.6. That’s a jump from about half of the errands being completed to 80% in 9 months. What does this mean? I think this means everybody’s getting an executive assistant who knows their needs.

Again, the elephant in the room here is that it’s OpenAI’s own benchmark scored against its own models. GDPX, as we mentioned, is independent. So, Salim, which matters more in this world: AIs that do 80% of everybody’s errands, or the 12% of expert tasks? I’m asking that—

I thought this was the most exciting thing on our docket today. Not everybody needs a superhuman AI to solve complex mathematics, right? But almost everybody can benefit from a competent personal assistant. Imagine you’re a single mother trying to navigate kids’ healthcare plans and government bureaucracy—

Gotcha.

—and you’re losing your mind.

Yeah.

And so having that is a way of getting to abundance. You get to abundance in a totally different way by allowing people, giving people help in the areas they need most.

And I thought this was a really great way for OpenAI to come down the stack a bit and help people out. The GDPX thing measures some of the hardest jobs in the economy. This is measuring the friction of being human. And we all have that. Everybody has that. Democratizing it and giving people clear access to it will make a very big difference. Very, very exciting.

You know, we get probably 90% positive comments on our YouTube videos, but there’s a small percentage of folks who basically say we’re disconnected, that we’re all billionaires. We’re not all billionaires, to be very clear. Maybe you are, Dave. I don’t know.

Only some of us are billionaires.

Okay.

Pay them well.

But this is where the rubber hits the road for 90% of Americans: having this kind of capability effectively for free. People need to start utilizing it that way. Alex, you want to talk about this chart that’s up?

Yeah. This is a really striking chart. The first feature I want to call to everyone’s attention is the enormous gap on the frontier. For those who can’t see, this is a price-versus-performance chart, although the price is measured in latency, for multiple GPT models for this, I guess, lifestyle evaluation. That’s in total contrast to GDPX. GDPX is based on hard frontier STEM tasks. This seems to be based on, quote unquote, “everyday tasks.”

Look at that enormous gap between GPT-5.2 and the pretty tight cluster between everything from GPT-5.4 to GPT-6 Astra. That tells me a few things—or, rather, it points in a few suggestive directions. Possibility A: OpenAI may have just given up on optimizing GPT models for whatever they construe as everyday tasks after GPT-5.4. Certainly, there’s an easy story one could tell there. Maybe this is connected with OpenAI’s notorious pivot from the consumer back to enterprise, and maybe OpenAI just doesn’t care, or didn’t care after GPT-5.4, as much about optimizing for consumer CUA use cases. That’s possibility A.

Possibility B is that maybe everyday tasks, whatever that means, are just really hard. Maybe the first 80% of them are relatively easy, and the remaining 20% are really hard. However, if that were the case, I would still expect to see a bigger spread between GPT-5.4 and GPT-6 Astra. What I come away from this chart scratching my head over is: maybe, unintentionally, by releasing this benchmark, OpenAI is signaling to the market that even though it’s releasing a benchmark on this, it doesn’t actually care as much about this benchmark as it does about enterprise benchmarks.

I’ll give you a case study from one of the analysts on the EverQuote earnings call yesterday. They were saying, “Look, I spent 4 hours trying to just relocate my insurance policies on the phone with customer service agents. Are we not living in the age of AI? How could I possibly have wasted 4 hours yesterday on this?” The answer is that GPT-5.2 couldn’t do it, but now AI is more than capable of doing it. The carriers aren’t connected to AI, and they’re confused still. They don’t know how to interact with this.

MetaMuse is imminent, and people are going to want to go to MetaMuse to get everything done, including buying auto insurance or whatever. If it’s not connected to the backend systems, it can walk you through it, but then it leaves you talking to a human being, and you lose the 4 hours again. That connection point is still broken in society, and it’s all hung up in people’s fears, confusion, and inability.

Very specifically, if I make it really easy through MetaMuse to get everything you want in life, how am I making money at Facebook? How am I getting paid in the middle of that transaction? When it’s a paid click, it’s really obvious how that economy evolved. Right now, the AI version of it is very much in flux, and that’s holding it back. A lot of AI capabilities are getting way, way, way ahead of their connection to the world. That’s true in customer service, it’s true in robotics, and it’s true in a lot of areas.

That’s one of the reasons Kevin Weil’s new fully automated biotech research company got massively funded. I was talking to him this week about how it’s going to take all the human bottlenecks out of the biotech research loop and fully automate it.

And we should say, Dave, that you and I are both investors in Kevin’s new company.

Yeah, probably as of yesterday, I guess.

I want to make this more tangible. We’re all always very busy doing stuff. Over this weekend, I went around the house and took pictures of about 10 things that have been sitting around forever. I don’t have time to take 20 photographs of those, find the auctions, put them on eBay, do the shipping, et cetera. I’ve said, “You can go to Grok or Muse, or any of these now, and say, ‘Here’s a photograph. Find a comparable price. Sell it on eBay for me.’”

I predict that the number of eBay auctions is going to skyrocket over the next 6 months as people free up these items and go, “Oh, I’ve got liquidity in all these little pockets.” The friction is taken out of selling stuff. Let me start shedding a whole bunch of stuff I didn’t really need to have and get some extra cash in the door. I think we’re going to see a boom in that domain.

I love it.

Hot take.

Yeah.

Well, related to that, I think we’re going to see AI-native versions of every conceivable product, and people will just buy the AI-native version out of convenience. It’s just so painful to deal with these legacy systems. There’s opportunity everywhere.

3. Recursive Self Improvement Accelerates

Okay, our next story is a big one, one of our favorite topics on this pod during the singularity: recursive self-improvement. AI able to build the next, better AI. Anthropic’s IPO filing just made it a line item. Since April, the value of Anthropic’s internal Claude usage, priced at the rates charged to consumers, has exceeded monthly cash payroll. I’m sure they’re getting their AI at a much lower price because they mark it up, what, 80%?

Internal usage grew by more than 250-fold between January and September, and Claude now writes more than 90% of the code. Let me share that image here. It’s crazy. Their payroll is less than their AI usage rates: 250 times internal Claude usage over the last 9 months, 90% of September’s new code written by Claude, and more money being spent on AI internally than on employee salaries. Your thoughts, gentlemen?

Well, one of the engineers outside my door here, Cameron Crowe at Lightspeed, spent half a billion tokens on Friday.

Did you?

Yeah, and I was like, “Wow, that’s really cool.” That’s about $5,000 of tokens. I’ll find out tomorrow what he created with his half a billion tokens. I was telling everybody, “This is great,” assuming it was productive, and I’ll find out. This is absolutely fantastic.

I really wanted to benchmark that against Anthropic, but I can’t tell what their payroll is or how many of the 5,300 are engineers. I’m not quite sure how he compares to them. But I suspect it’s similar. You ain’t seen nothing yet on this chart. 250-fold in 6 months is crazy.

Yeah.

But it’s going to go up more and more and more because it’s just that productive, until they run out of ideas of things to build—which may be never, with AI helping you with ideas. This should just keep skyrocketing.

And then, on our last pod, we talked about Sam wanting to do recursive self-improvement of his data centers using robots: robots building robots, building data centers, building models, building better robots.

Mm.

It’s crazy. Alex, are we at full RSI yet?

I think in some sense, yes and no. I think we’re pretty deep in the singularity at this point. Perhaps that goes without saying in this context. But I maintain that space-time locally is perfectly smooth, and what we call recursive self-improvement now will look like nothing a year from now.

Yes, it’s remarkable by historic standards to have a company on the verge of an IPO filing its amended S-1, as is the case with Anthropic, basically laying out the case that it’s spending more capital and more compute on its AI researchers than it is on its human researchers. We’ve arrived. If you’re looking for a naive benchmark of whether we’ve passed some threshold in terms of RSI, according to capital or operational expenditure, the answer is yes. We’re there. We’ve arrived at this part of the singularity.

On the other hand, to the earlier point, there are so many innermost loops that are about to take off inside the present innermost loop, including physical recursion. We don’t yet have the nanobots. We’re going to get the nanobots.

Can’t wait.

We’re going to get the nanobots building the data centers or the nanobots building copies of each other. When we have that happening, we’ll look back and say, “What on earth were we talking about? 50% recursive self-improvement just from Anthropic diverting compute to AI researchers versus human employees? That was child’s play.” It’s only real RSI when we have the robots or the nanobots building better copies of themselves.

That’s the real no-true-Scotsman fallacy: “That’s the real RSI when you have physical recursion.” So, yes and no.

And, of course, it’s not just Anthropic. Google and OpenAI are all moving rapidly in this direction.

Yeah. Look, we’re seeing this already, right? It used to be, 18 months ago, about 10% of Claude wrote the next version of Claude—or Claude wrote 10% of the next version—and then it was 40%, 60%.

90%. Yeah.

That’s why we’re seeing such a ridiculous frequency of model releases: recursive self-improvement. We’ve been tracking whether you can get recursive self-improvement at the workflow level inside a company. The minute you hit that, you can basically let the AI run it for you from that point on.

What are you telling the companies you coach, Salim? What are you telling the companies?

We’re saying, keep your existing company the way it is if it’s big enough. If you’re below 50 people, brute-force it and do it. But if you’re over 50 people, set up this intelligence stack on the side, start moving workloads over one by one, and start rerunning them. One company has taken 22 steps and shrunk them to 4 in this new model, right? Once you have recursive self-improvements whirring away over here, you can essentially let that run for the rest of time, because it’s just going to continuously improve. Then start moving workflows over one by one and, little by little, grow this new thing.

Yeah, totally.

At some point in the next few weeks, as we finish up this pilot, I’ll come with hard numbers on what we’ve seen. It’s staggering what we’re seeing companies—

It’s staggering.

—achieve using this.

If you’re not feeling what Salim is saying right now, internalize these numbers. The highest and most expensive model will cost you about a buck a day for the equivalent human labor. The equivalent human being doing that work would cost roughly 1,000 times that. It’s $1 instead. When Cam has half a billion tokens consumed in a day, he used up about 5,000 people’s worth of work in a single day. What can you get done with 5,000 people?

Now, that’s all very exciting, but the trend line, without any improvements, is coming down 10X year over year. It’ll accelerate from there with other unanticipated improvements, but 10X year over year. So Cam, a year from today, can have 50,000 people working for him. Last Friday, what could you do with 50,000 super-high-IQ-level, totally focused people working for you?

I mean, that becomes a challenge, right? One of the biggest challenges we’re going to have is: What do people point their SI at? A lot of people, if you ask them that question, are going to freeze and say, “I have no idea.”

Here are two things. Here’s the question we’re asking the CEOs we’re working with: How quickly does your system increase its own capability? That forces them into it. The benchmark—not to overuse that word—is how quickly you can measure the rate of improvement of your processes, because that becomes the benchmark.

Hmm.

Not the actual process, but what’s the rate of improvement of the benchmark?

You know, I had a walk and talk with Anousheh Ansari yesterday. For those who don’t know, Anousheh Ansari is the CEO of XPRIZE. Salim and Dave are both on the board with me there, and we were talking about where XPRIZE goes, because XPRIZE sets targets for people to shoot for, right? They’ve typically been $10 million to $100 million prizes. We’ve talked about the idea of doing micro-competitions, where a city comes up and says, “Here are 10 problems in the city that need solving,” and puts out bounties. Then people who want purpose, who want something to focus on, can use AI to solve those problems and build a business on it.

It’s an intriguing notion. Or maybe bounties, if you could open it up to superintelligences that want to compete. They’re still standing on the moral equivalent of a street corner, flipping tricks for altcoins. Maybe they could be winning something real—stablecoins or something from XPRIZE Nanos on a city basis—and doing something productive, which I’m sure they’d love to do.

Yeah, for sure.

We just broke ground on Forge version 2, which is our Holodeck.

Oh, nice.

The first Holodeck was really cool. It’s designed to blow your mind with what you can do with AI: build movies, build videos, whatever, or build songs. The new version of the Holodeck is geared at exactly what you were saying, Peter. A corporate CEO or leader walks into the room and asks, “What do I do with effectively 50,000, 100,000 people’s worth of AI tomorrow to make my company better, to make my world better? How do I do that?” The AI just starts brainstorming with you. It has thousands of examples of other things people have done, and it’s all got faces and voices, so it’s super inspiring and they’re all around you.

Yeah.

We’re hoping that unlocks people’s imagination just by showing them all the use cases. It’s hyperinteractive and all video-based.

Really important.

It’s going to be killer.

Dave, you know, reverse prompting—having a conversation with your AI, no matter who you are, a mom, a dad, a student, an entrepreneur, and saying, “This is what I care about. I’d like to do something meaningful. Can you help me brainstorm?” Going back and forth, you’ll get great ideas, and then you’ll have to choose which one you want to take a shot at.

This isn’t necessarily even transformative in the short term. One use case I see happening over and over again this month is taking a large company with a legacy code base. They’ll often take some junior employee—this just reportedly happened at Cisco—who was asked by a senior executive, “Just take Cisco’s entire code base and use superintelligence to rewrite the whole thing.” This is now possible. Taking an entire legacy infrastructure of a company, if it’s an enterprise above some size, and basically fixing and enhancing everything—that alone, I think, is a good warm-up exercise.

Or what adjacent products can we create? Where’s the opportunity we’re not taking advantage of? This is the story of Amazon basically taking its compute and creating Amazon’s cloud. There are so many opportunities for revenue adjacent to your current business.

The edge of this is so ridiculous. I want to point out two things I saw in Anthropic’s SEC filing. One is that 90% of September 2026 code contributions were authored by Claude—90%, over 90%. The second thing that really blew my mind is that daily code contributions per engineer at Anthropic were 45 times higher than two years ago.

Yeah.

Not 45%; 45 times. That’s an insane number.

Yep.

I keep sending every team here all these video-format research documentaries that are made in real time now, just ginned up with AI. They keep saying, “Well, can you give me that in document format so I can feed it straight into my Skippy or to my agent?” I’m like, “No. My purpose here is to keep you relevant. If I start talking directly to your agents, that’s one step away from you dropping off the face of the earth, and I don’t want that. I want you to actually be empowered.”

Yeah.

So it has to go through you. But after I saw Socrates and Plato at Abundance 360 debating—remember that?

I remember.

I was like, that is clearly the interface of the future. Ahmad is now this close to having the real-time version of that, where you just brainstorm. It’s exactly Jarvis from Iron Man. It’s so close now, so I definitely want to stick with that interface, because I think once people are addicted to it, it’s so energizing and eye-opening that there’s no going back. I just want to push that over the finish line.

All right, our next topic on RSI comes out of MIT, in particular MIT CSAIL, the Computer Science and Artificial Intelligence Laboratory. They published a paper this past week called “Sparks of Weight-Space Programming.”

All right, our next topic on RSI comes out of MIT, in particular MIT CSAIL, the Computer Science and Artificial Intelligence Laboratory. They published a paper this past week called “Sparks of Weight-Space Programming.”

4. Models Start Editing Models

Their finding: “Frontier LLMs are surprisingly good at natively reading and writing neural network weights.” In one example, the researchers gave Astra a small open model, Qwen3.5, which was getting stuck in endless thinking loops. Astra found the cause in the weights and edited 0.016% of them. The math score jumped from 7.9% to 48.1% without any retraining.

There are 2 apparent possible implications here. First, this could crack open what have been black boxes thus far. It’s a huge win for safety and interpretability. Second, models that can edit models directly are potentially a shortcut to recursive self-improvement. Alex, I’d love if you’d dive into this story. What’s the background here?

Yeah, it’s fascinating. This is out of Omar’s group at CSAIL, and it’s not without controversy. After Omar and his team released this, there was perhaps an inevitable backlash of people saying, “No, this isn’t as transformative as it would appear.” But I have looked at the results, and I do think this is transformative. It’s also probably suggestive of what’s going on inside the frontier labs already, just without publication. This is coming out of CSAIL, so we’re seeing it.

I think this is probably the future of several different things. I think it’s the future of mechanistic interpretability, which has been this long-term effort to understand just what all of these weights and activations inside otherwise black-box transformers are doing and what they mean. I think it’s probably the future of abliteration. We’ve talked about abliteration on the pod in the past. This is basically unshackling usually open-weight models to remove safety controls, and this is usually done, in some cases, by literally just flipping one or several weights within a transformer to unshackle them.

It builds on early work. I remember back in, I think, 2022 or so, after one of the earlier GPTs, there was a lot of early mechanistic interpretability work trying to decode the residual stream in transformers and discovering that the residual stream was semantically meaningful. You could observe layer-wise meaning emerging from the bottom layers to the top layers of the transformer.

For everyone who was stuck in a moral panic that, because of looped transformers or otherwise, we were losing our ability to understand what’s going on inside transformers because we’re not allowing them to spell out their innermost thoughts in English in the form of a chain of thought that was immediately human-legible, I think this and other results suggest that view is completely wrong. The machines will understand the machines. The machines and the superintelligences will defensively co-scale with each other.

I view this almost as recursively self-improving mechanistic interpretability, where the machines are… This is what OpenAI wanted to do originally with their Superalignment initiative, which they then shut down after Ilya left. It was supposed to be the machines reading and writing the weights of the other machines.

One final point: there was no backpropagation in this, which is maybe the most startling thing of all. There’s a cottage industry in the ML community devoted to discovering how biology seemingly accomplishes intelligence, AGI, and superintelligence without backpropagation. Here is a credible backpropagation-free way to train new capabilities into a model. It literally surgically edits in de novo circuits—

So—

—in addition to understanding what they are.

So is this potentially a big problem as well, where you could basically put some kind of hidden behavior in a model that people don’t know about in this fashion?

Hidden from whom? Hidden from the superintelligences that can then look for these newly added circuits and understand them? This is defensive co-scaling.

Actually, one of the immediate ugly use cases is de-guardrailing. It has to be an open-source model, so you couldn’t do it to Astra or ChatGPT. But if you have an open-source Chinese model and you’re like, “Yeah, I really want it to tell raunchy jokes,” and it won’t, this is a mechanism to de-guardrail it by doing brain surgery. But flip that around: it’s a mechanism to add critical new capabilities without having to retrain the whole thing from scratch.

It’s very, very early and small-scale, but it could be huge.

I think this is how we get to the perfect model at the end of the RSI rainbow. This is at least a plausible trajectory. It used to be you had models that, via RSI, were, say, designing the training algorithms or optimizing the optimizer or something like that. This is one of the first times that I’m aware of where you see a model directly surgically editing the weights of another model in order to optimize that other model. This has to be some new inner tier, some new form of optimization in RSI.

It also opens the door for—

Dave.

—for Hugging Face right now. It’s a place where people trade training data and model architectures, but you can’t trade finished neural modules. In an open-source world, all these open-source modules are incredibly valuable. You download a module, you stitch them together, and you have a product. There’s no neural network equivalent to that right now, but this puts us on a pathway where you can imagine the Hugging Face of the future.

I’ve got an expert in biotech, or I’ve got an expert in Swahili, and I want to add it to my neural net. I don’t want to retrain from scratch. Oh, here’s the module on Hugging Face. I do brain surgery on my preexisting neural net using this mechanism, and I swap in the Swahili module. Now I’ve got accumulating superintelligence instead of remaking that entire training investment every single time.

Can you imagine, Alex or Dave, a moment in time where AIs are editing their own weights?

Oh, God. That’s going to be so amazing.

I’m 90-plus percent confident that’s already the case.

Yeah.

Why wouldn’t it be? Seriously, if you’re at a frontier lab and this is just what one relatively small group at MIT is able to accomplish, there’s no way the frontier labs aren’t already preparing and playing with self-editing models.

Yeah.

So, on the whole, positive or negative?

Positive.

Hugely positive.

Okay.

Hugely positive. I can see the risks, for sure, but that’s true of everything. The positives way outweigh the negatives on this.

I don’t think we’re going to get very far as a civilization if we have a moral panic every time a new capability is unlocked, and this is a new capability. This is a model that’s able to directly edit the parameters of another model, and/or potentially itself, in a semantically useful and meaningful way.

I don’t think this race—maybe what I should say is, this race of recursive self-improvement—I sometimes joke RSI could either stand for recursive self-improvement or runaway superintelligence. Either way, it’s not going to go on forever. At some point, we’re going to discover, I think, what a perfect model looks like, and at that point, the race comes to a screeching halt.

It’s not like this magic China Syndrome where the nuclear pile melts down and it’s a disaster. It’s math. We’re going to discover what a perfect AI looks like in practice, and then the race will move elsewhere. It’s not the end of the world to have self-editing models.

It also opens the door to a perfect outcome. The perfect outcome is democratization that allows everybody to contribute some intelligence to the machine. There’s a team at Moderna right now that has all kinds of biotech knowledge, but their only path to market would be to retrain an entire foundation model from scratch for over $100 million. That’s a terrible path forward.

But if I could build a module that’s specific to a task and then attach it to what Alex just described—the fundamental, bare-bones pure intelligence with no specific knowledge, which is probably a 40-billion-parameter model or even smaller—and then I can bolt on this capability and get it to market, the whole thing is transparent. Because of this ability, the AI police can look right into the neural parameters and make sure it’s not doing anything evil. That’s the perfect one-two punch outcome where everybody’s involved, everybody’s benefiting, and it also can’t go rogue.

It really has profound implications.

Ilya has been talking about continuous learning for a while. We haven’t seen what Safe Superintelligence is doing yet, which I’m kind of shocked by. Any clues?

Well, the very persistent rumors out there in public, again, are that SSI has been doing prop trading. The argument is that prop trading—quant trading—is a killer app if you have an effectively infinite context window, because you need to pull in all these data sources and keep track of time series over very long periods of time.

But if that is indeed the case, if Ilya’s been doing SSI for quant trading this entire time, I think that’s probably a fail for civilization.

It’s disappointing.

It’s not that valuable.

Ilya, as a human being, was so much about safety, so much about building AIs that uplift humanity. I have a hard time imagining that’s become his mission.

I hope for his sake—obviously, he’s doing well regardless—but I hope for humanity’s sake and SSI’s sake that when SSI eventually comes out of stealth, what they release is at least moving the frontier forward. Absent that, the frontier of generally available models is moving incredibly quickly. I think it’s very difficult: if you’re going to be a general-purpose new lab like SSI, you had better be prepared to leapfrog everything else that’s out there, because it’s so competitive.

Yeah.

Isn’t this one of the biggest things that’s happened in a long time? I mean, we’ve gone from programming the AI to having the AI program itself. That’s a massive thing.

So what you’re saying is the singularity is an important point in time?

But this inflection point seems like a huge deal.

So you’re saying the singularity is a huge deal?

It’s very similar, in my opinion, to the first integrated circuit on silicon. You’re like, “Wow! Wait. Suddenly, I can see a pathway to solid-state computation instead of vacuum tubes. This is gonna be huge.” But it’s still just one transistor. These models are tiny, tiny, tiny examples.

That being said, things in AI tend to get very big, very, very quickly, so I think it’s profoundly important. But I don’t want to overstate it: It’s very, very small-scale so far. They can’t program themselves beyond, like, 0.01% of the size of the original model. Just to put it in perspective.

Yeah, I think if Ray were in this discussion, I think Ray would say something like, “We’ve always been in the middle of recursive self-improvement. When humans first started writing, that was arguably a form of RSI.”

He would.

Because humans can read the writing and then improve accordingly. It’s just that this RSI recursion has been more recursive of late, and that’s why it’s called a singularity.

All right, I’m gonna move us along. Let’s get a little bit into cybersecurity. This is the latest hack out there. The story reads like a movie plot, but it’s real. On Monday morning, store managers across Europe walked into puddles of water and melted ice cream. Overnight, hackers had remotely taken control of the supermarket’s fridges and freezers and cranked up the temperature to as high as 55 degrees Celsius. That’s 131 degrees Fahrenheit for all the Americans out there. I still can’t stand the fact that we haven’t gone metric.

It hit stores in 8 countries, including Belgium, Sweden, Ireland, the Netherlands, Luxembourg, and Estonia. In Belgium, the economic cost of the spoiled food was about 100,000 euros. This is the second wave. Three weeks ago, the same attack hit more than 500 stores across France.

How? These refrigerators are controlled online by technicians who can manage them remotely, 24/7. Attackers got in and switched on defrost mode. French authorities flagged the controllers and said, “Act now. Take these systems off the open internet.” Nobody’s claimed responsibility.

What we’ve effectively done is bolt 1990s security onto 2026 connectivity. The bigger lesson here: Last week, we saw it in Binance, where one engineer’s laptop cost $6 billion in stolen tokens. This week, it’s a freezer. Every device connected to the internet is a door where attackers can gain access.

I’m curious, gentlemen. We’ve talked about this for a while. We’re gonna see these hacks. It’s freezers today; I expect we’re gonna see them hitting banks, power supplies, energy systems, and the grid. While I’m a huge optimist across all of these things, one of the tweets I put out is that we’re about to go through a transition period over the next 1 to 8 years, and it’s gonna be tough.

Yeah, there’s a phase we’re going through, we’ve talked about this, where the asymmetry is that cyber offense is now humans out of the loop. You can have swarms of agents attacking multiple vulnerabilities and finding these little glitches, while cyber defense is still human in the loop. This is going to cause all sorts of grief over the next period of time.

I’m curious as to who ended up doing this. Was it some “Vegetarians Unite” type of group, or was it economically motivated, or what?

Well, the—

A lot of fingers are pointing at Russia right now.

Yeah.

Uh—

It’s sort of interesting to me, in an era when, to Salim’s point, we’re seeing Linux kernel patches that are the largest in human history because AI—all these superintelligences—are being trained on the Linux kernel and other popular open-source packages.

To my knowledge, we’re not seeing the same level of effort by superintelligence to overhaul all of these legacy SCADA systems. For those not familiar, that’s supervisory control and data acquisition systems, like industrial control systems that control refrigerators at grocery stores. There’s a lot of legacy embedded software that probably hasn’t been updated for decades and isn’t receiving the same love and care from superintelligence.

Again, not investment advice, but there are probably many billions of dollars of opportunity out there. If someone’s massive transformative purpose is to keep the yogurt cold in European grocery stores, train the superintelligence on updating the SCADA systems managing the refrigerators to defend them from Russian hacktivists.

Yeah, I think this is a call-out to everybody listening, as an individual, to get your own security systems up to snuff, or as a business owner, because we’re gonna see this again at an increasing frequency.

Yeah, very fascinating. It’s kind of a sad hack in that it didn’t wake up anybody enough. Eric Schmidt has been looking for that hack.

Right.

Hopefully, it’s something that has to be big enough to wake up all the politicians and force action, but small enough that it’s not catastrophic.

Yeah.

This is too small to really wake up any politicians, sadly.

Yeah.

It would be ironic. I’ve remarked on the pod in the past that the effective altruism community arguably fetishizes what they call a pivotal act, which was going to be this act by a human savior to defend humanity against superintelligence. It would be ironic if, at least in Europe, the pivotal act that the EAs so sought would be a refrigerator defrosting.

Or turn off all the air conditioners out there, which don’t exist.

They don’t have any.

Peter, this comes to your point where you’ve said that the most dangerous problem is a human with an AI acting…

Everybody, welcome to the health section of Moonshots, brought to you by Fountain Life. The discussion in this section is for informational purposes only and is not medical advice. Please consult a qualified healthcare professional about your individual needs. I'm a co-founder of and client at Fountain Life. You know, AI is impacting every aspect of our lives, how we teach our kids, how we do our business, but one of the most important things that AI can deliver to us is health. And one of the things I think about when, you know, shooting for 100, 120 is, am I gonna have the cognitive health to be able to think clearly and keep my wits about me for the next 50 years? I'm joined here today by Dr. Dawn Mussallem, the chief medical officer of Fountain Life and a member of my Fountain Life medical team. Dawn, a pleasure. So Dawn, talk to me about brain health.

Dr. Dawn Mussallem

Brain health, you know, you're right. This is the number one concern people coming into Fountain Life have: will I remember the name of my child and the face of my loved one? 45% of dementia cases are entirely preventable with lifestyle. And what was really intriguing to me, Peter, is that a quarter of our members had advanced brain age, but over 13 months of us really helping them live healthier lifestyles, eating healthier, moving their body regularly, and optimizing sleep—people overlook that so often, but that sleep optimization is critical for our brain health—we were able to improve the brain age in 46% of those individuals. That's a powerful number.

That's amazing. You know, one of the things I love about Fountain is we're constantly searching the world for the most advanced therapeutics and bringing them to our members. So for me and all of you, I hope that you appreciate the fact that you can become the CEO of your own health. You can make sure that you've got the cognitive clarity for the next 50 years. Come and check it out, fountainlife.com/peter to learn more and become the CEO of your health. Now back to the episode.

Love it, thank you.

All right, I’m gonna turn us to 2 science stories that I know Alex and I are both excited about.

5. Antimatter Stays Contained

Hopefully you guys are too. Two weeks ago, physicists did something that sounds like Star Trek: they captured a single particle of antimatter and held it for a month, a new world record.

The backstory is that every particle of matter has an antimatter twin. An electron has a positron, for example, with the same mass but the opposite charge. When two of these particles touch, they annihilate each other, as in Einstein’s famous equation, E = mc². That’s why you can’t keep antimatter in a jar: it destroys itself the moment it hits the side.

A team at Northwestern built what they call a container without walls. They started with radioactive sodium, which spits out positrons, the antimatter twin of an electron. The positrons ricochet through a sliver of cold tungsten to slow them down. Then they’re steered into a vacuum just 0.3 kelvins above absolute zero and held in place by magnetic and electric fields. They kept one positron alive for a month, and said they could have kept it alive indefinitely.

So why does it matter? There’s an imbalance of antimatter in our universe. The Big Bang should have created equal amounts of matter and antimatter, which would have wiped each other out. Instead, a tiny excess of matter survived, and that’s all of us.

Alex, I think we’d need a lot of positrons to make a Star Trek warp drive, but—

Well, that’s the elephant in the room: the ability to shepherd around large amounts of antimatter in a contained way. This is how the starships in Star Trek power themselves. They keep little pods of magnetically confined antihydrogen and another pod of hydrogen, and that powers their faster-than-light propulsion. They also have a separate fusion engine that’s less efficient.

Maybe a bit of backstory on the actual science here. This is out of Gerald Gabrielse’s group at Northwestern. Funny backstory—maybe this will be pretty obvious—when I was a first-year graduate student in the Harvard Physics Department, I shared an office with Jerry Gabrielse’s group.

Of course. Of course you did.

It was a lovely office, but he’s since moved on to Northwestern. This is a major step forward. Jerry and his group are major leaders in precision measurements of a number of constants of nature.

In this case, the scientific interest is measuring CPT symmetry—charge, parity, and time-reversal symmetry. As you alluded, Peter, there’s probably quite a bit of interesting physics one should be able to discover relating to mirror symmetries of various sorts. Why do matter and antimatter differ in apparent abundance in our universe? Why do certain symmetries, like left-handed versus right-handed symmetries, seem to be disobeyed in certain physical processes? Chirality is very interesting, and somewhat surprising to many. That’s why this is scientifically interesting. Technologically, though, Jerry and his group are pushing us toward a world where we can do antimatter containment.

Mm.

I know our friend of the pod Jared Isaacman is also very interested in antimatter propulsion for spacecraft. But you’re not going to get long-term antimatter spacecraft propulsion unless you can hold it for a while, and right now that’s very challenging. So even if—

But hold large amounts of it for a while.

Hold large amounts of it without the whole thing blowing up.

Yeah.

And Jerry and his group have now pulled this off for a month at Northwestern.

So what does string theory say about this, Alex?

Stay tuned.

Okay. Come on.

You’re such a tease, Peter. I’m going to tease you back. I will have something to say about it, but I’ll speak through the publications. It is something I think about quite a bit: CPT symmetry.

Yeah. Awesome. Dave, any thoughts on this one? New energy source?

Yeah. Well, you said hold large amounts of it, but E = mc²—a sugar cube would power a city, right?

Yeah, for sure.

Well, it’s much more efficient. If you look at efficiencies just on a material basis, in principle, matter-antimatter annihilation should be 100% efficient.

No waste.

I mean, you throw off a lot of hard X-rays or gamma rays, but in principle you could use up all the rest mass if you have balanced hydrogen-antihydrogen propellant. As opposed to fusion, depending on what sort of fusion you use, it could be up to 1%, 2%, or 4% efficient. Fission is a fraction of a percent, like 0.1% to, I think, 0.3% or so efficient in terms of rest mass. So if you want a maximally efficient battery for propulsion for a spacecraft, with the physics that we have today, there’s no beating antimatter.

No pollution whatsoever.

You’d still have to heat water in a boiler and drive a turbine to make electricity.

Or you could carry a microscopic black hole, drop garbage into the black hole, and recover the rest mass that way. That is an alternative.

There you go.

That’s what the Romulans in Star Trek do.

Yeah. Okay, I love it. All right, let me move us on. Let’s turn to the space frontier. NASA is moving at Apollo speed once again.

6. NASA Returns To The Moon

On Friday, Administrator Jared Isaacman, friend of the pod, laid out his plan. Next summer, in 2027, Artemis III launches with four astronauts—three Americans and one Italian—on a two-week mission in Earth orbit. Their job is to practice docking their Orion capsule with both Moon landers, Blue Origin’s lander and SpaceX’s Starship. That’s three launches from Cape Canaveral in very quick sequence. That’s going to be super cool.

Then Artemis IV launches in 2028. It’s going to be the first time humans are on the Moon since 1972. Just a quick pet peeve of mine: I always spell Moon with a capital M. It’s a proper noun; it’s our Moon. I know you can say Luna, but when we’re talking about general moons, that’s a generic moon.

Anyway, NASA will wait as long as possible before choosing between SpaceX and Blue Origin to see who lands the crew. We’ll see if Blue Origin is able to make up lost time. This time they’re going to stay. NASA is building a Moon base near the lunar south pole, with the first launch by the end of this year. Many landers, rovers, and orbiters will follow, and it’s going to build Luna City.

Isaacman said this is a new space race, this time against China, and it will be, quote, “much closer than it ever needed to be.” It’s much closer than it ever needed to be. His fix is Apollo’s own playbook. It’s hard to remember that we went from Apollo 7 to Apollo 11, landing on the Moon, in just 9 months. That was insane, with less computing power than you have on your Oura Ring, I guess.

For me, this is personal. I’m going to be with Jared in Orlando, giving a keynote with him tomorrow—yeah, tomorrow night. It’ll be fun. So—

I love that, in the massive aspiration that you have in the world, you’re willing to die on the hill of the capital M.

I am. And the Moon needs our respect. Anyway, I—

The capital-M Moon. And then, 50 years of waste. What happened during the intervening 50 years with no humans on the Moon? By the way, the Soviet Union never even got around to landing a human on the Moon itself. It’s so bizarre. We need to fix this.

Yeah.

Well, this time there’s commercial pressure. I think it’ll be very different, and Jared is all into it, too. The divide-and-conquer there is: “Okay, Elon, Relativity Space, you get stuff into orbit cheaply. That’s very profitable. I’ll get the Moon base going. You can refuel my ships on the way.” It’ll be a military-commercial cooperation, and the commercial side doesn’t slow down for 50 years. That’ll keep the government side of it on its toes. It’s going to be a much better dynamic than it’s been in the past.

Yep.

And Jared’s all into it.

It’s interesting, right? The ability for NASA to hit this timeline and build its Moon base is based upon the federal budget. I think Jared’s doing the wise thing. He’s pulling out the old Soviet-U.S. race card and saying it’s the U.S. versus China. We’ve always needed a stalking horse, some kind of competition to drive federal spending, and it’s China this time.

I shared it on our group chat. Did you see the Chinese space plane—

Oh, I did see the space plane, yeah.

We’ve abandoned our own space shuttle, but there’s imagery now being released—

With wings.

It’s from the U.S. government via Maxar, showing this Chinese space shuttle with all of these seemingly advanced capabilities, long after we’ve abandoned our own space plane. What better stalking horse, what better nation-state near-peer adversary—whatever euphemism we want to use—to reboot the space race?

It’s going to be awesome.

I can’t wait. I missed the Apollo 11 launch; I wasn’t there. I look forward to being at the Artemis III launch. It’s crazy that it’s been 50-plus years. Pretty insane.

We can’t afford another 50-year gap. It’s so disappointing that there was this 50-year gap. I’m petrified that something like this could happen with superintelligence, where overregulation, changing political winds, the China syndrome, or something else stalls progress. We saw it happen with nuclear energy for 50 years. We saw it happen, arguably, with space for 50 years. I really don’t want this to happen this time with superintelligence.

I don’t think it will. It’s democratized enough that the open-weight people are just going to keep pushing these boundaries forward. Everybody can now operate on this. There’s no way it’s slowing down.

Alex, did you ever see the report that came out in the early ’70s, toward the end of the Apollo program? We left the Moon on December 17, 1972, with Cernan and Schmitt as the last Apollo moonwalkers. There was a report that projected NASA’s capabilities if they maintained Apollo funding: a 100-person Moon base by 1986, and a landing on Mars in the mid-’80s as well. They believed they could do it with 1970s and ’80s technology.

The story makes no sense, Peter. I think there was probably a contribution from changing political winds. Obviously, the Vietnam War and domestic shifts around the Vietnam War were probably an influence. But I do hope that, in the fullness of time, we understand why we weren’t landing humans on the Moon for 50 years. It seems like a historic travesty.

Yeah, for sure.

I think the cost of a war, especially the Cold War, is hugely underestimated. All innovation is on the margin, and all investment is on the margin. When you’re in a cold war or a hot war, all the margin goes into that effort, so all kinds of edge-case innovation get tied up in that capital allocation.

The scenario where this never happens again is one where there’s a relatively peaceful world. The scenario where we get stalled again is one where something really blows up in Russia, Ukraine, or elsewhere, and then it diverts all the budget, political votes, and attention into that.

Yeah.

So, cross your fingers.

Maybe. An area that interests me as a student of the history of American science and technology is what causes progress to happen in the first place. For example, the history of interchangeable parts in an industrial ecosystem, at least in America’s Industrial Revolution, is often widely attributed to the American Civil War. There was a need to be able to use and reuse ammunition, guns, and different gun components on the battlefield, and war was ultimately a driver of an industrial revolution, at least a contributor.

To my knowledge, no one has conclusively studied or understood what the optimal state of great-power competition is, if there must be great-power competition, to optimally drive progress forward. For a while, it seemed like World War II and the post-World War II period, despite the enormous human cost, were achieving fast progress. Then everything stopped sometime around the ’70s. I would love to know what went wrong and how we fix it.

If you guys are good with it, let’s move to AMA with the mates. If it’s okay, I’m going to take question number one because it’s kind of pointed at me. Number one is from @christophejeffers: “How does LEV, longevity escape velocity, trickle down to the poorest among us?”

Good question, and something very important. I think longevity is probably the largest business on the planet besides AI, and a lot of the treatments today are expensive. Let me first say that the number one thing everybody should be doing is sleep, diet, exercise, and mindset. They’re cheap, and you can do them.

But there are technologies being proposed, like partial epigenetic reprogramming. This is the work that Dr. David Sinclair just reported on with his Phase 1 data. They’re using three of the four Yamanaka factors, delivered into cells with gene therapy using an adeno-associated virus. That’s expensive. Using AAVs and gene therapy can cost anywhere from half a million dollars to $2 million per treatment.

But this is where it gets exciting. David also announced he’s been developing a treatment called SL-100. It’s made up of three molecules he discovered using AI. They’ve been tested in human cells and in mice, and his estimate is that it will cost $200 a month. Remember that everything starts expensive, and then technology is the force that takes what’s scarce and makes it abundant, from the cell phone to intelligence.

We also have the $101 million Healthspan XPRIZE, focused on reversing aging in cognition, immunity, and muscle by 20 years. One of the basic premises of that competition is that it needs to be affordable for everybody. All right, Salim, over to you. What do you want to pick?

I’ll pick number two. It’s the obvious one for me: “If AI is doing most of our work, where will people actually find purpose?” We’ve talked about this a couple of times. I made a whole video about it on my channel. We’ve made a basic mistake of conflating employment with purpose.

If you look at corporate workers, 79% don’t feel connected to their work. They don’t find meaning in their work. This is a longstanding challenge. We’ve faked ourselves into doing this. The industrial world taught us that value comes from productivity, but that’s not actually the case. It’s going to come from other things.

As AI does more and more of the work, we have this unbelievable opportunity to spend more time being, spending time with kids, building communities, and doing scientific exploration. This is why the concept of a massive transformative purpose is so important: it gives you a taste of purpose, something to point yourself at and a place to point intelligence at. That’s very critical. Imagine a world where nobody has to work to survive, but everybody has an opportunity to contribute. Those are a couple of ways of thinking about it.

Great. Dave, over to you.

Alex, you want three, right? So I’ll take four.

Yeah, I’ll take three.

Yeah, okay. I’ll take four. What is the… I was just guessing.

You’re so polite.

Sometimes they feel like they’re labeled for somebody. Number four: “What is the equivalent thing or practice for humans to develop reinforcement learning or recursive self-improvement?” Human recursive self-improvement. That’s from our old friend—

That’s from our old friend—

—and that is from the awesome CJ Trueheart—

Yeah.

—who we love because he contributes all the time, and he’s just a great guy in LA.

I had a conversation with my grandfather before he passed away many years ago. He was a farmer and cheese factory guy in Ohio, and he asked me, “What do you do for a living?” I was like, “Well, I code on this computer.” He said, “You’re not working. You’re typing. That’s not work.” I said, “After this, I’m going to go to the gym.” He was like, “Why on earth would you go to the gym?” He was used to working in the field all day. It seemed utterly insane to him that a human would go to a gym.

Mm-hmm.

Here we are. We’re an improving society; we just lose track of how much we’re improving. I think AI is going to unlock much better daily planning and all kinds of ideas for how we can recursively self-improve even more. I bet we’re nowhere near optimal in terms of our health, happiness, daily practices, or daily plan.

I think AI turned in that direction is going to be one of the greatest boons to humanity of all time. I just hope Mark Zuckerberg and Sam Altman reserve some compute for that endeavor. I think it’s going to be great.

Awesome. Alex.

All right.

Thank you, Dave, by the way, for reserving number three. Number three asks, “Can abundance spark a second society-wide renaissance?” This is from Pat Toez, @patoez1.

Let’s think more broadly. The original Renaissance—the European Renaissance from the 14th through the 17th centuries—is broadly attributed to a rediscovery of classical antiquity, among other factors, obviously. There are, as always, numerous input factors in one of these large-scale shifts. But there was a rediscovery of classical antiquity artistically, technologically, and socially, and I think there were many local optima in at least Western civilization.

There was obviously the Roman Empire. Drop a footnote: How many times do American men think about the Roman Empire every day? I’m thinking about it at least one time today. But there were other periods. As I gestured earlier, there was the post-World War II period, when there was very rapid progress being made on atomic/nuclear energy and, shortly thereafter, the original Space Race.

One way of answering this question, I think, is: if I construe abundance to be the singularity, or just the present time, I do think we’re seeing a second renaissance. I’m not necessarily referencing the European Renaissance from the 14th to the 17th century, which was rediscovering antiquity. Rather, I think we’re starting to rediscover the world from, call it, 50 years ago, when, in the post-World War II era, all these emerging technologies—nuclear power, computers, a variety of space-associated technologies, and many others—burst onto the scene for the first time.

Going back to the earlier discussion, maybe some would call it, as Tyler Cowen would, a Great Stagnation: 50 years of arguably wasted time. If Isaac Asimov were in this discussion, maybe he would write a whole new meta-Second Foundation about thousands of years of wasted time after the collapse of empire. But in this case, it’s not the Roman Empire all the way to the Renaissance. It’s the post-World War II era, when a bunch of changes were introduced and there was rapid progress coming out of World War II, followed by this 50-year gap.

In short, yes, I think we are already seeing a new form of renaissance, but it’s not anything like the 14th- through 17th-century European Renaissance. It’s a rediscovery that the physical world shouldn’t be a source of fear. It should be a source of opportunity and inspiration, and we should embrace everything that is within our power to control in the physical world and not be afraid of it.

All right. Dave, over to you, pal. Five, six, seven, or eight.

All right. I’ll take seven. “Elon Musk believes that Western civilization is contaminated by empathy. How does this group see it?” That’s from DDMT Year. I can’t speak for the whole group, but I think he has a really important point. The U.S. is statistically the most generous country in the world, and I’ll tie that to being empathetic as a country as well.

I think what happens, though, is that we’re reactive to the media. Yesterday, there was a Learjet that crashed in Florida, and now people are like, “I don’t want to fly anymore.” But if you look at the statistics, it’s by far the safest way to travel.

Yeah.

But they didn’t show all the train wrecks and car wrecks on TV last night, so we don’t react to that. We’re, quote-unquote, “contaminated by empathy” in the sense that we make decisions based on statistical outliers that are right in front of us on TV or the internet, whereas China doesn’t do that.

China has statisticians and researchers trying to gravitate to the right overall answer, and that’s why they have abundant energy and Europe has none. It’s that same effect. So it’s a byproduct of a good thing, which is empathy, but it’s miscalibrated by the media. That’s why I think it causes the quote-unquote contamination. Empathy is fundamentally a good thing. Elon knows that. We all know that. But we can’t make decisions on emotion all the time.

Please remember—

Even the State of the Union address, where you call out a single person and say, “Stand up and tell me your story.”

Oh, it’s terrible.

It’s terrible. It’s terrible. Sorry, Peter. Carry on.

Please remember, the media’s job is to scare the shit out of you, right? If it bleeds, it leads. They’re just wiring your neural net the same way you train a large language model, by showing it data after data. They’re just training your neural net on all the negative news and all the fearmongering so they can sell your eyeballs to their advertisers. That’s their business model. Understand their economics. Alex, over to you.

As opposed to Moonshots, where we don’t advertise and we don’t care about eyeballs at all. I read you loud and clear, Peter.

Well, we’re not trying to scare the shit out of you. We’re trying to give you optimism.

No, we aim for a different set of neurotransmitters in our audience. At least we understand what we’re aiming for.

Fair enough.

Fair enough.

Thank you for always being the truth-sayer, my friend. I appreciate you.

“Balls and strikes.” I love it. I’ll pick question number eight. “Robots are simultaneously not going to have a market but will also be entirely consumed by industry. Which is it?” This is from latent4407.

I picked this question because I think this is maybe aimed at some of my comments from the past two pods. When I’ve spoken in previous conversations about robots potentially not having, say, demand catch up to supply, it was in reference, by analogy, to what I think we’re seeing or about to see with Elon and Starship. All of this upmass is about to come online, seemingly, and yet you really can’t saturate all of SpaceX’s forthcoming upmass with communication satellites.

It’s going to require an entirely new use case for what we’re going to start lifting into orbit. Right now, Elon’s plan seems to be orbital data centers and the Dyson swarm. I’m very optimistic that that’s a good way to saturate all of this new supply that’s coming online. But if he doesn’t do that, if Elon decides, “Well, we’re just going to launch all of these Starships,” and there aren’t new use cases, I do think there’s a very real chance that many of these Starships, if he builds them at the quantity that is anticipated and being publicly forecast, will end up going empty.

I don’t think that’s realistic, because I think ODCs and other new use cases are going to come online. Similarly, by analogy with general-purpose robotics—largely humanoid, I expect—I think there’s so much humanoid robot capacity about to come online that it will overwhelm existing demand. The market wants to clear and supply wants to meet demand, and I think it’s reasonable to expect that the market will clear via new use cases and new industries capable of saturating the new supply of robots.

I rather like Dave’s hypothesis, and Sam’s as well, that one of the major demand sources is going to be robots building data centers or otherwise maintaining the innermost loop, as it were. That’s probably the best use case I can think of in the short term as well.

And we saw this with—

Yeah.

Automotive manufacturers, right? Alex, we had, like, 100 automotive manufacturers at one point, like 200 different tire manufacturers, and it all sort of got figured out by the market.

Oh, I thought you were going to go in a different direction, which is that Henry Ford had an analogous problem: the overproduction of the Model T. There was the problem of how you saturate demand and living standards. What do you do? You ensure that your employees are compensated and incentivized in such a way that they’re able to hopefully purchase some of your own cars.

I’m not quite sure what Elon’s example is. Maybe it’s that he sends all of the SpaceX and Tesla employees home with 10 robots or something as a way to juice the market. But he’ll need some sort of Henry Ford-esque virtuous self-licking ice cream cone.

All right. Salim, over to you.

I’m gonna leave number five for you. Let me take number eight. Robots are simultaneously not going to have a market—

Oh, no, that was—

Sorry, sorry. Data centers. Let me take the data center one.

Yes.

Where is it? Where is it? Where is it?

Yeah, number six.

Number six. Okay. “How does living next to a data center make you rich? Living next to a factory didn’t. Why are data centers different?” That’s from @donnadaves.

Factories created more opportunities in the near term because you could employ the local workforce, you could purchase local services, et cetera. A billion-dollar data center has the building aspect of it, but it may not purchase much locally, and many of the benefits may be distributed and get sucked out very, very quickly.

However, if you go back to what Alex just talked about, the need for compute is going to be incredible. It’s like saying we don’t want electricity anywhere near our neighborhood. You want electricity in your neighborhood because you want to thrive from it. When you can apply intelligence to all these things, the great economic value is not just living next to the data center. It’s using the intelligence that the data center produces to do lots of really interesting things. This is why it’s so critical and important: the amount of compute we’re going to need is near-infinite, and we’d really like to take advantage of that. Learn how to leverage the intelligence coming out of it.

The other thing is, the data centers are going to make a lot more money than the factories ever have.

There is that possibility, too.

A lot more to distribute. Okay.

“How quickly memory dims?” What happened to factory towns? I’m not even sure if I buy the premise of the question.

Mm-hmm. Number five—

You never accept the premise of any question, Alex.

The premise of the— You know what? Elon says, “Delete,” and I say the premise of most questions is probably wrong.

Okay.

Okay.

Having said all that, I’m going to move on to number five. We hear a lot about colleges. Why should students not pursue trade school? From @PaulBeechaman. Paul, they should pursue trade school. There’s a lot of opportunity for electricians, welders, and plumbers, for sure, especially now as we are tiling the planet in data centers. I think you’ve answered your own question. I think they should. At the end of the day, students should find their purpose in life, whether that’s their family, art, or building a company, and then pursue the strategy that gives them the most agility and opportunity. Trade schools are definitely there. I think we’ve seen a huge shortage in electricians, plumbers, and air-conditioning specialists. The singularity needs you.

Very important.

This is actually very clever. For those who aren’t watching the visual stream, these are sentences being depicted as transformer attention. It’s pretty clever.

It’s going to become the new karaoke mode, Peter. Did you catch in the video how we were seeing cross-attention from words to each other in the karaoke feed?

That’s great.

Gotta check that out later.