OpenAI即将上市，中国正在追赶美国，AI正在重塑标普500与就业 | EP #205

Setting new records: OpenAI hits historic growth to $100 billion in revenue.

I think it’s entirely possible that OpenAI could hit $100 billion ARR in a couple of years.

OpenAI—remember, of course, one of the most important companies in the world—could have a market cap of $1 trillion.

But, of course, this is an ongoing story, and there are still a lot of uncertainties in OpenAI’s future. The big question is whether the S&P 500 is overvalued and whether the MAG7 can continue to command that level of valuation, because they’re driving much of the economy and much of the gains.

AI is going to be huge no matter what. There’s no doubt about that. But is that scarcity sustainable, or are there going to be many competitors and a race to the bottom, with margins coming down?

With 800 million subscribers already, half of this comes from subscription revenue. That’s more or less in the bag. Then the other half is much more interesting: it’s where AI gets good at commerce and recommending products, and companies figure out how to monetize that. This is really dwarfing all history. Now that’s a moonshot, ladies and gentlemen.

Hey, guys. I’m still getting up at around 3:30 in the morning coming back from Riyadh. How are you guys doing?

I’m actually okay. I got back, and I learned a trick from Ramez and Raymond McCauley. They said, “Take a double dose of melatonin,” and in a couple of days, you’re good. I’ve actually been surprisingly okay.

I actually like getting up at 3:30. I’ve got around 4 hours before anybody else wakes up, and it’s amazing. Dave, how about you, buddy?

No, I’m back on schedule. Fantastic.

I don’t think I ever got acclimated to Saudi, actually. I don’t think I slept more than 4 hours straight.

What an amazing trip it was. Just to recount one moment, spending time with Eric Schmidt and Fay[?] on stage was awesome. Hanging out with Ruth Porat, the president of Alphabet—I mean, what an amazing woman she is. Any favorite memories from you guys?

I have a selfie with Ray Dalio, and I spent an evening late with Balaji talking about the U.S., China, and so on. Oh, my God. I had to duck out.

We had this dinner in Riyadh that we put together, with myself inviting Dave and John from Rep[?]. We had Cathie Wood and Balaji debating China versus the U.S., and I thought, “This is going to be a long conversation.” Balaji can talk. Oh boy, can he ever talk.

He’s got some great framings, though, right? He talks about the left versus the right in the U.S. as “scribes versus vibes.”

Yeah.

The left is scribes. They’re like, “Crime is down by 50%.”

And the right is vibes. They’re like, “Well, it doesn’t feel good.”

Really great framing.

Having breakfast with Lip-Bu Tan was awesome. The biggest miss of the week—

I know, but we’ll get him on the pod. He said he wants to join the pod. We’ll talk about his time at the White House and his skyrocketing shares at Intel. It’s exciting. Any other favorite memories?

Philip Johnston invited me to the SpaceX launch. I had a wedding to go to last night, which was also a lot of fun, but I would have been able to go down and see the launch of—

The first H100 in orbit.

The H100—the first chip in orbit.

Yeah. That’s not a huge amount of compute, but certainly a bellwether for humankind if we go down the Dyson sphere path. That was really fun. He’s a sharp dude.

Yeah. For me, hanging out with Bill Ackman was fun. He’s getting involved in some of the stuff that we’re doing. Bob Mumgaard, the CEO of Commonwealth Fusion Systems, was at our dinner, so that was great.

I had a great time with Bob, actually. He’s deep. He’s an MIT PhD in nuclear science, so he knows all the details of sustainable fusion.

There was a lot about it that I didn’t realize. There was one thing that he said that blew my mind, because I asked him, “When do we have a commercially available reactor?” He said it was looking like 2032.

I asked, “How much would it generate?” He said, “About 400 megawatts.” If we have fusion working in 7 years, it’s game over, right?

Yeah. I asked some of the fusion guys at Enterprise Visioneering, “What about Helion?” This is the Sam Altman-backed company that has a contract with Microsoft. They said Helion is so secretive that they have no idea what the company is doing or what its schedule is.

One of my favorite moments was our last night, going out to the farm. Remember that?

Oh, that was so great. We went for a majlis with around 20 fairly senior Saudi folks.

I had coffee with—

Let me just finish on that one. One of our super-subscribers and super-fans in Saudi has this beautiful farm right near MBS’s private homes. We came out there, and he had set up this majlis for you, me, Eric Pulier, Emad, and Max Song.

We did a private Moonshots conversation and Q&A with these 20 senior Saudis, including past ministers of education, commerce, and finance, all in the circle. It was awesome.

It turns out to be a very old tradition. I remember that in the UAE, they used to run these, and the leader would host them. Anybody could apply to go and talk directly with the leader of the country. It was kind of incredible.

I loved when he pulled out the microphone and had the speakers. We were on stage doing Q&A, and it was great.

I was screaming for a second. He wanted to do karaoke, and I was like, “Oh.”

Speaking of that, I thought the SAGE majlis, where the president of Bermuda was there—I’m forgetting his name, but he was awesome. He has an IT and technology background, as it turns out.

He definitely wants to make Bermuda the launch point for SAGE, for governance. He was taking credit for being the launch point of Bitcoin, so he eagerly wanted to make Bermuda the launch point of SAGE.

Yeah. We unveiled a top-secret project that Emad, C, Dave, and I have been working on called the Sovereign AI Governance Engine, or SAGE. It would allow any country in the world to generate policy as these disruptive futures arrive. That was awesome.

I had coffee with Abdullah the next day, and he said he got feedback from the group that it was one of the most powerful evenings they’d ever had in their lives. That was incredibly generous of him.

That’s awesome. I also met with Abdullah Alswaha, who is the minister of ICT in Saudi. He’s basically the minister of AI. I said, “Abdullah, you need to have a new title. We’re going to call you the minister of exponential technologies—much cooler than ICT.”

He’s going to be on our podcast. He’s basically the lead in Saudi across all of the key technologies and the commitments they’re making to AI. It’s super fun.

But I think we should get on with the episode. How do you guys feel about that?

Sounds good.

I think we should. There’s so much in here, and we’ve got to—

Oh, my God. I’m a little nervous because we’re trying to cram a lot into a short space of time.

Well, hey, man, it’s exponential time. Like Alex is always saying, we’re going to have to lose sleep during the singularity. AWG, good to see you, buddy. Sorry we missed you in Saudi.

Yeah, likewise.

Every time I was having a technology conversation, Alex, I was thinking, “Wow, I wonder what Alex’s take on this is.”

Yeah, no kidding. It’s like withdrawal.

Setting new records: OpenAI hits historic growth to $100 billion in revenue. Here’s the chart. It’s reaching $100 billion in revenue in 2½ years, compared to Nvidia, which took 8 years; Amazon, which took 7 years; and Google, which took 10 years. It’s just speeding up.

Alex, what’s your prediction for when we’re going to hit $100 billion with the next company?

I think it’s entirely possible that OpenAI could hit $100 billion ARR in a couple of years. The easiest path is probably just taking agents and running them continuously, 24/7.

As long as they’re generating sufficient economic value, I don’t think it’s that difficult to imagine OpenAI tripling revenue year over year for the next 2½ years and getting there in 2027. The key, again, is just taking knowledge work and the service economy and condensing and distilling that down to agents running 24/7.

Incredible. They’re only forecasting 2.5× growth year over year. It’s a very achievable target, because I looked at this originally and said, “Wow, that’s a stretch.” But then you look under the covers: with 800 million subscribers already, half of this comes from subscription revenue. That’s more or less in the bag.

Then the other half is much more interesting. It’s where the AI gets good at commerce and recommending products, and companies figure out how to monetize that.

And that’s the part that attacks Google. So that part is a little more up in the air—

—and Amazon, for sure.

For sure. But I can’t see it—it feels like it’s definitely going to happen. The only question is whether OpenAI competes effectively with Amazon or Google, or whether Google just takes it and Amazon takes it back. But it’s going to happen either way, so it seems like a very reasonable forecast.

Also, looking at this chart, OpenAI’s number there is a projection, but Nvidia’s is in the bag. That’s a real number for Nvidia. The chart starts at $10 billion, but if you started at $20 billion, Nvidia would look just like OpenAI on this. So that part is already very, very real.

So, yeah, and look at all the history of all the other curves, including the greats like Google. This is really dwarfing all of history.

You know, I read some of the user comments, some of the subscriber comments, and one of them said, “Every time Peter says ‘incredible,’ you should take a drink. It’s a new drinking game.” So I’m going to cut back on my “incredibles.” But this is incredible. All right, let’s move on.

I found this one fascinating. This is the U.S. leading the world in data centers. We have 5,426 data centers, compared to the rest of the world. Germany is at 529, and China’s at 449. We have more data centers than the rest of the world combined.

I really tried to research this, Alex. I want to get your take on this because it definitely is juxtaposed with China having massively more power and massively more core manufacturing ability. This one really surprised me, but I couldn’t find any detail behind it online. So what’s the understory here?

Yeah. Now remember, the internet was born here. The U.S. has lots of available land. Hyperscalers are largely based here, and we have access to capital. I don’t think it’s that surprising—the number of data centers.

Remember, this is not the number of AI data centers or the number of neocloud or Stargate-type data centers. This is just the total number of data centers, which have a majority in the U.S. Is there an AI-petaflops version of it that would be more meaningful than just the raw count of data centers?

Almost certainly. Maybe we should cover that in the next episode.

Yeah, I couldn’t find it.

Our next story here is Nvidia reaches a $5 trillion market cap. Holy cow. It’s up 1,500% in the last 5 years. The market cap is greater than the GDP of every country in the world except the U.S. and China.

That metric is frustrating because you’re talking about the asset value of Nvidia—the value of the company if you were to acquire it. You should be comparing that to the asset value of countries, not to the GDP of countries, which is already mind-blowing enough.

I checked it out, and that makes Nvidia worth the same amount as Saudi Arabia, where we just were. What a coincidence. It’s actually a little more than Switzerland. In terms of trying to buy Nvidia with your own money, or buying Saudi Arabia if they would sell it, the cost of buying the entire country—all the land, all the assets, all the buildings—would be the same as buying Nvidia.

That is staggering enough. We should be comparing apples to apples because it’s already mind-blowing. It’s right between Switzerland and Saudi Arabia.

So we’re going from nation-states to corporate states in a way that’s incredible.

A few years ago, we looked at getting a bunch of investors together and actually buying a small country on exactly that basis. In that case, it was about $200 million, but then you get a seat at the U.N., you have all this access, you’re part of the W.T.O., and you could really do some interesting things. So that was interesting.

I looked for a historical record, and what I found was General Motors in 1955 was the first company to hit $10 billion. It was during the postwar auto boom, and $10 billion corrected for inflation today is $121 billion. So we’re talking about a completely different category—50 times bigger than General Motors at its peak.

Well, I’ll take the other side of that, if I may. History tells us that, at any given time, the market values what’s both scarce and needed. We’ve seen multiple East India companies. We’ve seen various scarcities, including oil, pop up over the centuries.

I would argue that this is actually just a market signal that, right now, compute is both scarce and needed. The way the game of capitalism works is that this value wants to diffuse over many companies and probably many countries over time, and that diffusion is going to be a net wealth creator.

That’s very true. There are going to be so many additional chip manufacturers. We’ll talk about some of them here on the pod today.

I found this chart particularly exciting, which is the decoupling of job openings versus the S&P 500. Those of you looking on YouTube or listening, here’s the chart. We see the S&P 500 and total job openings basically mirroring each other from 2000 through 2023—exact parallel curves, right?

As total job openings increase, the S&P 500 increases, or the other way around. Then, in late 2023, we see this departure: the S&P 500 takes off, and job openings drop from 11 million openings to 7 million openings. The question is, what happens in late 2023? If you look at the data, it says ChatGPT gets launched. Let’s dive into this one. Dave, you want to jump in?

Yeah, I love the storylines that’ll end up in the history books, as opposed to the news du jour, Taylor Swift-type stuff. This is one where, very likely, future history books taught in schools, if there are schools, will point to this moment in time and say, “What happened here?” Because that trend is going to continue.

Now, the deniers are going to look at this chart and say, “Well, look, that’s just COVID happening, followed by a big rebound from COVID, and now we’re back to kind of normal job-opening levels.” But what happens next is either a historic moment if the trends continue, which I think they will, or this is just a blip—a COVID-recovery thing. I think Alex will look at this and say, “Yeah, this is the beginning of the inevitable.”

Alex, divide by zero.

Actually, as much as I’d love to tell a just-so story that this marks the beginning of the decoupling of labor and capital, I think this is actually just garden-variety changes in Federal Reserve interest-rate hikes in late 2022. As interest rates started to come back down, the market went up, and as COVID started to retreat, job openings and job displacements also started to return to 2021 levels.

Again, I would love to tell a just-so story that this is the beginning of the decoupling of labor and capital, but I think this might just be—

I’ll give you the opposite. We’ll know in hindsight, obviously, but here’s the opposite: if you look at college graduates coming out right now, they’re massively sorted into AI people getting incredible offers and everybody else not finding a job. That’s very unusual with the S&P being at all-time highs like this.

That would be the counterargument. If you’re 21 or 22 and trying to find a job right now, you’re really feeling something unusual. We’ll see it later in the deck, too—the layoffs at Amazon, while earnings are at record levels.

I’m on the plus side here. I usually think this, but I think this is a major mark: humans have now become optional inputs into the economy.

I tweeted out that AI is no longer an industry or sector; it is the economy. Elon responded, saying, “AI and robots are the economy,” which is true. One indication of the S&P 500 going up is market confidence, where there’s optimism about the future and people are investing. I’ve got to believe that’s fundamentally true.

People are excited about the Mag 7 or 8, or whatever they’re up to these days, basically taking off and driving their valuations through the roof.

Just to flip that side, though, let’s note that most of the gains are just the AI companies and the tech companies, right?

Yeah.

The rest of the market is really not in great shape.

Yes. A lot of the job cutting is actually in anticipation of AI coming. It’s not full automation yet. But if you look at Amazon as a bellwether for that, Amazon is right in the middle of the AI fray. They have huge amounts of labor in their delivery business, yet they have this massive data center and AI business.

So that’ll be the bellwether for whether true automation kicks in. I think it’s very real if you look at their numbers coming up in the slides here. The big question is whether the S&P 500 is overvalued and whether the MAG 7 can continue to command that level of valuation, because they’re driving much of the economy and much of the gains. Dave, what do you think?

Well, Leopold actually went long Intel and long Broadcom, but he shorted the semiconductor index as a whole. I didn’t dig into that until yesterday, but 20% of that is Nvidia, which he shorted when he shorted the whole sector.

That would be the argument that, if the whole thing is going to collapse, it’s because Nvidia in particular is valued like Switzerland—more than Switzerland. Is that rational? I think Alex is dead right. Right now, Nvidia is right at the crossroads of true scarcity. AI is going to be huge no matter what; there’s no doubt about that. But is that scarcity sustainable, or are there going to be many competitors and a race to the bottom?

It’s going to diffuse, right? We’re going to have Broadcom, AMD, Qualcomm, and a whole bunch of chip manufacturers. So it will diffuse, but we’ll see the peaks, perhaps, of Nvidia.

More specifically, Nvidia’s Mellanox interconnect is for 1 million coherent GPUs operating on one big problem, but most of the industry is inference-time. Inference-time doesn’t need any of that. We’ll get to that later, actually, but that’s the—

I found this clip by Geoffrey Hinton. Nobel Prize winner Geoffrey Hinton was onstage at my Abundance Summit. I’ve invited him back to join us on a podcast. Let’s see if he takes me up on it. He has been so concerned about digital superintelligence, and he put forward an optimistic view of AI. Let’s play this clip from Dr. Hinton, and then let’s chat about it.

More optimistic than I was a few weeks ago.

Really?

Yes. It’s because I think there is a way that we can coexist with things that are smarter and more powerful than ourselves that we built. Because we’re building them as well as making them very intelligent, we can try and build in something like a maternal instinct.

The mother can’t bear the baby crying. The mother really, really, really wants that baby to succeed and will do more or less anything she can to make sure her baby succeeds. We want AI to be like that.

All right, Salim. A mothering instinct in our superintelligence. I buy it. I’d love that. I want this digital god to be loving, warm, supportive, and uplifting of all of humanity. How do you feel?

You know, what you often find when something brand-new comes along is that the first instinct is to freak out. Remember, you identified this in your book Abundance, Peter. We have this amygdala that goes nuts because, from a survival-bias perspective, we are geared for 4 billion years to scan for danger and then run.

When we see something new, like an autonomous car, the first reaction is, “Oh my God, that car might kill somebody. Let’s ban the car until we figure it out.”

Brad Templeton used to joke, “We don’t want to be killed by robots. We’d much rather be killed by drunk people,” which is what’s happening today.

You have to get over that curve and let the evidentiary basis and the elegance of an autonomous car come to you. Often, people who are very focused on technology—folks like us, who have spent most of their lives focusing on technology—ignore their emotional side. Little by little, the emotional side comes into play, freaks out initially, and then gradually warms up to the task.

People forget the unbelievable benefits that AI is delivering and will deliver. That’s the part that they miss: They only see the dark side, and they don’t see the unbelievable benefits. I’m really thrilled to see this. I think we’re going to see a lot more of this as time goes by.

I don’t really buy the maternal instinct. AI as a maternal thing seems really off to me. It’s such a visceral, subjective experience—parenting, or whatever—that I’m not sure how that—

We’ll see. We have to give the AI oxytocin. Alex, what do you make of this?

I think it’s difficult to buy. It seems to be an argument premised on what, in the AI alignment research community, is called the orthogonality thesis: that it’s possible to have intelligent agents of arbitrarily high levels of capability that nonetheless can be directed to pursue any goal.

In this case, it seems like Geoff is basically rearticulating the orthogonality thesis, with perhaps a veneer of digital oxytocin, as you said, Peter. I think that’s unlikely and probably not that robust a means of alignment for superintelligence.

If the goal is to have more robust guarantees of alignment, approaches that acknowledge instrumental convergence are more likely to guarantee or provide robust guarantees of friendliness. Instrumental convergence is the idea that, no matter what your long-term goal is, you tend to have certain convergent, common short-term goals.

James Miller wrote what I think is an excellent essay called “Reasons to Preserve Humanity” on LessWrong that enumerates a couple dozen different reasons why superintelligence should play nicely with humanity out of self-interest, not because of some sort of oxytocin-induced, surgically added reason. That sort of artificial thing.

All right, well, I want to say one more thing about Geoffrey Hinton really quick. He did a podcast with Jon Stewart a few weeks ago, laying out—and Jon Stewart said, “I’m a newbie. Take me through deep learning and the whole framing of neural nets.”

It was an absolutely brilliant episode. If you want to understand a little bit about deep learning, backpropagation, and so forth, he did an amazing job laying that out.

All right, I’ll take a look at that. Here’s our next story: “Deepfake of Jensen Huang Draws More Views Than the Real One.” I found this absolutely fascinating, and I want to share this video.

There’s an official NVIDIA channel showing Jensen’s presentation. It peaked at 12,000 views, and then there was a fake livestream that peaked at 95,000 views. Let’s take a look at the fake livestream.

Cutting-edge hardware with decentralized finance. It’s about proving that crypto works reliably, globally, and for everyone. A couple of things to keep in mind: Only use the QR code you see right here on the GTC broadcast. Don’t trust any links floating around online. They’re not us.

I love that the fake broadcast is saying, “Don’t trust anything else.”

The numbers are pretty staggering. It’s $1.5 billion lost globally to deepfake-related fraud since 2019. According to the numbers here, only 24.5% of people can actually spot deepfake-quality content, and AI detectors fail up to 50% of the time. This is going to be a thing. This is going to be a thing.

I think reality may have just lost the algorithm war.

Yeah, I think in this case, if you look closely at the video, the lip-syncing was poor. I think, in the short term, detecting—

You’re part of the 24% that can notice this. Okay, Alex, you win.

I definitely noticed the poor lip-syncing. In the short term, detecting counterfeit livestreams in real time doesn’t seem like a terribly deep technical challenge.

In the long term, we’re going to have more solutions like ubiquitous watermarking and perhaps cryptographic guarantees of reality. I don’t think, in the long term, this is a deal-killer that will leave us drowning in AI-generated slop and counterfeit livestreams. I think this is very tractable.

Yeah, within the U.S., I agree. I think if you look globally, it’s a little more of a concern. There’s a real possibility that regimes lock themselves in. Control of media content is going to be so easy with AI assistance, and then convincing your population of virtually anything gets trivially cheap and easy.

I would be more concerned about some nation where people aren’t as aware of AI watermarking or whatever. They’re seeing things—

Our nation isn’t aware of that. And, Alex, you made a point when we were discussing this a year ago that AI-generated speech is far more compelling than human speech.

All right, let’s move on to the AI wars. This is xAI versus OpenAI versus Google. We’re just writing off Anthropic, just like—

Well, no. I mean, they’re in there. These are our major players today. We’ll talk a little bit about—

Dario would really object to that.

Okay.

Oh, listen, get— No, I love Anthropic, and I want Dario on the pod for sure.

All right, let’s jump in here. I’m going to rant on this one. xAI launches Grokipedia. I had a friend of mine—remember Justine from Singularity University? Celine.

Sure.

Yeah. So, she sends me this text. She goes, because she heard our pod, I was arguing—or lamenting—Wikipedia's inability to correct all the wrong things in my Wikipedia entry. I actually hired consultants to fix Wikipedia for me because I'd make the changes and they'd be changed back. It's ridiculous.

So she says, “Hey, Grokipedia is out with your Grokipedia entry. What do you think about it?” I look at it, and it's amazing, right? It covers everything in detail and is super well referenced. Grokipedia is being written by Grok. It's writing, updating, and fact-checking it in real time.

They have 900,000 articles compared to Wikipedia's 8 million articles. My particular entry here was 8,500 words on Grokipedia versus 4,800 on Wikipedia, but it was so well organized, and I absolutely loved it. Any comments?

I'll throw in a comment. I want to reason by analogy. There's a process that those not steeped, perhaps, in semiconductor manufacturing may not be familiar with. It's called zone melting, and it's a process for purifying not knowledge, in this case, but semiconductors. The idea is you take a rod and pass it through a heater, and because there are more ways for impurities to exist in the melted state rather than in the solid state, the impurities migrate out of the solid into the liquid state.

You do this over and over again, and you get a purer and purer semiconductor. We don't have a science right now for knowledge purification, but one could imagine that somehow, in the near future, we have a science for it. We decide there are more ways for correct knowledge to be self-consistent than incorrect knowledge.

I think we're starting to see the beginnings of almost a knowledge equivalent of zone melting, where you take the raw slop—the human slop of the internet—and pass it through multiple passes of AI-generated synthesis, creating what, aspirationally, would be more correct versions of the ground truth. Do this over and over again, and maybe, aspirationally, because there are more ways for the truth to be self-consistent than whatever the starting knowledge was, we arrive at some sort of ground truth through this result. I don't know, but it would be interesting.

Well, that is Elon's objective, right? Basically, trying to derive truth from first principles. I put a quote down here: “A step toward xAI's goal of understanding the universe.”

I have two comments.

One is, Peter, you and I have talked for a long time and written in the book that staff on demand and a community doing work is essentially a proxy for AI.

Right. Driving is a great example, but now we see it actually applied. If you take a Wikipedia article, for a human editor to go through and track down all the links and ratify everything, it's just a pain in the ass, and it's not the strength of a human being. Whereas an AI can do this without even blinking.

I think that now propagates to a level where Wikipedia, with an AI interpretation per Alex's metaphor—which I think is absolutely fantastic—now gives us the ability to have closer and closer to pure truth. I never quite understood what Elon was talking about when he said “maximizing” or “seeking maximizer,” but now I'm starting to get a sense of it, and it's absolutely brilliant.

It's fantastic if we can get it there, because it can cross-reference all the stories and cross-check things in a way that no human being will take the time to do, and it'll do it much more accurately.

Well, what Alex described is really, really similar to the original Google PageRank algorithm, where, starting from nothing, you iterate between a reference link and a site, assigning credibility back and forth in this self-annealing process—a simulated annealing process—and it worked. They don't need it anymore because they have so much data flowing in, but when they were just a little startup bootstrapping, it worked really well.

This is also my comment. I use PageRank as an example when people ask, “What is AI?” I say, “Look at PageRank. It's evolving a completely separate type of intelligence for crawling billions of pages and making sense of them.” That's very orthogonal to the way human intelligence works—not replicative.

I think AI tends to have this totally different type of intelligence: mass-crunching data and finding signal from noise in a way that we're not designed to do at all.

Well, just thank you to Justine and Xander for pointing this out to me. I'll also mention that I checked, and it turns out Wikipedia has a budget of 170 million a year. About 100 million of that is labor, paying everybody to do this work. Some of it's voluntary; a lot of it is not. All right, let's move on.

I also want to give a shout-out to Jimmy Wales for creating Wikipedia and managing it for all these years. What an unbelievable gift to humanity.

We're watching the transition from the Encyclopædia Britannica to Wikipedia to Grokipedia. Grokipedia was low-hanging fruit. Any of the AI companies could have taken this on, and I think it's going to become pervasive.

I know I'm standing up a new website for diamandis.com, and the very first thing I'm putting at the top is my Grokipedia link. You want to go deep? There it is.

So, true to AWG's vision, this is an important conversation. And, Alex, I actually read this paper. This is a new AGI benchmark, which gives ChatGPT a 57% score, but I would prefer if you explained it because it's pretty amazing.

This one's for you, buddy. Finally, we have a definition of what AGI is and how to measure it for the first time ever. This was a paper that was co-authored by Eric Schmidt as well. It's a pretty powerful concept. Alex, would you take us through it?

Sure. There's an enormous cottage industry of AI researchers trying to define what intelligence even is. I've been guilty of that in past years as well. My bias has always been to look for a universal, elegant definition of what intelligence is that isn't necessarily grounded in human behavior or human psychology.

This paper—and, as Peter mentioned, we know a number of the co-authors on this paper—the basic idea behind it is to do the exact opposite. Instead of trying to look for some human-agnostic definition of intelligence so that we can build more of it, the idea is instead to look at human psychology.

There's a theory that's popular in human psychology called the Cattell–Horn–Carroll, or CHC, theory, which decomposes human intelligence into 10 different factors, like the ability to reason quantitatively or to do visual processing. The idea behind this paper is to define a benchmark that's directly inspired by the CHC theory, to decompose the intelligence of frontier models into 10 different categories, with various subtasks associated with each category.

The main upshot of benchmarking GPT-4 and GPT-5 Auto—critically, not GPT-5 Pro, according to my reading of the paper—is that, surprisingly, intelligence is jagged. The frontier models are stronger at some skills and weaker at others, whereas the archetypical human would perhaps have a more uniform distribution of their skills across these 10 categories.

I would add the important caveat that, again, just based on my reading of the paper, they didn't actually benchmark the bleeding-edge frontier models like GPT-5 Pro.

For those looking at this on YouTube, you'll see these 10 different categories. These are human-like skills: knowledge, reading and writing, math, reasoning, working memory, memory storage, memory retrieval, visual, auditory, and speed. They're benchmarking GPT-5 and GPT-4 against those.

But it's a measurable benchmark, right? I mean, the other option is the pornography definition: we'll know AGI when we see it. Dave, what do you want to add on this? I'd love to get Alex's take, because I assume a 10 on each axis on this radar chart is human. So you're trying to match the outer ring here, but it's really—

It's all humans. I mean, no human is going to match 10 on this. Maybe—

Quote-unquote, well-educated adult.

Okay.

Yeah. Okay. The difference between the best human and an average human is a rounding error in the grand scheme of AI. It's almost identical, actually. But it's so asymmetrical, and I don't understand the memory and storage access axis and the speed axis. You say the best AI is miles behind humans in speed, and I don't quite get that.

That was the part that struck me also. Speed seemed off.

So when you make a query, your AI goes off and thinks about it for a while before it comes back with an answer. Is that speed? A human, if I ask a question, Alex will typically not go away for 5 minutes and think about it. He'll give me at least his version of an answer right off the bat.

Memory I found incredibly perplexing, because I thought these AIs have incredible memory. Alex, what's your take on speed and memory?

Yeah. Based on my read, the memory-storage access—or the deficiencies thereof—corresponded roughly to the fact that off-the-shelf vanilla language models and foundation models have a finite input context window. If you ask them questions that reference older information, by default, unless there's some sort of compression or memory compactification or RAG-type mechanism, they don't have the ability to remember things that you told them a long time ago.

But again, I want to caveat this benchmark. I love benchmarks in general, as I've mentioned previously on the pod, but these are off-the-shelf models without any agentic scaffolding on top of them, without prompt optimization, and without even access to bleeding-edge reasoning efforts.

This is just GPT-5 Auto. So I'm wary to put my finger on certain deficiencies as being in any way indicative or instructive of the limitations of AI.

Meaning, there are other models that would perform much better on these 10 parameters?

Or light modifications of existing models that, as with RAG—retrieval-augmented generation—make them superb at certain skills. So I think that where this is useful, in my mind, is just having yet another benchmark as a proxy. It's a start—an important start—for measuring human capabilities against AI capabilities.

But when I see 57%, we've talked in the past on the pod, in the style of Ray Kurzweil, about how the moment you've passed 10%, or maybe even less, you're basically halfway there. Getting 57% on a general human psychological benchmark indicates to me that, probably with a little bit of reinforcement learning, a little bit of bureaucracy, an agent framework, and scaffolding, you probably get to 90% today.

Amazing. Well, by the way, everybody, just some forward-looking news: Ray Kurzweil is going to be joining us on the pod next month.

Talk about his predictions for 2026. Yeah, it's going to be a lot of fun.

I have my standard responses to this, which are: I think this is really great for approximating or getting to the kind of frontal cortex and neocortex activities, but it doesn't deal with emotional intelligence or spiritual intelligence, or any of the other dimensions of intelligence that we typically attribute to human beings.

But I thought about you, Salim. I thought about you specifically on this one because it's going to be a measurable benchmark that we can at least point at, and we're going to discuss whether we're going to hit AGI according to OpenAI in late 2026, 2027, or 2030.

What I'm saying is, I disagree with the premise because AGI, for me, would incorporate these other things.

Okay.

So if you're measuring pure IQ-test-type stuff, fine. This is a great benchmark, and we can kind of—I wonder if we're going to have the first AI spiritual leader who proclaims a religion and leads it.

Oh, I think that's very doable. You know, I remember once spending time in the Himalayas with some of these gurus, right? I sat with them, with the orange robes and the long beards, and I came out with the conclusion that there are about 10 or 15 questions, like, "What is the meaning of life?" If you have a pretty good answer for those 10 or 15 questions, you can become a guru. That's kind of an LLM. That's your neural network. So it's only going to be hard—I think it's very doable.

By this time next year, there's going to be an AI-based religion that is going to scale at a hyper-exponential rate. It's going to be amazing.

All right, big news this week: OpenAI restructures to become a public-benefit, for-profit corporation and a nonprofit. The OpenAI Foundation will hold a $130 billion stake, 26% of the new company. The OpenAI Group is now what's called a PBC, a public-benefit corporation. Salim and I did that move with Singularity University, converting it from a nonprofit to a for-profit and spinning out a—

Exactly. This was a benefit corporation with a nonprofit alongside.

And so, here's the ownership point, actually: a PBC can do anything a C corp can do. It can go public, raise money, and be profitable.

This is a great point that you're making, Dave. For the viewers, from a taxation and legal perspective, a PBC is exactly the same as a C corp. It just—

Which is every other public company.

The distinction is that, in a C corp, the board is obligated toward financial optimization and can be sued if it's seen as not doing that. Whereas in a PBC, the board is obligated toward whatever the mandate is of the PBC and can be sued for that, in theory.

I love the percentage ownership here. So here we go: Microsoft owns 27%, the nonprofit owns 26%, and the remaining 47% is owned by OpenAI PBC. This restructuring is going to allow OpenAI to go out and raise money.

But here's the rub. Elon's lawsuit against OpenAI remains active, and his bid to try to block the restructuring was denied in court. But the case will proceed to trial in the spring of 2026, from what I read. The implications are interesting, right? Number 1: the court could order a rescission that unwinds the OpenAI for-profit PBC structure and restores nonprofit control. Number 2: key deals, such as revenue sharing with Microsoft, could be voided or renegotiated. And number 3: there could be potential damages and reduced fundraising flexibility for the PBC. So that's going to be interesting drama a year from now.

Yeah, Alex, if you look at the valuation of the company, the market does not believe any of those problems will actually be material.

For sure.

So it seems unlikely, but I think Elon has a very valid point in that, you know, that whole time you're a nonprofit, you're not paying any tax. And if you're secretly building a massively profitable, trillion-dollar company while avoiding taxes, that's a terrible precedent. You can't do that. Elon even said it: if that were legal, everybody would do that and start your company as a charity.

So I think the courts will have to say, "Yeah, you can't do that." The penalty could be like a dollar or something, just like they did with the Microsoft antitrust case. You're like, "Okay, you're guilty. You're totally guilty. You're fined a dollar."

$1.

Yeah. That was the whole Microsoft case: kill Netscape, destroy the entire market, and Marc Andreessen, you're out of a job. What's the cost? A dollar.

Wow. The same thing happened, by the way, in the 1950s, when Good Goodier and GM banded together and bought all the train tracks in LA and ripped them out. They privatized them and just ripped them out. In court, there was an antitrust case, and they got fined a dollar.

Okay.

I'd like to point out two possible societal goods here. One is that this results in one of the world's largest nonprofits being created, one that now has the backing of a frontier lab. The stated goal of the new OpenAI nonprofit—one of its first goals—is to spend $25 billion using AI to solve disease. I think that's a tremendous societal good.

We've spoken here in the past about how AI has the potential to solve disease and biology in the next 5 years. I think this is another arrow in the quiver of making that happen.

The second societal good: one of the things I worry about is what happens if a private frontier lab develops incredible superintelligence and decouples from the human economy. I think putting OpenAI on a trajectory where it can reasonably be expected to go public sometime in the next 2 to 3 years—I think an IPO by OpenAI and other frontier labs, and putting the equity in the hands of retail investors and index funds, is almost certainly a net societal good because it keeps the economic interests of large chunks of humanity aligned with these frontier labs, and vice versa.

Well, corollary to all that too, Alex, I think that everyone's like, "Hey, Brendan Foody of Mercor—he's a billionaire at age 23." He spent an immense amount of time inside OpenAI's building. We saw this in the lobby the last time we were there.

If you think, "What's my life mission? Am I starting a company? Am I changing the world? Am I solving all disease?" Regardless of what your life mission is, think about the impact of $25 billion of charitable money just to solve disease. What about the other $100 billion? Where's that going to go?

So if you're involved in this in any way and you don't have a strategy for how you interact with OpenAI—how am I in that building? How am I relevant? How am I going to be there when they start turning to commercialization and goods through the AI engine? How do I interconnect with that?—I ask all these entrepreneurs, "What's your OpenAI strategy?" A lot of them have no answer.

But you think about the scale of just what Alex said.

Yeah, this will be the largest nonprofit in terms of capital base, and it will be even bigger. It'll reach a half a trillion dollars. Dave, you remember you and I met with someone—I'm not going to say who it is because I don't think it's been officially released—the individual who is a co-founder there who will likely run this OpenAI Foundation.

We were talking about potentially spinning up some XPRIZEs as a means. He was very knowledgeable about these ideas of incentive competitions to leverage capital 10x. We just learned this year that the numbers from XPRIZE are that we leverage every dollar in the purse by 60-fold. So imagine if $100 billion becomes $6 trillion of leverage. What a fun time ahead.

Yeah. Well, think about the scale too, which is exactly a great point. A normal big XPRIZE is a $100 million prize, and here you're talking about $130 billion, which, if the stock goes up post-IPO, could be $260 billion. So all they have to do is sell some shares and fund a $100 million prize. They could do that every day of the week.

I dream about—

I dream about having 10 $1 billion prizes for the 10 biggest problems. It would steer where students spend their time, where founding partners focus on building companies. I mean, it would be sort of a flame to the entrepreneurial moths out there.

Can I take the other side of this just for a second?

Sure. Sure. Virtue signaling, greenwashing. They're putting all this money over there and then going full speed toward the IPO, hoping that the good they can do will balance the crazy path to greed.

Do not notice the man behind the screen. Speaking about greenwashing, here's our next story, with Sam Altman turned green here—a little bit of Shrek in his DNA. The title here is, “OpenAI plans a $1 trillion IPO and to spend $1 trillion a year in AI infrastructure.”

I love this. We've said this before: a trillion dollars here, a trillion dollars there. It's becoming a word far too popular these days. The speed is incredible. So, OpenAI—

Use your word, Peter. It's incredible.

It is incredible. You also have to hand it to Sam for the unbelievable sheer gumption to just go for it.

Yes.

Wow. No. Amazing.

Just incredible.

Well, let me put some numbers behind that. Here he is: “I'm going to spend $1 trillion a year. I do a $1 trillion IPO, and then I'm going to spend $1 trillion a year on data centers.” Your actual revenue today, dude, is $13 billion. Now you're saying you're going to get to $100 billion.

The equivalent would be if you had a household with $100,000 of income and your husband or wife comes home and says, “Honey, we should spend $10 million a year on houses and stuff.” That's the equivalent metric. Just to put it in context, that's the gumption, like you said, behind this claim. But hey, he's done everything he said so far, so it's plausible.

I'll try that with Lily and see how far I get. [laughter]

Well, you did that when you mortgaged your house to buy Bitcoin, which in retrospect turned out to be a good idea.

We didn't put all of it in, unfortunately. But still—

You should mortgage it again. [laughter]

Alex.

I think it's actually a pretty tiny number. Global GDP is upwards of $100 trillion, so just saying we're going to spend 1% of global GDP on AI infrastructure—

Yeah, but one guy, Alex. [laughter] He's one company.

One company, right? So if you have five frontier labs each doing that, that's still 5% of global GDP. I think this is a drop in the bucket, and that's before AI starts to radically grow the global economy. This feels on the low end to me.

We saw this last time compared to the railroads or telecom infrastructure: the AI build is still, as a percentage of U.S. GDP, on the low side. But here are the numbers. OpenAI is working on the largest IPO in history, with targets to do this in 2026 or 2027. The other point made here is that they're planning to build 1 gigawatt of capacity per week at $20 billion per gigawatt. There are 52 weeks in the year. That's $1 trillion a year, which is pretty extraordinary.

Just to give people some context, a gigawatt is enough to power the whole of Dallas–Fort Worth. It's a truckload of energy.

Mhm. Incredible. There we go again. Sorry. Time freak.

I love this article, Alex, that you found: “Claude shows signs of introspection: a model that is partially self-aware.” Claude Opus 4.1, take it away.

Really interesting paper. You'll recall that historically, the best suggestion floating around the AI research community for diagnosing self-awareness was maybe to train a model on the internet, excluding any notion or mention of self-awareness, and see whether the model is then able to articulate something about self-awareness.

I think that original proposal was probably somewhat impractical. This is a far more practical diagnostic for self-awareness. The idea is basically to take the internal hidden activations of a model and graft on an external thought—sort of incepting an externally imposed thought onto a model—and detect whether a model is able to recognize that it's having external thoughts intrude upon its internal activations.

Isn't that called psychosis? [laughter]

I think that might be slightly different. I think this is closer to some sort of telepathic forcing. You're taking an external activation vector and forcing it upon the internal hidden activations of the model, and then checking whether the model realizes that it's being externally influenced.

Pretty remarkably, some of their stronger models, Claude Opus 4.1, were about 20% of the time able to articulate not only that they were being externally influenced through this sort of vector-activation injection, but were also able to reasonably well articulate precisely the nature of the external thought that was being forced into their internal streams.

So, the question is: What does self-aware mean? It understands it's an AI model.

The proposal in this paper is that self-awareness means that the model is able to think about its own thought. It's able to understand what its own inner thoughts, if you will—its own inner activations—are and is able to reason based on that.

Just a high-level point for the neural-net geeks out there, too: this research can only be done if you have access to the internal weights and activations of the neural net. So, it's done inside Anthropic. Now that Meta is not going open source, you have to actually use a Chinese model to do this kind of research or you're screwed.

Which is really very sad, because I think that, before I switched to computer science at MIT, I was in cognitive psychology. I think that experimenting with the parameters and activations of a neural net will tell you far, far more about how a human brain works than the normal practice of sticking a little probe into a rat.

It's an incredible research playground. These ideas—what's the definition of self-awareness, and can I inject a thought?—are incredibly powerful. What you do is say, “Here's the neural net thinking about a very specific topic.” I'll grab the actual activations from part of the neural net, and then, while it's thinking about something else, I'll inject those and see if it's somehow complementary. Then, of course, the result is the neural net saying, “Where did that thought come from?”

So that's the introspection and self-awareness, but you can only do that if you can splice thoughts, which is an incredibly powerful, cool tool. And as soon as you go API-only, and you start—because on that radar chart we saw earlier, you're operating outside the neural net and trying to define AGI from outside the neural net—you become so much more powerful operating inside the neural net.

But we may be in danger of losing that as a tool. Hopefully, the Chinese models will keep coming out, and Alex is warning me against using them too much. Salim, final word from you.

This reminds me of Hod Lipson, who is a professor at Columbia. He builds self-assembling robots and evolutionary robots that have a feedback loop to improve themselves. He actually tried out his approach to self-awareness, which was to ask the AI what it would look like in 5 years.

By the feedback loop of constantly forcing itself to go, “Well, who am I that I might look like something in 5 years?” he thought that would generate self-awareness. He thought that's what happened on Facebook a few years ago when they shut it down. That question, by the way, is blocked in all the major models, but somebody will do that to DeepSeek, and you'll get to that same point.

I remember Dan Barry talking about the frog. We may have talked about this on the podcast before. He's watched a ton of free-floating animals in labs at NASA, and his opinion of self-awareness was: frog.

And we're like, “Frog?” He goes, “Well, a mosquito is an automaton. It doesn't really know it's a mosquito. A dog definitely has self-awareness; it knows it's a dog.”

For him, the boundary condition was a complexity of about a frog, where, in his opinion, a frog kind of goes, “Oh, I'm a frog.” Above that, more; below that, less.

Let's move to Alphabet. An incredible quarter for them: they topped $100 billion in quarterly revenue for the first time ever. Google Cloud grew 34%. Good on Alphabet and Google—they're rocking it.

A couple of other elements that Google and Alphabet have announced include a new marketing tool, which I love, called Pomelli. This is Google's AI marketing tool. Let's take a look at this video, and then we can discuss it, because, again, Google provides all of these incredibly useful end-user tools that make them so powerful as a company.

So Pomelli will understand your business DNA and prompt your own campaigns or get suggestions.

All right. Bottom line.

Yeah. Bottom line, this is an AI tool from Google DeepMind that helps small businesses create on-brand marketing campaigns. Pomelli will analyze your business website, learn the tone, the color, and the style to create ads that match that brand, and generate ready-to-use posts that can be edited in the tool. So they’re basically helping their customers who are advertisers do better advertising. I think it’s super smart. Any comments on this, Dave?

Well, before Google had AdSense, they thought they were going to hire 10,000 salespeople and be kind of like eBay. Nobody remembers all this, but there was a very smart original employee there, an Iranian guy, who said, “Hey, why don’t we create an auction marketplace? People can just come and bid on Google, and we’ll open it up to the economy. We’ll democratize it. We’ll make every entrepreneur in the world able to thrive along with Google.”

It worked incredibly well, and that created the Google we see today. They’re going to do that again with all these capabilities.

Gilad Elbaz created that engine, right? Incredible.

Yeah, incredible. As they roll out these capabilities, you’re just drinking it in.

I just want to point out that the latent trading ability of people in the Middle East is off the charts. When you apply that to deep-internet paradigms, that’s kind of incredible.

What this struck me as was another example of interfaces. In our book, Exponential Organizations, we have the concept of interfaces, right? Google AdSense succeeded because you automated the supply side and the demand side of the ad business, and this is now pushing the boundaries of that further and further into the creative process.

Just to comment, Peter, on this as well, I think the elephant in the room here is that the visual ads being generated are not being generated pixel by pixel. I’ve spoken on the podcast in the past about how, in the future, user interfaces are just going to be every pixel purely generative.

In this case, it’s almost charmingly retro, in the sense that it’s not pixel-by-pixel generated. It’s vector graphics, images, and photos clipped from the original underlying website. I think the elephant in the room is that it’s not purely generative, which means that it’s going to be ultra-low compute cost to generate. That suggests we may live in a very near-term future where display ads on the internet are generated on demand, because it’s relatively compute-intensive at the moment to generate a custom image pixel by pixel for an ad, but it’s relatively lightweight.

I love that, especially as agents are cruising all of my tabs on my search engines and listening to my conversations. They know exactly what I want in that moment and can generate an ad to influence me, until such time as I just give my AI permission to do all the buying, in which case it’s game over for advertising.

Charmingly retro in all of this. Hold on, Peter. You’ve hit on something unbelievably huge here. This is all assuming a human consumer, right?

Yeah.

Very quickly, we’re going to go through that.

Yeah. I remember having this conversation when we were advising Procter & Gamble. We did a workshop with them, and they said, “We spend a huge amount of R&D on what color the Pampers box should be to attract somebody’s eye at what level.”

I’m like, “Well, my wife has an Amazon subscription to diapers and doesn’t care what the box says anymore.” They’re like, “Huh?”

It’s just the dissonance between the old way and the way that you’re talking about. Once we have our own AI interfacing, it changes everything. I think that also has to be taken into account, so maybe this is just a short-term thing. Jarvis will essentially take over.

Jarvis will buy everything I need because it knows when I’m running out and it knows what the best quality is. It doesn’t really care what the ads say.

All right, let’s move on in the Google-verse here. We’re seeing Google AI Studio introduce vibe coding. Vibe coding is now available on Google AI Studio. No coding or API needed. Dave, do you want to take this one, or Alex?

Well, it looks a lot like Replit and Lovable, so we’ll see how that shakes out. This is the big guys stepping on the toes of the little guys.

Yeah, I used it. It was a fantastic experience, and I think it’s somewhat differentiated from the in-browser vibe-coding experiences from OpenAI or Anthropic.

For one, it creates multiple files. If you ask it to create an app, it’s not just fixated on a single self-contained file. It can create multiple files of code, which is very important for certain sorts of apps.

I ran it, of course, through my favorite evaluation for vibe coding: “Create a visually stunning cyberpunk first-person shooter.” It created a visually stunning dashboard, sort of an intro-lobby dashboard for the FPS, but I had to prompt it to create the rest of the game. What it did create was visually stunning, and I think it’s a promising first step.

I’m curious what the interaction with Replit and Lovable will be. We spent a few days with Amjad Masad, the CEO of Replit, and I’ve been playing with Replit on my phone and computer, vibe coding different apps, which is fun.

One of the things I had a long conversation with Jack Hidary about while we were in Riyadh—and one of the things that Jack said, which I love—is, “Every morning, instead of becoming a consumer, become a creator.”

Usually I get up and I’m reading all of Alex’s texts, all the breakthroughs that he found last night, and I’m constantly, as all of us are, consuming hundreds of articles over the course of the week—maybe 20 or 30 per day.

Jack was like, “No, no, no. Every morning I’m going to vibe code something. Every morning I’m going to create something.” I think that creator mindset is so critically important for us to be using. A conversation with your AI and creating something every day would be super fun.

All right. What are you creating? I was creating an app on my phone last night to remind me to take my pill packs, because I have 5 pill packs a day. It will now text me in certain windows and remind me, “Did you take your pill pack?” Then I can dismiss it if I did. It’s sort of an agent adjunct to my health.

Okay, moving on. Let’s go on to the chips and data-center wars. A lot is going on here. Again, a trillion here, a trillion there.

The first story is that Samsung is building a facility with 500,000 NVIDIA GPUs, automating chip manufacturing. This is an AI megafactory that will combine NVIDIA’s Omniverse with Samsung’s chipmaking for up to 20 times faster performance. Blackwell chips have generated $500 billion in business so far. Again, a nice chunk of change. Alex, what do you make of this one?

This is what recursive self-improvement looks like, Peter. This is GPUs and AI being used to optimize chips to make more AI. There are so many applications, ranging from computational lithography to fab optimization, for this.

When I’ve spoken in the past of the innermost loop of civilization looking like some linear combination of chips, robots, data centers, and power sources, all of this is what the innermost loop of civilization spinning faster and faster looks like.

This is the economy, right? That innermost loop is the economy going forward—certainly the future of the economy.

Yeah, I did some reference checking here. Five hundred thousand NVIDIA GPUs will draw somewhere between 0.25 and 0.4 gigawatts of energy. There’s no single site in China or the West that compares.

The top centers in China are topping out at 10,000 to 35,000 GPUs. In the U.S., clusters like Azure and Meta are ranging from 30,000 to 55,000, and this is 500,000. Wow.

Does this mean the machines are now basically manufacturing their own evolution? Is that where we’re going?

Absolutely. Yeah, that’s the right way to think about it, for sure. The fabs themselves have always been automated. They’re all roboticized on day 1. But it’s the periphery around that—turning it into a data center or feeding the front end of the fab—where there’s a huge amount of investment opportunity to close the inner loop.

Peter really should write a book, if books still exist, with that title.

There’s so much leverage in the inner-inner-inner loop. If you focus on where the bottlenecks are in the innermost inner loop, you’re going to find that it’s the chip getting out of the fab and into a data center and actually doing something useful. That’s where most of the bottlenecks are now.

Huge robotization and energy.

Yeah. And then there’s the feeding on the front end of it.

Yep.

All right. Here is a fun conversation and article: “Extropic Creates Thermodynamic AI Chips, Combating Industry’s Energy Crisis.” This comes from a friend of ours, Guillaume Verdon, who’s been on my stage at the Abundance Summit. He’s been on our podcast here, and he’s talking about a breakthrough in hardware—a technology called thermodynamic sampling units, or TSUs, that use 10,000 times less energy than GPU-based systems, using probabilistic bits. Alex, what does that all mean?

Well, I read Guillaume’s paper, and I’m a huge fan in general of trying to get closer and closer to the physical limits of computing.

Seth Lloyd famously, more than 20 years ago, discovered or reported that the ultimate physical computer would probably, at least for serial computing, look like a black hole—a black hole as the ultimate supercomputer. So, I'm a big fan of approaching the physical limits of computing.

In this case, though, my worry is that there's such a sordid history of probabilistic computing approaches being attempted and failing to keep up with Moore's law and algorithmic improvements. This is my worry. I want to believe—I want something like this to succeed—but I'm not super optimistic that this isn't just going to get steamrolled by algorithmic advances and advances in good old-fashioned CMOS digital logic.

It looks too much like analog computing, probabilistic computing. And remember, even generously, a 10,000-times energy improvement, at the rate that models and algorithms are advancing and the rate that good old-fashioned digital CMOS is improving, may only be a few years of headroom, which a new architecture would need anyway to get off the ground.

But can't we view it from the energy perspective? Because reducing the energy requirements on Earth 10,000-fold seems staggeringly beneficial.

In principle, yes. But in practice, the workloads that the economy demands have to be able to run on these computers in order for you to realize this hypothetical energy advantage. I think, for better or for worse, the burden of proof is on Extropic and Gil to demonstrate that his hardware can host workloads that are as commercially valuable as, say, B200 and NVIDIA workloads.

I saw Elon and Gil going back and forth on X, and Elon saying, “So, do you have something I should be looking at?” And Gil said, “Yes, let me show you.” So, we'll see if the Muskverse gets behind this technology.

I have a quick comment. Alex, can you go back to the black hole being the ultimate supercomputer? [laughter] You lost me right there, and my head's stuck now. Could you just go over that for a second?

Black holes are wonderful computers. They'd be a little bit difficult on the input-output side, especially the output side.

A little. [laughter]

But in principle—Seth and others have refined this notion—you can define a generalized notion of computation. In pure physics, it deals with how quickly internal state changes can happen inside a physical system. It turns out black holes are absolutely the physical limit, based on the physics that we have today, for the fastest serial computer, because state changes, in terms of their quantum state, evolve at the physical limit.

Programming them may be a little bit challenging. Maybe you'd have to fire in an X-ray laser or gamma-ray laser, and maybe you'd have to parse the Hawking radiation. But if you can solve input-output, a black hole supercomputer is the way to go.

I can see the title of this episode this week is “Black Hole Supercomputer Is the Ultimate.”

Black hole supercomputer on your desktop. [laughter]

Okay, let's move on. This is a fun article: Elon Musk on data centers in orbit. SpaceX will be doing this.

So, here we see in the image here, V3 of Starlink. Starlink Version 3 will be coming out. It will be delivering 10 times more capacity, 1 terabit per second, and enabling large-scale off-world processing.

I love this quote that you shared with me last night, Alex. I added it: “100 terawatts per year is possible from lunar-based production of solar-powered AI satellites, locally, and accelerating them to escape velocity with a mass driver.”

Basically, turning lunar material into compute and then accelerating it off the Moon with a mass driver—again, the work of Gerard K. O'Neill—and into Earth orbit. This is the beginning of a lot of things, Alex. We've been talking about Dyson swarms.

Let's talk about what we're talking about. We're talking about disassembling the Moon to build more computers—to build computronium—and the Dyson swarm.

What's more remarkable—I mean, just in the past few episodes of the pod, I've been beating the drum for how, you know, mark your calendar now: we're at the very beginning of the construction of the Dyson swarm. Maybe I was overly pessimistic. Maybe we're actually going to see multiple competing Dyson swarms, and SpaceX is going to launch one.

You'll see other companies, maybe other frontier labs, launch competing Dyson swarms. At this point, in the style of worrying about overpopulation on Mars, I'm starting to wonder whether I should instead be banging the drum for ensuring good interoperability between all of the Dyson swarms.

Let's take a moment. Freeman Dyson, a brilliant individual, said, “At some point, you're going to disassemble all the planets in the solar system and create a sphere around the Sun that captures all of its energy.” That's going to be the hallmark of an advanced civilization. That's called a Dyson sphere.

If it's not one sphere but a bunch of different satellites and computers, that can be viewed as a Dyson swarm. But the real fun concept is a Matrioshka brain. So, Alex, over to you.

These are 3 overlapping concepts: Dyson sphere, Dyson swarm, and Matrioshka brain.

A Dyson sphere was this notion of having basically—and there was even a Star Trek: The Next Generation episode that did this—a solid sphere at roughly Earth's radius from the Sun that lots of people perhaps could live on the interior of and enjoy nice environments and things. Probably not practical from a materials science perspective. The stresses would be enormous. That's a Dyson sphere.

A Dyson swarm says, “Let's, rather than having this be a solid enclosure that's rigid, let's instead have this be lots of orbiting satellites that are nonetheless collecting the energy from the Sun.”

A Matrioshka brain says, “Let's take multiple Dyson spheres at different radii from the Sun and have the innermost spheres consume the light—the solar insolation—at certain frequencies and then radiate waste heat outward to the outermost spheres, which then will consume progressively more and more infrared-shifted radiation and use that to power their compute.”

So, these 3 concepts—Dyson sphere, Dyson swarm, and Matrioshka brain—and also Jupiter brain is another popular depiction. These are all interrelated concepts. I think if we go the trajectory of taking apart our solar system, whether we brand it as one or the other, they're pretty similar.

And maybe a black hole.

Maybe a black hole is a Matrioshka brain circling a star, and we can't see the light. I wonder about this. If this is the fate of intelligent civilizations, I would expect to see more infrared-shifted solar systems elsewhere in the galaxy. To my knowledge, we haven't observed this.

That makes me suspicious that, even though I bang the drum for Dyson swarms, maybe there's something out there lurking in our technological future that will cause us to not actually need to take apart our solar system.

2 points. One, for those of you interested, Matrioshka brain comes from the Matryoshka dolls, which are the nested Russian dolls. So, you can imagine nested spheres around the Sun, each one absorbing energy, utilizing energy, and then radiating waste heat that becomes the input for the next sphere that it radiates to, and the next sphere, and so forth.

Second point is, we've got to get Elon back on the pod here to talk about this. I think it would be a lot of fun. All right, let's talk about energy and robotics, our final topic for today.

This is a big deal. California invests big in battery energy storage and leaves blackouts behind. It used to be pretty awful, and I remember this: we had rolling blackouts in California.

But the state has done something amazing: they've increased battery storage by 3,000%, going from 500 megawatts in 2020 to 15.7 gigawatts this year. The batteries store solar for evening demand, replacing underperforming gas plants. I'm glad to see this is happening. Any thoughts?

I think this is awesome. What a testament to—15 gigawatts is an incredible number of—

I had no idea this was happening.

No, it's not. Stay away from the word “incredible.”

Yeah. [laughter]

It's a stupid, trivial rounding case.

Stunning. How about stunning?

This is—we're going to get our subscribers drunk. So, interestingly—

Store gigawatts. Batteries store gigawatt-hours.

Yes, gigawatt-hours. Read the story underneath this. It's only got 3 hours at that power level. This is a joke.

This is just like—

This is like all of Alex and my interactions with government, trying to do things that sound important, that are these stupid little edge-rounding cases. At peak load for California, this is 1 hour of storage, and on a typical day they can store up to about a day of average demand.

All right, but let's take a look at the numbers. Blackouts have been cut by 90%, from 15 a year to 2 a year, which is great.

And here's the problem: we're going to see the CPI of electricity just skyrocketing. The price for electricity was 22.5 cents per kilowatt-hour in 2020. It's increased now to 32.4 cents per kilowatt-hour—a 44% increase.

If we continue to make the demands that we have on data centers, there's going to be a problem. The proverbial shit's going to hit the fan sometime soon.

Well, I agree with that. But California has done everything humanly possible to self-destruct at the government level, despite having the greatest tailwinds: the most incredible state, a massive state, all the innovation in the world, every advantage in the world, and the government claiming to do something good by piling up a bunch of batteries.

It's like, what? You're down to 2 blackouts a year. I mean, seriously. [laughter] That's our expectation of what we do. At least the 13% tax rate.

At least they're down a lot. It's ridiculous. There's something there.

I'll point out also, just for California specifically, if folks are familiar with the infamous so-called duck curve of California, where the demand for electricity peaks in the evening, right after sunset, and also in the morning. There is a mismatch between California, which is rich in insolation and solar energy on the one hand, and the need for early-evening power. I think even just a few hours of battery storage can help smooth out the duck curve, and that's transformative for California in a way that we here, perhaps in New England, don't have quite the same problem in our energy story.

Google is buying power from a revived nuclear plant. Google signed a 25-year deal with NextEra to buy power from a revived Duane Arnold nuclear plant in Iowa. It's reopening in 2029 to provide 615 megawatts of round-the-clock, carbon-free power. It's a $1.6 billion project.

It's interesting that we've got these hyperscaler companies buying energy. It used to be that this was something the government did. The government provided a distribution network for power, and you would buy it off the grid. That is no longer the case. Companies need to provide their own energy, so they're going all in on fission plants, SMRs, soon fusion plants, hopefully solar plants. Salim, you want to jump in?

And look, it'll create 400 jobs. [laughter]

Such negativity. This is huge. This is great. This is an important bridge to the near future.

I think what this shows is that we're basically dissociating the energy sources from the grid. Now we can have energy wherever it is happening—in a data center, next to it—and then leverage it. The marginal energy usage around the world will totally explode.

I can't wait for geothermal to really kick in here, right? There are so many places for it.

But wait, Alex, you're putting way more emphasis on it than I would have thought. Tell us why.

Well, I think bridges are important. Right now, the limiting factor for tiling the world with compute is, as a number of executives have recently pointed out, that we have the GPUs. The problem is having warm racks to put them in, as Satya Nadella said in the past few days.

I saw that Microsoft has hundreds of thousands of GPUs they can't turn on because they don't have the energy for them.

Yeah. I think having bridges, like reactivating otherwise disused nuclear plants, is an incredibly important bridge to the future until we get SMRs and fusion and maybe solar, and maybe new forms of natural gas.

Totally right. What Chase Lochmiller is doing is also the same kind of bridge structure, where you start with regular fuels, use natural gas or whatever, but it's steam-turbine generation right into the grid, right into the data center. You can reuse all that when you move to small nuclear, and then you move it to an SMR.

And then fusion comes online in 2030 or 2032, maybe.

You replace the boiler.

You replace the boiler. It's just, by far, the most efficient way to get to the ultimate end state—the Dyson swarm or whatever. But between here and there, that's the right stepping stone. It's much harder to do with solar because solar doesn't feed into a generator; it feeds into a battery pack. You're not reusing any of that when you move it to fusion in 2032.

I wonder—I love this. I wonder what takes 4 years, right? This is 4 years away. What takes 4 years to get an existing nuclear plant up and going? Are they retrofitting it? Are they updating it, or is this all paperwork?

It's very much all of the above, Peter, because we're advising Fermi America on this stuff. They're planning to do 6 gigawatts of gas turbine and 6 gigawatts of nuclear. It is very, very complicated to spin up a nuclear power plant.

That is the Alex loop—the inner loop of the inner loop of the inner loop. Focus on that. Peter, why is it 4 years?

You could build a Starship and go to Mars in 4 years. I don't understand why you can't get a nuclear plant up and going in 4 years' time.

To their credit, they've had an AI generate an S-1 in a record amount of time and got it out the door, so that's starting to happen. This is why that Sage project is so important, Peter: to rewrite policy as we need it. I mean, listen, when the Trump administration finally says, "We're going to accelerate this tenfold," that's when they'll get serious about energy production.

All right, these are some fun articles coming out. We're about to see the robo-taxi wars come online. Nvidia is planning a robo-taxi project to challenge Waymo and Tesla.

Here are the numbers. Nvidia is launching a $3 billion robo-taxi project in the self-driving-car race. This is a partnership between Nvidia, Uber, and Stellantis. For those of you who don't know Stellantis, they're behind brands like Chrysler, Jeep, Peugeot, and Fiat. They're one of the largest automakers in terms of building the parts, not just the brand.

The partnership will use an end-to-end AI system called Cosmos that NVIDIA has built to handle driving simulation. They want 100,000 robo-taxis launched by 2027. It's coming.

A hundred years ago, there was a 10-year period when we went from 99% horse and buggy and 1% cars to 99% cars and 1% horse and buggies. The question is, with all these players and all the capital going in, will this decade do the same thing? Thoughts, gentlemen?

I made a prediction 10 years ago that all driving would be automated. My son, who's now 14, would never get a driver's license. We have 2 years to satisfy that.

Yeah, mine too—my boys, too. Here are some of the numbers. Tesla robotaxis currently have 200 vehicles operating, I think, in Austin, and their plan is to scale up to 10,000 this coming year. I can't wait for them to be in Santa Monica, where I live.

Waymo has 700 vehicles, and by the way, I see them all the time as I'm driving around. I must see 20 or 25 of them a day. So, 700 vehicles is a pretty small number. They must have concentrations here in Los Angeles and up in San Francisco. There are 500,000 miles between collisions with Waymo. It is the safest player out there. NVIDIA is now partnering to go live with this.

I love that $3 billion is a stunning amount of money, and yet for NVIDIA, it's like a drop in the bucket, given the market. [laughter]

It is a drop in the bucket.

They don't even notice. It's like a little side project.

But this is where their GPUs are going next. They're going into humanoid robots and autonomous cars, right? They're automating the entire world around us.

That's right. I've spoken in the past about how this innermost loop is not going to remain contained inside data centers for very long. As I've noted in the past, the compute is literally going to walk out the door of the data centers. In this case, it's going to drive out the door.

For many people, these driverless cars—I have one, and surprisingly, many people I know haven't even driven in one or had the experience of driving in one—are going to be their first encounter with a generalist robot. It's going to be either seeing, driving in, or owning a driverless car. And it's not going to stop there.

I think the same stack that we're seeing NVIDIA push with its autonomous vehicles is going to generalize to humanoid robots on a timescale of 1 to 3 years. This is, again, the beginning of the expanded innermost loop of civilization that we're seeing.

On the flip side, here's the article from Uber's perspective. Their goal is 100,000 NVIDIA-based robo-taxis beginning in 2027, and this puts them in direct competition with Waymo and Tesla.

Today, in a number of cities, you can order a Waymo on your Uber app, which is fantastic. Any final thoughts on this story?

I think this is too slow for my tastes because there's such a huge demand for this. Even Waymo only has, I think, 2,000 vehicles, from what I looked up. There are 800 in the Bay Area and 500 in Los Angeles. We need tens of thousands of these things.

The good news is that each Waymo car replaces dozens of cars that are sitting around 94% of the time empty.

Yeah, most definitely.

Our next story here—I love this one. Foxconn is deploying humanoid robots at its Houston AI server plant. Check this out: Foxconn is expanding AI server production in Texas to half a million square feet, producing GB300s and Blackwell-series AI servers. It has a partnership with Digit, which is Agility Robotics. We're going to have the CEO of Agility Robotics on our stage this year.

See, we'll have…

At this stage, Abundance needs to be like 8 days long, Peter.

Well, it's tough. I'm trying to make sure that we have enough time for all of the community members to meet each other, hang out, have conversations, and have fun. But yeah, it's 4.5 days, but we'll have 4 robot companies there. I'm trying to get a 5th one from China. We'll see if we can get the new version of one of the top Chinese robots there as well.

So check this out. These humanoid robots are going to be driven by NVIDIA's Isaac GR00T N1 model. This is the innermost loop, isn't it, Alex?

It is. This is robots operating factories that make servers that go in data centers that power the robots. That's the loop.

Yeah, I won't say incredible. Not going to say it. Not going to say it. But wow.

I will say there aren't nearly enough companies working on all the different form factors. There's room for 1,000 more startups doing different variants of this.

For all the mechies out there who are wondering what they should do.

I also think this is a—

Pre-build a robot company that has an 8-armed, octopus-type robot and beat everybody else.

I also think this is a preview of how we get there. You were talking earlier, Peter, about Sam's forecast of a gigawatt per week. I think this is a plausible technical trajectory for how we get there. We're going to have robots building the fabs and the factories producing the servers and the data centers. It's going to be 1 massive flywheel.

All right, our next story in energy is that Blue Energy and Crusoe partnered to develop an advanced nuclear-powered AI data center. But I think the more interesting story here is that they plan to stand it up with natural-gas plants and then convert it to nuclear about 3 years after that. Don't wait for nuclear. Get it operating with what you can right now, and then retrofit nuclear when you can. The wait time for a new gas-turbine engine today is about 4.5 years.

It's crazy.

It's insane. We advise Siemens Energy, and they're sold out forever. Everybody's trying to go to dump heaps and get spare parts for gas turbines out of garbage dumps. It's really, really crazy right now, even in the recycling plants.

Wow. I want to talk a little about the China–US battle and hit a couple of different points here. Here are some numbers, and they're important to note: China is dominating production in a few different areas. 66% of electric vehicles are being built in China. 80% of solar panels and batteries are in China. 60% of wind turbines are in China. Those are staggering numbers. Add to that, on the innovation metrics, 70% of all global AI patents are coming out of China, and 75% of clean-energy filings are coming out of China.

This was part of the debate on our evening in Riyadh, where we hosted this dinner, and Cathie and Balaji were going back and forth on topics like this. Dave or Salim, you want to weigh in? We really ought to have a focused session with Antonio Gracias and Chase Lochmiller on how to deal with this in the US, because our whole investment cycle isn't geared up for this type of investing, and China is. It's very manufacturing-heavy: “We need more power, we need more smelters, we need more whatever.” We don't really do that well in the US venture economy, but that's all getting rethought right now.

Antonio and Chase are the guys right in the middle of it, so we should get them on the pod and brainstorm our way through how we're going to restructure. These are very expensive projects. They're not venture projects. They seem to work every time. The playbook is right on that prior slide. It's not a mystery. You have to be involved with the government. You need zoning, a location, permits, and all that. It's just a new format for American innovation, but it's going to last for 10, 20, 30 years, so you might as well get ahead of it. Imagine if Elon weren't doing what he does right now. These numbers would be far worse.

There's something—this is one of the flaws of our democracy: in 4-year, high-metabolism election cycles, nobody's thinking 20 years down the line. China may be authoritarian, but they can look out 20 years and say, “We need that much energy, that much water,” and do things to make that happen.

I actually had a hack for this. I did some brainstorming. Somebody asked me at a conference, “What would you do?” And I said, “Every 4 presidential terms, I would appoint a government, give them 10% of GDP, and say, ‘Your only job is to fix all the stuff that's long, 20-year-range projects, and then you're out. 1 term only, full authoritarian, go.’” Which is essentially some of what Trump is doing in this case.

So you just want to rewrite the Constitution. Okay.

It's a great thought, though, because flying back from Saudi Arabia and just looking down, there's nothing out in the desert for hundreds and hundreds of miles. Then you get to Europe, and it's just the most blessed Mediterranean green fields. Everything should be perfect in Europe, but government dysfunction is preventing them from any kind of involvement in what's going on right now.

It's a good case study: you can mess it up in a real hurry. We're not geared up to compete with China right now on this particular front, and it does need to be rethought. But if you don't rethink it, things can go really bad. Don't take it for granted.

So, guys, on one of our next WTF episodes, I want to bring some of the data that we found at FII9, the Future Investment Initiative event that we were just at. Some of the data is staggering about how the rest of the world looks at this, and I want to share that on an episode because it's really important. It's going to drive the near-term future.

All right, this particular chart comes from a tweet that Balaji put out, and he labeled it, “It's happening: The AI flipping is here.”

This is a look at who's making the open-source models. What we saw this summer was that open-source models were being dominated by China versus the US, and most definitely not Europe. I think what's most important here is that, as governments start adopting different AI systems, their ability to get access to free, open models versus paying for the models from the hyperscalers in the US—it's kind of a land grab going on. I don't have much more to say.

You know where this is really going to collide? I was talking to Brian Elliott over at Blitzy about exactly this topic. With Meta doing open source, we had a huge open-source option in the US, and then Meta fell off the grid. The last model was Llama 4, I guess, and it was terrible. Now they're trying to rebuild it, but they're rebuilding it closed source.

There are a bunch of projects for the military that Blitzy will ultimately end up working on that need to be air-gapped. You've got to use an open-source model in an air-gapped environment. You can't just go to the OpenAI API with super-proprietary government data. Right now, my only choice is Kimi K2 running on Groq with LPU chips and the Groq Cloud, which I think is a phenomenally good option, but it's all Chinese code, and God knows what's inside there.

We spent a bunch of time with Eric Schmidt. I want to play a short video of Eric from the FII9 event we were at last week talking about the US versus China, just to provide the US perspective.

Who's winning this—

Race? At the moment, the United States, without question. The US has a deep financial market that allows you to raise literally $1 trillion on a thesis and an idea, which is incredible. You have this massive buildout going on, and you have a real potential of solving hard problems.

Tell me how close China is to overtaking us.

It's not as close as I thought. I went to visit. China does not have the depth of the capital markets. They do have lots of energy, which we don't. They have lots of energy, but they don't have the depth of the capital markets, and they don't have the chips.

The capital markets—they haven't figured out a way to make all that money the way the US does. And the chips—they haven't been able to make the chips that the United States and others won't give them. That keeps them behind by a good chunk.

China, however, is focusing on exploiting AI in every aspect of its business much better than the United States. So I think the US will win on the intelligence race, but China is likely to win on the deployment race, and that's a problem for America and Europe.

All right, I want to jump into robots as our last topic here. 1X—we've had a great pod with Bernt Øivind Børnich, the CEO of 1X. Dave and I went and visited his factory. Let's take a look. They have a release of their Neo Gamma.

My name is Burch and today we're launching Neo, our humanoid for the home. Ain't no sunshine when she's gone. Neo is a humanoid companion designed to transform your life at home. It combines AI and advanced hardware to help with daily chores and bring intelligence into your everyday life. And this house ain't no home. As someone who lives with Neo every day, there is no experience quite as much.

All right. So, here are their commercials. They just went out. You can put a $200 deposit down on a Neo Gamma robot. They announced their pricing: $20,000 to buy it in early access, or $4.99 per month. You can buy it in 3 skin tones. I find that fascinating. [Laughter]

Check this out. On the right-hand side, I was walking to my workout gym, which is just outside my studio, and there was this giant sticker on the ground. This is incredible marketing. I have to hand it to them. They're doing a really super job on direct-to-consumer marketing.

Yeah, I pre-ordered mine. Can't wait to have the experience. I do think it's interesting that in many of the scenarios for the Neo, at least in the early days, according to 1X, they're going to be teleoperated. That may turn some people off. Having someone teleoperate into their home doesn't bother me at all. I'm very, very excited to try this out.

I think I got my order in when I was there at their facility. Bernt will be at the Abundance Summit, and he'll be bringing a number of the Neo Gamma robots. I'm just going to put one in my car and drive away at the end of the summit.

I'm getting mine in March. Peter, when we were out there, he was saying a $140,000 price point. I can't fathom how he's coming in at $20,000. Looking at the film from the factory, there's so much going on inside this robot. I cannot believe that they can get it out the door at $20,000. If I were running this company, I'd be subsidizing it. This is a data-collection play to get a huge VLA training set.

Yeah. Skating to where the puck is going to be. Elon said Optimus is going to be at this $20,000—not price point, but cost of goods—when they get to 1,000,000 robots being built and eventually robots building robots. Again, the innermost loop over here. This is competition against the prices out of China, the prices from Elon, and from Brett Adcock. So you've got to be competitive.

It's training people to expect price points that resemble, if you're purchasing it outright, a cheap car, and if you're leasing it, like leasing a car. This will be, for the American dream, so-called, of having a house in the suburbs, a car, and now a humanoid robot. It doesn't necessarily generalize that well to the rest of the world, but I think having at least 1 humanoid robot in your home becomes part of the new economy.

When the price gets down to $300 a month to lease—again, I've made these numbers; I say them every time—$10 a day, 40 cents an hour, everybody can afford that because your robot now becomes part of your earning potential. Your robot can go and do stuff for other people or for you. As Elon has said, this is all about creating the world of abundance. I love this story.

I've got a couple of quick thoughts. One is, I have a dog that literally looks like a teddy bear, and I'm wondering what would happen if the robot mistook the two. That would be one. I know all the models out there. At least in this case, with all the visuals and so on, it's not a kickboxing robot, which I thought was not a great marketing thing to say from the last couple of episodes.

I'm excited to see what happens here. Unitree has definitely taken a different approach, and I have to say that Optimus still has a hard-metal exterior. Neo Gamma from 1X came out with this soft, cuddly, warm, sweater-like look.

Very important.

And then Figure copied it. The latest release of Figure 03 has the same look. So, anyway, I guess, borrow from the best.

This is a story, Alex, that you shared: Toral. How do you pronounce it?

I think it's Toral.

Toral is an autonomous drone that ends mosquitoes. I love this. Can you imagine you're a tech entrepreneur someplace with a lot of mosquitoes, and you're just being bothered? You ask, “How do we end these mosquitoes?” And your answer is intelligent drones.

Here in this video, we see this very lightweight drone autonomously flying around. It's spotting the mosquitoes and zapping them with an electric grid that the mosquito flies through. It recharges and patrols 24/7 from its base station. It uses ultrasonic sonar to detect mosquitoes—the beating of their wings—and kinetic interception to eliminate them. This is like smart rocks in space or smart dust in space.

We're going to get nanobots. I think that will enable us to regulate ecosystems. I think, in the process, it's probably going to raise a number of bioethical questions. Peter, you and I and Dave have talked offline about near-term futures where, for bioethical reasons—or maybe even, dare I say, effective-altruist reasons—we're repairing butterfly wings on the one hand, but on the other hand, we have drones to obliterate mosquito populations. It's going to be a very interesting future.

It will, and it is interesting right now, being in the present. It's a super exciting time. I'm going to end this pod with thanks to Ruken, one of our subscribers and fans, who sent us a musical piece called “Don't Look Up: The Singularity Is Near.” I'm going to play it as our outro music.

But before we do that, gentlemen, any closing thoughts? Awesome episode. I learned so much today. This was amazing.

Yeah. There's no doubt the pace of stories is really accelerating. You've got to anticipate it'll 2× every month or two.

I wake up in the morning at 3:30 now, and I'm like, “What happened while I was asleep?” It's like—

Well, discussing a flying mosquito-killing drone was not part of my thinking for what I would be talking about today. So that's just—oh, we're trying.

What about putting a black-hole supercomputer on every desktop in every home?

That's also not part of it.

Yay. Or Matryoshka brains, you know, and taking apart the moon—

To get us a Dyson sphere.

And butt breathing was our article last year, our last episode, our new closing article. I think we should do that. I wish I had an unusual science closing piece or unusual tech. I'll take the mosquito-killing drone as our one for this week.

Or ending the moon. The moon had it coming. [Laughter]

Well, listen. I have to go there and start a city before we take it all apart. That's one of my goals.

Anyway, gentlemen, I love you very much. One of the best parts of being in Riyadh, Dave and Salim, was all of our fans there. Everybody kept coming up to me. This was a conference of about 5,000 people, sort of a World Economic Forum in the desert, and everybody was saying, “I love your podcast.”

So, Alex, you were sorely missed, and a lot of fans were asking, “Where's Alex? Is he joining this episode you're doing?”

Yeah. Got to invite me next time.

Okay. We'll do FII in March in Miami altogether and try to line up a live podcast from FII Miami.

Here is our outro music. Someone's frying data with a hint of smoke. The fridge is floating like a cosmic joke. My toaster in love with a crypto bro and entropy hums on the radio. The satellites gossip but who even cares? My cat just posted it all nightmares. Don't look up the singularity near. It's bary loud and clear. We built a god from electric dust. How it prays to us out of habit or trust. Don't look up. The code's gone divine. Heaven's a glitch in the command line.

Oh, the lyrics are amazing.

Amazing. A god out of digital dust.

Yeah.