OpenAI互联网浏览器已到来：ChatGPT Atlas，与Dave Blundin和Alexander Wissner-Gross对谈

OpenAI has launched a full-blown browser. The competitive positioning versus Google is basically all-out war.

Today, we're going to launch ChatGPT Atlas. This is an AI-powered web browser built around ChatGPT. We think that AI represents a rare, once-a-decade opportunity to rethink what a browser can be about.

Okay, great. But Google is going to come in and do at least this and take back any market share they lose.

I don't think we should think of it as a product. I think we should think of it as a distribution channel for OpenAI's superintelligence. Having a local agent mode, I think, is potentially transformative.

If Sam wins the data aggregation race, if he falls behind for a month or a year in the AI race, he still has your data.

We're going to have an AI that is our personal portal into everything. I'm not going to care what browser I use. I'm just going to be able to have a conversation with my AI, and it will pull up the data from wherever it is, whether it's using superintelligence from OpenAI or Google.

That's a moonshot.

Everybody, welcome to Moonshots, another episode of WTF Just Happen in Tech. I'm here with my Moonshot mates, Dave Blundin and Alexander Wissner-Gross. Good morning, gentlemen.

Hey, good morning. And a huge shout-out to the team. We were going to shoot this podcast last night, and Alex had so much material from the last 3 days that we just needed to get in here. Things are changing so quickly. Basically, the team pulled an all-nighter last night to pull together these stories, and it's epic. So, thank you to the team behind the scenes.

Yeah. Right now, our fourth Moonshot mate, Seem [?], is on an airplane. I just spent the last 4 days with him here at Calamigos in Malibu for XPRIZE Visioneering 2025, which is a story I want to open up with. Dave, I wish you were here. Alex, I wish you were here. It was awesome.

For those of you who don't know, XPRIZE gets together our brain trust and our benefactors every year, and we debate and discuss what problems aren't being solved that need to be solved, or what challenges are too far out and need to be accelerated and brought forward. That's Visioneering.

It was an amazing 2½ days—really 4 days in total—but 2½ days in which we raised $3.5 million in capital. Dave, you're on my board here at XPRIZE.

You do that every night. That's $1 billion a year.

That would be awesome. Someday we will.

Just so the audience knows the sacrifices Peter makes to bring you all of this information: He was onstage all day yesterday, and tomorrow he boards a flight to Riyadh. He'll be in Saudi Arabia, and he'll be onstage the day after that with Eric Schmidt, kicking off that event. That's 10 time zones away. Watch the footage of him from Riyadh and see what that looks like.

How wired will I be on caffeine?

Oh, my God. It's great.

We announced our impact report for XPRIZE yesterday, and the numbers are staggering. We have a massive, detailed report, and we found that for every dollar invested in a prize, we get a 60x return. A $1 million prize drives $60 million of R&D invested by all the teams. They're all optimistic. They all think they can win, and they're all part of a sort of Darwinian evolution to solve these problems.

I'm super pumped about that. But I want to report this because this is the first group to hear about it: We just determined who won XPRIZE Visioneering. We entered the 2½-day program with about 20 concepts. We have 5 different domains, 5 different grand challenge areas, and 4 concepts per area. We narrow it down to 2, and then down to 1, which leaves us with 5 that enter the battle royale, as we call it. We go from 5 to 3, and then last night we got down—well, let me just show you the numbers here.

The XPRIZE Visioneering winners for 2025: We were expecting just 1 of these prizes to get funded and go into development. It turned out that all 3 of these got funded to go into development. Let me mention what they are, because I'm very proud of them.

The first prize is called Abundance, and you have to love the name. It was actually proposed by 2 of our Abundance 360 members, who raised the capital to get it going. What is the Abundance XPRIZE? Deliver food, water, housing, electricity, and bandwidth to a community for $250 a month. That's the goal—everything that you basically need.

The conversation last night was about the potential for a lot of civil unrest as people start losing jobs and subgroups start becoming wealthier. We've talked about this. I'm absolutely clear that in the next decade we're going to have extraordinary abundance uplifting everybody, but it's this turbulent period over the next 2, 3, 4, or 5 years that's concerning.

The idea here is that if all of a sudden moms and dads have all of their bases covered—the basics of life for $250 a month—then they can start to think, “Okay, how do I use AI? How do I use this technology to be an entrepreneur and create a better life?” Any thoughts on that, Dave?

Especially that last fundamental of food, water, shelter, and bandwidth: If you're going to contribute to this global revolution, I love the fact that they added that as a fundamental necessity inside the $250 limit. That's just such a great idea.

It unlocks your ability to contribute, to make a living, and to get educated. All education will move to AI. Healthcare—all of that ties to bandwidth. It really is a fundamental necessity. I love it.

Alex, any thoughts?

Yeah. This sounds a lot like a universal basic services concept. UBS is sort of the symmetric dual to UBI, universal basic income. I'm very bullish on universal basic services in general.

I would expect it's an artifact of a mature economy that the cost of living can be driven down to near zero as part of a sort of lifestyle subscription—Amazon Super Prime, if you will.

I'm super excited about it. There are a lot of studies that say universal basic income backfires in that it causes depression, alcoholism, and drug use. But services, where you actually get the things you need to survive, still encourage you to work and contribute on top of the service. It's a much better idea.

We learned in our pregame here that Alex doesn't even use caffeine. I don't know how that's possible.

Caffeine is a universal basic service, for sure.

This won the most capital last night, and it's going into prize development. I'll report on it, and we'll have this team at the Abundance Summit. Both of them are Abundance members, and we'll talk about it.

The second prize, surprisingly, that got top honors and received enough capital to go into development is a Fusion XPRIZE. I was thinking, “Okay, there are 37 venture-backed fusion companies, and about $10 billion invested in fusion. What do they need a fusion prize for?”

Amazingly, I met with 4 fusion companies—solidly funded, ongoing fusion companies—as well as some of the top faculty, including 1 professor at MIT, and they said, “No, no, we need an XPRIZE to move this forward. We need the public to understand how important this is and how the government needs to come in and support it.”

This one isn't fully defined as a prize, but $500,000 was committed to develop it and move it forward into a potential prize. Alex, I think you have some feelings about this one.

I think fusion is already well capitalized, but ultimately, to the extent that the limits of economic growth are bound by our ability to solve fusion, I think on the margin it would be more helpful to allocate more capital toward fusion energy sources. Perhaps this helps with that.

Alex, you know what happens after this Visioneering phase: The world's greatest experts on the topic all get together in the Peter-verse, and they contribute all their ideas. Not all of them get from there to actually being a prize, but you learn so much about the state of what's happening along the way.

I love it when a topic like fusion gets through this part of the funnel, regardless of how it ends up, because the amount of information we'll bring back into the podcast will be immense.

The CEO of Commonwealth Fusion Systems, Bob Mumgaard, is going to be with us in Riyadh, and he's going to be onstage with me at the Abundance 360 Summit in March. I was on the phone with him getting ready for what we're going to be doing in Riyadh next week, and he said, “Listen, I heard that you're talking about a Fusion XPRIZE. I am so excited about that.”

Here we have the best-funded, most advanced fusion company actually excited about a Fusion XPRIZE. So, I'm excited to dig in further.

The third prize is something I love. It's called WALL-E. We'll have to be in debate and discussion with Disney about this, but here's the prize: Dump a machine into a garbage dump. The machine sorts the trash and generates piles of metals, foods, and paper. Basically, can we take our current landfills and actually reutilize them?

I have a way I would actually win it, but I don't know. I think this is a convergence of technologies.

It's going to be AI, robotics, and materials science. Any thoughts, Dave?

Alex keeps bringing us deal after deal after deal, and every one of them so far has been a winner. So it's really exciting. But, Alex, you brought us that rare-earth company. You want to talk about that? I learned a lot about this just studying that company.

Maybe just a broader comment on the space. I think there is such a long tail of physical-world service jobs that are ripe for automation, not just limited to repurposing junkyards, as it were. If you look around the world today, I often look out in the street and ponder, where are all the robots? We're supposed to be living in the future.

Why haven't we seen anything that looks visually transformative when you look out in the street? I think in the next 5 to 10 years, we will look out onto the street and see an abundance of robots and physical automation that enables communities to be visually transformed. Aesthetics that would otherwise be out of reach for an economy our size, as the economy starts to grow radically, will become possible as we deploy robots everywhere, even for the most minor tasks that would otherwise be economically inaccessible today.

I think this is just a special case of a much broader opportunity over the next 5 to 10 years: deploying automation everywhere.

Every week my team and I study the top 10 technology meta trends that will transform industries over the decade ahead. I cover trends ranging from humanoid robotics, AGI, and quantum computing to transport, energy, longevity, and more. There's no fluff, only the most important stuff that matters and impacts our lives, our companies, and our careers. If you want me to share these meta trends with you, I write a newsletter twice a week, sending it out as a short two-minute read via email. And if you want to discover the most important meta trends 10 years before anyone else, this report's for you. Readers include founders and CEOs from the world's most disruptive companies and entrepreneurs building the world's most disruptive tech. It's not for you if you don't want to be informed about what's coming, why it matters, and how you can benefit from it. To subscribe for free, go to dmmandis.com/metatrends to gain access to the trends 10 years before anyone else. All right, now back to this episode.

I just wanted a robot walking up and down the I-10 freeway, the 405, picking up the trash on the side of the road. But the idea that we can actually take our landfills, which have so many different problems—from methane production to disease—and we're sending so much of our trash overseas to Southeast Asia with heavy metals, is amazing. The idea that we can actually use it as a feedstock is amazing.

I don't want to belabor the point. Congratulations to the teams that won XPRIZE Visioneering. Congratulations to Nusan and Shari and the entire leadership team of XPRIZE. It was an awesome 2.5 days.

We recorded a podcast, which we dropped a couple of days ago. We had Immodust and Eric Pulier and myself on a podcast. Hopefully, everyone listening has heard that one. We're going to be expanding on some of the ideas because I want to make sure to bring in the brilliance and vision of AWG and Dave.

All right, let's move on. The main course today: AI chips and data centers, as it is every day. Do you want to introduce this video, Dave or AWG?

Yeah. So this is one of the reasons we needed to get together quickly. This just came out: OpenAI has launched a full-blown browser. The functionality won't blow you away yet, but the positioning—the competitive positioning versus Google—is basically all-out war.

I went back and researched this. Google launched Chrome. Chrome was not in the world; people don't remember this. They leveraged their user base to install it, and now they have two-thirds market share of browsers. This is people's point of contact with information; it goes through Google. They get to see everything you do.

Then later on, they turned on Chrome Sync, so they watch everywhere you navigate. All that information goes back into Google's great AI machine and serves you ads. Brilliant and kind of scary.

Sam, being the strategic genius that he is, says, "Okay, this is one of those fundamental Bill Gates-style points of control we absolutely have to play in the browser game." So we're going to launch the Atlas browser, and what's going to make it better than Chrome is that it's going to learn what you like and don't like far, far better and use our AI advantage to serve up better ideas.

The integration of GPT with what you're browsing will be completely seamless. It'll be advising you, taking you to the next website, and curating your news, all through that integrated browser.

It'll be taking your data.

Yeah, yeah. That too. I mean, that's the key, right?

Let's watch this short video of Sam and his team announcing Atlas, and then we'll talk about it.

We're going to launch ChatGPT Atlas, our new web browser. We think that AI represents a rare, once-a-decade opportunity to rethink what a browser can be about, how to use one, and how to use the web most productively and pleasantly.

There are 3 special core features of Atlas that Ryan's going to walk you through in a bit. The first is that ChatGPT comes with you anywhere as you go on the web. The second big feature is browser memory. The third, which we're really excited about and Justin's going to show later, is agent mode, which is in Atlas. ChatGPT can now take actions for you. It can do things.

All right. So my first reaction is, okay, great, but Google's going to come in and do at least this and take back any market share they lose. I don't know. Do you agree with that, Alex? What are your thoughts?

I think there's a misconception that Atlas is a product. I don't think we should think of it as a product. I think we should think of it as a distribution channel for OpenAI's superintelligence.

I think all of these discrete products are just going to dissolve over the next few years into a uniform medium of distribution for superintelligence. Whether it's one browser on the desktop versus another browser competing, I almost think it's the wrong question. The right question is: What form of backend superintelligence is being surfaced via which channels?

A browser is one. Intelligent code editor environments are another. I think robots and various wearable devices are going to be another over the next few years. I think it's really the superintelligence at the end of the day that's the differentiation, less the particular Chrome, if you will. That's just an embodiment of it to deliver it to the user.

Along those lines, the most interesting part of the Atlas launch for me was agent mode, less so the other features. Having a local agent mode, I think, is potentially transformative for a number of use cases and feels a little bit more sophisticated than prior agent launches that we've seen from OpenAI. If you remember Operator, or if you remember the cloud-based ChatGPT agent, this one is at least partially local.

So you've got the big guys with infinite budgets. You've got Google, you've got Zuck, and you've got Elon. But then you've got the 2 little startup, super-hyper-creative startup guys. That's Dario and Sam.

You've got Sam and OpenAI playing a very different game from Dario. Dario is relying on exactly what you just said: "I'm going to build a more intelligent fundamental machine, and because it's more intelligent, people will navigate to it, and we'll go out through corporate channels."

Sam is playing the old Bill Gates game where I'm not going to take for granted that my AI is better than Google's. But right now, I have twice as big an installed base as Google does. So what can I add to protect my position that makes me the default choice in the case where the 2 AIs are roughly on par?

He gets Jony Ive to build a device. He's building his own data centers with Broadcom, and now he's adding a browser. He'll add everything that Bill Gates would have added that's a user point of control or an entry point into the use of AI in order to defend that turf and encourage more of the innovation to come through him rather than work around him through Google.

It's like all that warfare all of a sudden between Google and OpenAI, and it's just really fun to watch.

Dave and Alex, my favorite model for this is still JARVIS from Iron Man. We're going to have an AI that is our personal companion and our portal into everything. I'm not going to care what browser I use. I'm just going to be able to have a conversation with my AI, and it will pull up the data from wherever it is, whether it's using superintelligence from OpenAI or Google.

The only thing I'd add to that is that very, very soon, that data will be your personal health data, your personal preferences—everything about yourself. When you have your virtual girlfriend or boyfriend, everything you like and don't like in life will be in there.

If Sam wins the data-aggregation race, even if he falls behind for a month or a year in the AI race, he still has your data. That personalization might create a much more compelling experience and allow him to catch up again. The personal-data warfare is kicking off in a huge way right now. You mentioned it a second ago, Peter.

What's the downside of what we see here with Atlas? I mean, we have the ability for OpenAI to not only look at the data you have on your browser, but probably every tab that you have open and everything you have going on in your computer. And they're not promising to keep it confidential. Thoughts on that, Alex?

I think we'll see forcing functions for greater forms of confidentiality and privacy, but I'm just reminded: Do you remember the browser wars?

Yeah, of course. Right. And Google won with 70% market share today.

Yeah.

Right. So there's sort of a long history of sleepy periods of relatively low innovation separated by Cambrian explosions of functionality. I remember all of the browser wars, and I think a browser war today over competing, among other factors, on whose browser is the most private while also being AI-agentic—I think that's a valid front for competition, and I welcome the competition.

Amazing. Alex, would you introduce this next slide here? You built a chess game. But before I play it, explain what you built here.

Yeah. So with computer-use agents, or CUAs, of which this new ChatGPT Atlas agent mode is arguably one example, I have my own evals. One of my favorite evals for testing these CUAs is to see whether they can win at simple or complicated single-player web games.

A favorite easy example is to see whether I can turn Atlas loose on a single-player—not two-player—game of web chess and see whether it can win. Historically, I've used this eval against Operator from OpenAI. What we're seeing here is a time lapse of it just being asked—I turned it loose on a single-player web chess game and asked it to win.

Interestingly, this is the best performance I've seen to date from a web-based CUA turned loose on a game. Sometimes I'll turn it loose on a game of web Civilization, if folks are familiar with the Civilization franchise. But in this case, intriguingly, it asked for hints, which I've never seen before.

It used the helpline built into the web game to ask for hints and was winning at the end of the day. I think this is a preview. In short,

Did it ask you for hints, or did it ask someone else?

It asked the website for hints once it discovered—which it did pretty quickly—that it could ask for hints. It asked for hints and used that to win the game. And I think this is a preview of CUAs for everything, not just winning easy games of chess.

Amazing. Amazing. All right, I'm going to jump into Anthropic. This is a conversation between Jonah Cool, who's the head of life sciences partnership and development, and Eric Chow Abrams, who's the head of biology and life sciences research.

You know, in January at the World Economic Forum, we heard Dario Amodei, the CEO of Anthropic, talk about one of his passions, which is the ability of AI to accelerate biology and longevity. Very famously, he said, "If we're able to hit the targets we have for AI, we could see the doubling of the human lifespan in the next 5 to 10 years," which perked everybody's ears up, including mine. Are we going to see longevity escape velocity within this decade? Increasingly, the answer is yes. Let's take a listen to Jonah and Eric having this conversation.

I'll start with why we're focused on the life sciences when we talk about the beneficial use cases of AI and all the amazing things that we can do in the world with the frontier AI that we're developing. Actually, the number 1 place that we at Anthropic are excited about applying it is within biology and the life sciences. If you read our foundational material, that's the primary area where we're really focused on delivering the beneficial impact.

We need Claude to be conversant with all of the tools that scientists are using every day. There's a whole ecosystem of important tools and partners out there that we are integrating with. We talk about Benchling on the experiment-management and lab-notebook side of things; 10x Genomics with Cell Ranger, an incredibly important platform for analyzing single-cell experiments; and then PubMed, for example, for being able to query the literature.

These are just 3 incredibly important partners in a much larger ecosystem. That base level is that we need to make sure that Claude can talk to all the major sources that scientists are using throughout their daily work. We want to bring Claude to performing at the level of a superhuman research assistant that can assist you as a scientist throughout all stages of your project. Alex,

I speak from time to time on this pod about superintelligence solving math, science, engineering, and medicine. I think this is likely how biology gets solved. I was talking a moment ago about computer-use agents, or CUAs. I think we're entering the era of CUAs for biology, where we have baby superintelligences that are completely fluent and well-versed in the tools of computational biology, are able to read PubMed fluently, and then even go and perform experiments. I think this is what solving biology with AI looks like.

Yeah. You know, there's a company I just recently invested in that I'm very excited about. It's called Lila Sciences—L-I-L-A. People can look it up. It's out of MIT and Harvard. George Church is the chief scientist. Jeffrey von Maltzahn is the CEO.

What they're doing in a similar fashion, but I think more advanced, is they've set up these science data factories. They have a superintelligence model they're building, and these science data factories are basically 24/7, lights-out robotic farms looking for information out of nature.

If you imagine it, the superintelligence will come up with a scientific theory or proposed research. They'll program the robots to go do the research at night, gather the data, bring it back, check their theory, iterate, put the next experiment forward, and run this 24/7 cycle to sort of mine data out of science itself, focusing on biology first and foremost, but also chemistry and materials science.

And I love this, as we're searching for new data out there in the world to help us understand what's going on in our 40 billion cells. You know, it's 5 to 10 billion chemical reactions per second per cell. We need to be able to reach in and get the data out to build our models even better.

I think that, as you might know, Peter, Jeff was a labmate of mine when we were undergrads at MIT. And I'm a huge evangelist for dark labs. I would like to see dark labs for everything.

Yeah. Well, and Jeffrey von Maltzahn—I know it's a harder name to find on the internet than Jonah Cool—[laughter] but definitely look him up. The guy is going to be huge. You can see it coming, and Alex will reaffirm this, but he will be one of the key figures cracking life sciences.

And I'll tell you what else. You know, we'll see later in the pod, there are some people saying, "Look, we've got to slow down AI. We've got to stop. It's not going to actually happen." We're going to move full throttle, and there are 2 reasons: one is China; the other one is this—people are not going to sit and let people die unnecessarily from illnesses if AI can discover solutions to that. That's not going to happen.

So that's why the AI labs are talking about this use case so much, because it's life; it's preserving lives.

Yeah. And, by the way, Jeffrey von Maltzahn and Lila Sciences will be at the Abundance360 Summit. I'm super excited for him to present. Our theme in March of 2026 at the summit is "superintelligence and the rise of humanoid robots." He said, "Okay, that's definitely a subject I want to cover."

All right, let's move on. Wikipedia says human traffic has been dropping 8% year on year. Fewer humans are coming to Wikipedia. We can dive into this. I'm still waiting for Grokipedia to come online.

Alex, what are your thoughts here?

I get asked the question a lot: How do we incentivize humans to create new knowledge in an era of generative AI? And I suspect the question itself is probably faulty. I think knowledge gathering is likely itself to transition to AI.

I think we'll see investigative reporting that's AI-based. I'm not losing sleep over human traffic dropping in an era when knowledge synthesis is abundant, but knowledge generation by AI is not yet abundant. I think AI-generated knowledge is right around the corner.

Okay, Dave, I'm going to go ahead, and then I have a rant on this.

All right. Well, this is right in my wheelhouse, so I need to wax poetic for a minute on this topic. I've been the founder of 20 direct-to-consumer AI companies. First and foremost, every time someone complains about their traffic going down, it's going somewhere else. It's not going away. Overall traffic is going up very, very quickly.

I'm involved in a company I can't name right now that's gone from nothing to $600 million of revenue purely from online traffic, $100 million of profit on the bottom line. And so, when Wikipedia says, "Hey, traffic is going down," it's going to some other place.

The formula for getting the traffic is well known now. First and foremost, you need to create huge amounts of AI-generated content, but it has to be good content. You also have to pay the man. You have to pay Google. You have to pay Facebook. If you do that concurrently with putting your content out there, they'll give you the traffic.

Also, you need to reformat your content so it’s easily readable and interpretable by AI—hence GEO at the bottom of the slide: generative engine optimization. In the future, people will not go to Wikipedia for their content. They’ll just ask the AI. The AI has all the information, but it still needs to be factually accurate and correct.

And so that role—and I’m a big Wikipedia fan—but, you know, I was at the Washington Post when it was getting obliterated by the internet, and it felt like, “Hey, we’re important for the country. We’re factual.” It doesn’t matter. You’re going away. And so that’s what’s happening.

My rant on this: I’ve been trying to update my Wikipedia page for literally 2 years. I hired consultants to update my Wikipedia page, and every time it’s updated, they bring it back to what it was. It’s so stuck 20 years ago.

I used to use Wikipedia. I don’t anymore. The ability for an AI to actually search the web and get consistent, relevant, and accurate information about me is important. So I think maybe Grokipedia will be a solution here, or in fact any AI that just says, “Spin up a page on Dave Blundin.” That’s going to be awesome.

I’ll give you one other pro tip. Get a Similarweb account at Similarweb.com, and you can see exactly where that user went. The guy who would have gone to Wikipedia yesterday—where did he go instead today? If you track where it’s all moving, replicate that behavior, and you’ll succeed.

Amazing. All right. The next article here is “GPT-5 Rediscovers Long-Forgotten Math Connections.” This has Alex Wissner-Gross written all over it. Dr. Gross, please tell us.

Peter, I talk frequently about how superintelligence is—and will be—solving math, science, engineering, medicine, and other fields. There was a lot of hand-wringing over the past week-plus about a specific set of math problems and whether AI in general, and GPT-5 specifically, were actually uncovering new math. I think this story beautifully encapsulates the fog of war we’re in right now.

The water level of intelligence is rising day by day, and some of the earliest open math problems to be solved, I think, will be math problems where the solutions were known to a subset of humanity but not to all of humanity. We’re going to wring our hands collectively as a civilization quite a bit over whether this open problem in math was really open, whether it was solved, or whether it was half-open—where some people knew how to solve it and other people didn’t know that it had even been solved. That’s the fog-of-war phase that we’re in.

There was a lot of discussion over the past week about whether this was a real accomplishment—a real discovery in math by AI—such as one of the Erdős problems, number 143, for example. But there was ultimately a lot of really revealing discussion and commentary on this particular problem and on other Erdős problems showing that this is just a phase right now.

In these early days, we’re still cleaning up house, as it were, in terms of understanding even which problems are open, closed, or somewhere in between. After this phase, I predict we’ll get to a phase where a lot of the uncertainty is reduced regarding whether a given problem is actually open or not.

Yeah.

Open means solved, right? You mean solved.

Open means unsolved. Closed means solved.

I think this is also a great little case study in how the academic world is like, “Well, this proves that it didn’t really solve it. It looked up an ancient...” When you’re trying to do something, you don’t care a wit how it solved it. It came back with the right answer.

This is a lot like ThinkStruct[?] in our lab. It’s a company that does academic research and now patent research using AI. Nikki Abate[?] and Julius Hidekutter[?] are actually turning out to have a really good hybrid of writing your patent application while doing all the background research for all prior applications and all prior knowledge. Those 2 things are integrated. This is where you’re seeing AI being superhuman, because normally you’d say, “Oh, well, research of old documents is this guy, but thinking of new things is this other guy.”

The AI doesn’t care. It just does both.

I’m so excited about the use of AI in writing up and submitting patents. Talk about something that is extraordinary. One of my favorite applications of AI and patents was the following.

This was a conversation with an Abundance member who was like, “I want to figure out how to use these technologies in my business.” I said, “Well, why don’t you just ask?” What I showed her was, “Here are 3 patents you’re interested in. Put them in the browser and say, ‘This is my business. How would I combine these 3 patents together to make a new product or service in my business?’”

Oh my God, it’s extraordinary. This is literally a creative engine.

All right.

Well, anyone who’s a real fan of this podcast by now has to have read Accelerando, because Alexander Wissner-Gross says it’s the best piece of writing in the history of humanity. If you heard that and then didn’t read it, something’s wrong with you. The very first chapter, the opening scene, is exactly what we’re talking about right now. The lead character makes a living with AI-generated patent filing.

Yes, consistently, and gives it away. Anyway, let’s not go there.

All right. Our next article here is “Uber Tests Microwork for Drivers to Train AI.” Uber is paying between 50 cents and $1 per task that can take 2 to 3 minutes and get processed within 24 hours. Is this sort of a digital TaskRabbit? What is this, Dave?

This is really, really cool, because Mercor is almost closing in on $1 billion in revenue, going all over the world, grabbing expertise and getting it into a format where the AI can assimilate it, and then the AI can be an expert in that topic, too.

You’ve got all these Uber drivers driving around. They’re sitting around a lot of the time. Do they have knowledge that may be a contributor back into the great AI machine? A lot of what’s missing is physical motion, common sense, and just all this information. Why not use that same platform you’ve already got to be another Mercor-type AI data-gathering machine?

Yeah, I think this points directionally to the future of the gig economy. The gig economy historically was focused on the physical world and physical tasks, including driving other people to their locations or driving food to a person’s location.

I think this points toward a near future where training robots to perform service-economy tasks is the new de facto gig economy.

Yeah. It’s so fascinating that Uber turned this way. It’s all about the relationships it has, right? It has a relationship with a large number of people who want to earn money on the margin. We’ll probably see other companies follow suit as well.

Well, you know, during COVID, Uber got annihilated and Uber did fine because they had launched Uber Eats. They’re very, very thoughtful about this. In fact, I don’t know if you remember Travis Kalanick when Uber was going public, but he got on stage and said, “Uber is not a ridesharing or hailing-cab company. We’re an internet fabric.” It was some really ethereal statement, but now they’re actually doing it. It makes sense in hindsight.

They don’t view their platform as being about cars and rides. They view it as—

We’re going to spend time with Dara, the CEO of Uber. He’s going to be on stage with us at Abundance360. We’ll talk about what he’s doing on the data side, but they’re also now partnered with Waymo. In certain places, you can hire a Waymo through Uber, and they’re hooking up, I think, with Joby on the flying cars—let’s call it that for the moment.

Uber has been an incredible platform for experimentation and the integration of various exponential technologies.

So that’ll be fun.

All right.

Alex, I’m going to turn to you on this one. DeepSeek is packing text into images. Talk about this, pal. What’s this significant transformation?

Yeah, this is a major advance from DeepSeek. It’s a new model that DeepSeek announced: DeepSeek OCR.

Maybe a bit of background first. Foundation models—frontier models like GPT—aren’t thought to perceive text in the way that humans perceive text. Humans look at text on a page and see text visually. The frontier models and foundation models, most of them, are believed to still consume text in the form of chunks of letters called tokens.

Based on publicly available information, they don’t have any visual perception of letters on a page. They don’t visually see the shape of a character, formatting, or desktop-publishing-type layout on a page. They perceive none of that. At best, they may perceive HTML formatting instructions.

I think DeepSeek OCR, if you squint at the model architecture, is sort of an autoencoder that does optical character recognition after a fashion, but in a really interesting way. It consumes raw images of entire pages and encodes those as image tokens, not as text tokens, and then tries to decode those image tokens into text tokens.

A few things fall out of this. One is optical character recognition at high accuracy rates, which is pretty incredible.

But secondly, this is able to perceive formatting in the way humans do. I think the practical upshot of this would be better grounding. Wouldn't it be wonderful if we could have desktop-publishing-type formatting of outputs from frontier models with beautiful layouts? I would expect that to fall out for free.

I would also expect better understanding of mathematical equations that are dependent on the way the equations are written and how they appear visually. I think better understanding of fonts—all of these things—will eventually fall out of this line of research.

Interesting. We're going to be seeing an article later about Amazon getting into the AR-glass marketplace. We're going to see Meta, Google, and probably OpenAI—all of them—transforming from a phone as the medium of interface to glasses as the medium of interface.

I'm assuming this kind of technology is going to help your glasses effectively translate everything you're seeing into something that can be understood, read, and responded to. I think that's table stakes. So, yes to that, but also having AI that understands, at a visual level, all the text—I think that is going to be quite transformative. Dave, you want to comment on this?

Well, I'm still blown away that when I'm writing code in Cursor or Windsurf and I take a screenshot and say, “Hey, there's a bug in here somewhere. It's an image, not text,” and I just slap that right back into Cursor, it has no problem with it at all.

Now, I know under the covers it's not doing this. It's actually converting it to text and then moving forward from there. So this will put the AI engine much more in tune with human thinking because you're using the same exact pixel-by-pixel interface that we use with our eyeballs for everything, whether it's text or images or whatever. It will be a big advantage in multimodal. But what works already is just mind-blowing to me.

Do you expect, Alex, that we're going to see this type of OCR come into all the models next?

Yeah, I think we're moving toward a near future with universal tokens—tokens that span modalities. I've long thought, wouldn't it be wonderful, aspirationally, if we had just a single modality that everything else flowed through?

Rather than having a text modality, images, audio, and video, if we just had a single, universal, maybe video-style modality that everything else flowed through, it might have certain benefits.

What's really interesting about that, Alex, is that that's happening, and at the same time it's going the other direction in very specific domains, like magnetic bottles and quantum computing, where the knowledge is so far out of the human domain that you want completely different data representations at the front end of the funnel.

These first models are going to use the second models as tools. It's really cool to watch the two spread apart and think about how they're going to end up interacting.

All right, let's jump into our next article here. This is “OpenAI Hires Bankers to Automate Junior Work.” OpenAI has hired over 100 bankers, paying them $150 an hour to train AIs on M&A, LBOs, and IPOs. Effectively, these bankers are, in a sort of way, helping to eliminate the jobs of fellow bankers.

My first take on this is that OpenAI is basically eliminating what I would have imagined an entrepreneurial startup would do. I imagine lots of startups are looking to do this, and this is sort of a shot over the bow. Is OpenAI going to do this for every field? Get rid of white-collar work across the board?

The answer to that is absolutely yes. They're going to do this for every field, and they're going to do it quickly. I just had this conversation with 2 of our companies. Make sure that you're the guy that Sam calls. They're like, “Well, Sam's not going to call me. Why would Sam call me?”

Eliminate everybody else. He's not going to call State Street Bank. He doesn't want to talk to big, legacy, bloated entities. He's also not likely to call 2 22-year-olds out of Y Combinator who haven't even gotten to market yet. So if you're somewhere between those 2 things, you just need to position yourself like Brendan Foody did at Mercor: get into the building and be the person who solves that problem for OpenAI.

But yes is the answer. He'll do this in absolutely every category of human endeavor.

So I would imagine this kind of a vertical would be something an entrepreneur would say: “Okay, we're going to go and do this ourselves for whatever field.” But, again, we talked about this when we were up with Kevin Weil. Is OpenAI going to be moving into and eliminating all the entrepreneurial vertical channels? I find this fascinating.

Yeah, go ahead.

Yeah, I think the superficial story is that this is what the end of so-called white-collar work looks like: vertical by vertical, labor category by labor category. Each existing form of the service economy, each manifestation of it, gets digested and turned into AI automation.

But I think that creates enormous entrepreneurial opportunities for everyone. There are thousands, if not tens of thousands, of labor categories with domain-specific knowledge that will require automation. The same is true with industry subverticals.

Every platform company always instills maybe a modicum of fear in other companies: “Oh, well, the platform will just absorb what I'm doing, and we'll lose our footing.” But I don't think that's an accurate representation of the real economy, where there are just tens of trillions of dollars of service-economy labor that can be automated.

I do not expect a singleton scenario where any one company, any one platform, or any one model just consumes the entire economy. We'll have, I think, a completely heterogeneous economy indefinitely into the future.

So the bottom line here is that this effort by OpenAI could eliminate between a quarter and a half of the junior headcount across Wall Street within 2 years.

All right, moving on. I love this article. Google is prepping Genie 3 for public experiments. Genie 3 is going to let users create interactive worlds with text prompts. We talked about this—extraordinarily powerful, persistent, consistent worlds generated from a text prompt that are photorealistic, that you can get into and use for a variety of different areas.

Alex, do you want to jump in?

Well, first of all, you have to admire that the user interface, which we're now seeing previews of, looks almost identical to the grid of the Holodeck in Star Trek.

Yes, I love that.

You have to admire that we're catching up with the future. It's very exciting.

At a deeper level, though, I do think so-called world models right now are going to merge with the foundation models. I think this is very likely to be an instrumental element of general-purpose, generalist foundation models and frontier models: you'll not just be able to have text-based or audio-based conversations with them. They'll create entire worlds that you'll be able to walk around in.

On the one hand, that's the consumer use case. The enterprise use case is that these fully interactive world models will enable us to create new inventions and new products. This is the mode through which AI understands—and will understand—the physical world and be able to create economically transformative inventions.

It's the democratization of interactive content creation at a level of reality and resolution that is shocking. Just to hit on some of the ideas: for an individual thinking about this, how would I use it? You can build personalized gaming, creative storytelling, and customized education.

For companies, I think a lot of them could be using this for game development and edtech. I personally think the most extraordinary way to educate and learn about something is to dive into that world. I've used this example so many times: if you want to learn Greek history, you can read a very dry textbook. You can even watch a movie.

But imagine being able to drop into ancient Greece. You see a guy in a toga on a chunk of marble, and you walk over and he says, “Hey, I'm Socrates. Let's go for a walk. Let me show you around. Meet my friends.”

That kind of immersive experience is the future of education, without, for me, any question at all.

Well, just not to disappoint everybody, but this is going to be another one of those things that everybody instantly loves.

Just like Deep Research. If you've tried using Gemini Deep Research lately, you'll sit there for 10 or 15 minutes unnecessarily, and then it'll give you something great back, but it's just enough to frustrate the hell out of you.

More GPUs.

More GPUs, man. Keep tiling, because people are going to love it. It's incredible. Unless you buy your own NVIDIA box, you're not going to be able to get the speed you want.

All right, our next article here is “Meta borrows $27 billion to build an AI data center.” Meta’s SPV is borrowing this money at 6.8% to fund a multi-gigawatt Louisiana data center. Dave, you had some thoughts on this one.

Yeah, absolutely. Mark Zuckerberg has taken every penny of cash flow from one of the biggest tech companies on the planet—from Meta, Facebook—and pumped it all into this AI initiative, and now he's borrowing and going to the next level. The stock market loves it.

What does that tell you as a CEO? If you're a true AI company on a true AI mission, you can invest like crazy from your public capital or from your venture capitalists, and they love it because they see the future is here. But I don't think it's probably unprecedented in history for a company that used to be an absolute bottom-line cash cow, producing huge amounts of EBIT, to take every penny of it and then borrow even more to pump it into an initiative.

I mean, Mark has said over and over again that he will do whatever he needs to get to digital superintelligence first.

It's like his war cry. Alex—

I think it's also worth adding that the credit markets are just as interested in financing this project—call it tiling the Earth with compute—as the equity markets. The fixed-income, credit, and debt markets are enormous.

I think we're starting to see this financial model where the lower half of the AI infrastructure stack is being funded by credit, as we're seeing here, and then the upper half, where the models and the applications live, is being funded by equity. We're seeing a whole-of-economy financing model emerge for this full AI infrastructure.

The implication of this, though, is that capital from public equities, from sovereigns, and from debt is all flowing into AI, to the exclusion of so many other technology areas. One of our best partners, Kush Parikh, a phenomenal guy, just co-founded a company called Oryn. Check it out.

My point in this is that people who are technical and engineering-minded wouldn't normally get into the finance side of things. But, like Chase Lochmiller, they're getting drawn into this in a big way, and it's a very good strategic move.

If you have any interest in finance whatsoever and you understand GPUs, chips, data centers, or just math, it's a great direction to go. It's just a huge amount of capital redirecting into this direction, and making it move intelligently—the right investments, the right locations—is not a trivial problem at all. If you have an engineering mindset and you're interested in this area, you can really do well.

Yeah. Amazing. Okay, in the data center world, here's the news from Oracle. Oracle is planning a 16-zettaflop AI supercomputer. We don't talk about zettaflops all that often. It's announced a next-generation cloud computer designed to scale to 800,000 GPUs.

You know, is zettaflop here? Is zettaflop there? [laughter] All right, I'm going to feed Alex on this one. I can't wait to hear what he has to say. I remember on that podcast a couple of months ago, we were talking about 10^26 FLOPs. So, you know, 10^26—that's the regulatory definition of AGI, superintelligent, register it with the government type thing. So that's 1e26.

A zettaflop is 1e21. That's per second, though—that's that many FLOPs per second.

So we're talking about an exponent of—

Yes. Go ahead.

Yeah. To get from 10^16 to 10^21, you need 5 more. So that's 100,000.

Those are orders of magnitude, just to translate.

5 more. 5 more. So, 100,000x. Every 100,000 seconds, a 1-zettaflop computer can create a frontier-level foundation AI model. So 100,000 seconds is 1.1 days, as it turns out. Every day you get a new foundation model, and that's at 1 zettaflop.

This is 16 zettaflops, so 16 times a day you build a frontier-level foundation model. Does that sound right, Alex? Did I get any of that wrong?

I need to double-check, but it sounds approximately right. [laughter] I would add that Oracle—this is all in the public reporting—is both financing and operating Stargate Abilene. Stargate Abilene, together with this 16-zettaflop supercluster, is emblematic of a new form factor for computing.

The personal computer was a major new form factor. The smartphone was arguably a major new form factor. These superclusters, with approximately 1 million GPUs and tens of zettaflops, are a fundamentally new form factor for computing, with high-speed interconnect—which we're not talking about, but which is arguably just as important as the raw compute power—being a key architectural innovation.

This form factor, again in the spirit of tiling the Earth with compute, is not going to stop with Stargate Abilene. Unless something radical changes, we're going to tile the Earth and maybe the near-Earth solar system with this type of new computer form factor.

Incredible. All right, continuing on this conversation: Anthropic to expand to 1 million TPUs on Google Cloud. Their goal is to bring this compute online by 2026. There's, I think, a very loving relationship between Google and Anthropic. Anthropic is sort of the little brother there, and they're growing closer and closer. Alex, what do you make of this?

Well, I want to say something a little bit glib, perhaps, which is that when you have superintelligence that's incredibly thirsty for compute, it makes for some interesting combinations in the market. I think the thirst for compute is creating enormous pressure on the frontier labs to diversify their infrastructure.

We're seeing NVIDIA GPUs up against Google TPUs, up against Amazon Trainium chips, up against ASICs, including frontier-lab-specific ASICs. I think these are all in the mix. For those who are worried about some sort of architectural monopoly or singleton where only one GPU or accelerated-compute architecture completely dominates the market, I think this is a healthful dose of both diversity and reality. Now, actually, we're seeing heterogeneous architectural combinations at multiple levels of the stack.

The future light cone of compute architectures is not going to be dominated by any single company.

Alex, for those who don't know the difference between TPUs and GPUs, would you give us a 101 here?

GPUs—this is branding that was popularized by NVIDIA—means graphics processing unit. This was originally conceived and went to market for accelerating video games, where NVIDIA was arguably the chief actor for accelerating compute specifically for video game purposes and professional graphics as well.

Then eventually it found its way to Bitcoin and other crypto mining. Fortunately, the need and thirst for accelerated compute for AI arrived just in time to recover from a bit of a mini crypto winter and step in.

Meanwhile, TPUs—tensor processing units—is a term from Google, but the underlying architecture is pretty similar to the way GPUs from NVIDIA and other firms handle AI operations. The tensor refers to this idea that the central operation they need to perform in support of AI and machine learning is taking large matrices which, if generalized, become tensors—sort of high-dimensional matrices of numbers—and multiplying and adding them.

To oversimplify, that's the core operation of accelerated compute for machine learning: just taking matrices of numbers and multiplying them.

That's a great point. Appreciate that. We talked about this on the podcast we recorded at Visioneering a few days ago, but I wanted to bring Alex and Dave into the conversation here.

This is Starcloud bringing data centers to space. I'm going to play a short video from Philip Johnston. Philip, who's the co-founder and CEO of Starcloud, was here with me for the last few days, so it was fun to hear his points of view. Let's play the video, and we'll talk about it afterward.

The reason we're building data centers in space is mainly for the energy that we can draw from solar energy in space. There's almost unlimited access to abundant solar energy in space. The problem on Earth is we're very quickly running out of space and, actually, energy on Earth to build large data centers.

In space, we can have these enormous solar panels which can power these data centers. Another advantage is that we can then run large radiators to dissipate that heat as infrared out into the vacuum of space.

So, it's interesting. Philip Johnston was on stage pitching a prize called the Space Cool XPRIZE. It was something like that. Basically, one of the challenges they still have is radiative cooling.

Space is very cold, but there are very few atoms to carry the heat away. So you're focused on infrared radiative cooling, which is a challenge. I'm so curious, Alex: what do you make of this? Is this the future, or is this something that isn't going to happen?

Well, I think at the heart of this is what I would argue is one of the most important civilizational questions that we face.

We don't know the answer, but the question is: Does a mature, intelligent civilization build a Dyson swarm or not? A Dyson swarm means taking apart the planets in our solar system to build lots of computers that orbit the Sun. I don't know the answer. I suspect the answer will depend on physics discoveries that haven't happened yet.

Jumping out a few decades, playing this tape forward, as it were—playing the recording forward—I think if humanity ends up being permanently latency-constrained, we're probably going to do it. This is probably then the beginning of the construction of a Dyson swarm. On the other hand, if physics makes it ergonomic to easily travel to other star systems, presumably with physics that we're not aware of yet, then I could imagine scenarios where actually building a Dyson swarm—turning Starcloud and other orbital computing platforms into a full-on Dyson swarm—probably doesn't make that much sense.

One could also imagine other contingencies. Maybe the demand, as unconscionable as it is right now—the demand for accelerated compute—might peak at some point in the future, if that ever happens. I could also imagine we don't build the Dyson swarm. Otherwise, I think, straight shot, this is the beginning of a long-term trend. Mark this point in time: we're at the beginning, unless something changes, of the construction of a Dyson swarm.

Yeah. Just to clue folks in, the term Dyson swarm comes from Dr. Freeman Dyson, who was at the Institute for Advanced Study at Princeton. He basically said that, as you become an advanced civilization, you're going to want to capture all of the energy coming out of your star. So you'll dismantle your solar system, and you'll basically build a shell around the star that captures all of it.

This is the earliest days. I just want to point out, I had this conversation with Philip. We have 8,000 times more energy that hits the surface of the Earth today than we consume as a species. And the challenge is, can we build the square meterage of solar and dissipation arrays in space? There's going to be a lot of robotics required to do that. And when do we get there? Is it 10 years from now? 20 years from now? We're going to find out.

Along these lines, we saw Crusoe basically announce that they plan to support this by 2027. I'm not exactly sure what they mean by supporting it. They're going to put an H100 up in space, and an H100 in space represents 100 times more compute than any other satellite has had. But it's a single H100. It's not a cloud, not a Crusoe Cloud. Alex, did you dig into this further?

Yeah. I would maybe also just comment on time scales. Putting a single H100 in low Earth orbit, or LEO, may not sound like that much now, but if you just start from physics—if we have this notion that we know, or at least have a prediction, that the end state of all of this is taking apart our solar system—you could actually just do a few calculations to figure out the time scale for when that would happen.

One of my favorite statistics: If we could completely encircle the Sun with solar collectors, capture all of its luminosity, and channel all of that power to, say, unbinding Jupiter—basically disassembling Jupiter—Jupiter's created its own gravity well, so the term of art would be unbinding it from its own gravity—it would only take approximately 2 centuries, if we captured all the light from the Sun, to disassemble or unbind Jupiter. So I view one H100 going into space in the next couple of years as the first step in a potentially 2-century journey to deploy compute at scale in our solar system.

Exponential growth: Double something 30 times, and you get a billionfold increase. Dave, what are your thoughts on this?

Well, Peter, you said that the sunlight hitting the Earth every day is 8,000 times more energy than we consume. But have you ever done the math on the fraction of all the Sun's energy that hits the Earth in the first place?

Oh, yeah. It's far, far less. It's a fraction of 1%.

Yeah, I know. I don't know how many decimal points are in there, but there's a monster amount of energy in that Dyson sphere, Dyson swarm view. So, yeah, 200 years, sure, why not? What's interesting in the short term is this could be a great idea or a terrible idea for Crusoe, and it depends entirely on the timeline to diffusion, which we're about to talk about.

That's an interesting factor in all this. It's worth pointing out: While the term Starcloud sounds like it's got Musk behind it, Elon is not involved in this. He did retweet the Starcloud announcement, but I love Elon. He's incredibly brilliant, but at the end of the day, if he were to take this on, he would probably do it on his own. That's my experience.

All right. Moving forward. Okay, now on to Elon here. Elon says the AI5 chip, by some metrics, will be 40 times better than AI4. “We deleted the legacy GPU. It's basically a GPU. I poured so much life energy into this personally. It'll be a real winner.”

We've seen this before, where Elon goes heads-down and focuses on a very specific element, all the way down to the engineers, scientists, and the production line. Alex, you've been tracking this. What does the AI5 mean for Tesla, for Optimus, and for xAI?

I've spoken on the pod in the past about this notion that superintelligence is not going to stay just bottled up in the data centers. I've argued in the past that it is literally going to walk out the doors of the data centers in humanoid robotic form, in driverless car form.

I think what's most intriguing about the AI5 architecture is that it's a unified architecture. This is a single accelerator that is planned for use both on the data center side and on the robotic/car side—a single chip. This is something new that the world hasn't seen before: a single, unified architecture for both cloud data center compute and also embodied in robots and cars. So I think this is quite literally potentially the embodiment of intelligence walking out the door of the data center into our homes and into our lives.

Well, this ties back to our last story, too. All the big guys now have their own chips, as Sam announced in our last podcast that he has his own Broadcom custom designs. Anthropic is the one exception, and they're going to adopt Google's TPUs. That was in that other slide.

But that's not a very comfortable place to be if all the other competitors have their own chip designs, because as they're modifying their algorithms, they're tweaking the chip design. Once you're in bed with Samsung or TSMC or Intel and you have your whole supply chain going right into your own data centers, you can innovate, redesign the chip, and get it back into production very, very quickly. Google's already got that cycle time way down. So it leaves Anthropic in kind of this uncomfortable position: Well, we're buddying up with Google. Yeah, but you're on their TPUs. They're going to give you whatever they want to give you.

Fascinating. But all of this comes back to TSMC production capability, right? At Samsung, there are basically choke points.

Yeah, there's no doubt that any one of these companies would be buying TSMC, Intel, or Samsung tomorrow if the regulators would let it happen, because that's the choke point, and they all know it. So all these really high-level partnerships and relationships are forming, and it's a very competitive playing field.

Week by week, we're seeing the shifting relationships and capital flow here. All right, this next article comes from Amazon and their new delivery glasses. Let's take a look at the video here and then talk about the implications for this. It's fascinating what this means for labor.

Well, check out these nerdy smart glasses. These are smart glasses developed by Amazon for their delivery drivers. They're just in development now, but basically they use technology, like a heads-up display, to show you what you need to do.

So in this case, instead of using your mobile phone as a driver to scan the parcels, you simply look at them and work out which parcel needs to go where. But then when you head out to deliver, it gives you actual information about the place you're delivering. It gives you warnings about dogs and things and shows you exactly where to leave it. And it's all done. Even the photos are taken, and you never need to use the mobile phone.

So, cool technology, very much like Meta's Ray-Ban glasses or maybe Apple Vision Pro.

Okay, so this is what I think is going on. This is put forward as, “We're going to help our drivers, keep them away from barking dogs, and help them do this with hands-free delivery.” I think this is a mechanism by which Amazon uses the drivers to collect a lot of information to train their delivery robots. This is just like Tesla with its cameras training its full self-driving models. Dave, what do you think?

Yeah, you're exactly right. It shows you how the technologies interact, too, because the glasses will be profitable instantaneously within their internal use case. They can perfect them, and then they can decide later.

You remember there was a Kindle phone, Kindle Fire phone. It didn't succeed, but they've tried before to compete with Apple and Android in the device hardware game.

So this is a great stepping stone for them to make money and perfect the device while gathering all the data, which will then feed their robotics initiative, but also the consumer glasses initiative, which will come later. So you're exactly right.

Yeah. Do you want to add anything, Alex?

I'll just add that I think this functionality can generalize well to non-delivery functions as well. I think this is the tip of the iceberg for using wearables to automate and, even before we get to automation, to capture telemetry and training data for the entire services economy. So I think we're going to see this across many other verticals: healthcare, energy, and hospitality. Expect smart glasses and wearables for building pre-training datasets and post-training datasets across every possible vertical.

Also construction. Construction, you know, we're doing the biggest construction buildout in the history of America and, certainly, probably the world. It's all electricity, plumbing, buildings, and everything. But because those are AI-forward projects, like Chase Lochmiller at Crusoe and Project Stargate, they're going to be early adopters of exactly the same thing you're talking about for construction. Construction is a huge fraction of the global economy.

And for me, the most important thing for an aging population is going to be memory augmentation: using these glasses to remember who you're talking to and the last conversation you had. Personally, I can't wait. I meet so many people, and I love being able to remember the details, but sometimes it's just a challenge.

All right, we're going to go into a subject we covered on the last pod with Emad in particular and Eric Boulier, but I cannot wait to hear the take that you, Alex, and you, Dave, have on this. Again, this is Google's quantum breakthrough nearing real-world use. This is the Willow quantum chip. A friend in Santa Barbara, Hartmut Neven, who heads the Google quantum team—congratulations. But at the end of the day, Alex, what does this mean?

Well, first, maybe a little bit of the background. I read the core Nature paper behind this announcement. Very interesting. This was the Google team and its collaborators.

And by the way, Alex, I have to say I really appreciate the fact that you dig in—

On everyone's podcast to go and read the actual science, you know.

Well, it's difficult to comment on it if I haven't read it, but thank you.

I understand that.

Well, everybody else on the planet is commenting on it without reading it. You're the only one doing it.

I've seen these comments on YouTube that Alex is an AI. I've seen him glitch. God knows.

We want to use this as our cold open.

I'm not going to disclose any details.

But maybe we'll see you in real life. Anyway, dive in, please. You read the paper. What does it say?

Right. I read the core Nature paper behind this announcement. It's very interesting. The premise is that there's a certain physical quantity, and in the case of this announcement, it's called a second-order out-of-time-order correlator. This is basically a measure of quantum chaos. It measures how chaotic a given quantum system is, and the Google team and its collaborators showed that it would be very challenging for a classical computer—that is, a non-quantum computer—to compute it.

I think it's very interesting. It's nice progress in terms of demonstrating quantum speedups, or quantum advantages, versus classical computers. What I'm still waiting for, though, if I got my wish, is a more economically transformative quantum algorithm. What I'm waiting for, what I'm hoping for, is that sometime in the next few years we will achieve a definitive breakthrough speedup for quantum acceleration of AI.

I think applications in quantum simulation, quantum chemistry, simulating materials, and optimizing molecules are great. I don't think they are necessarily world-changing. The world-changing use case for quantum acceleration, if the physics of our universe are so kind as to allow it, would be something like achieving orders-of-magnitude speedup in training or inference for a frontier model. I think that would be utterly game-changing.

Amazing. The term “quantum advantage” was coined a few years ago as the point at which a quantum computer demonstrates the ability to do a real-world thing better than any classical computer, right? With 1s and 0s. People have been chasing this idea of a quantum advantage to rationalize the massive investments and actually get traction. Now we have a number of public quantum companies wanting to get revenues.

I think one of the other important things to note here is the concept of error rates in quantum computers. How do we get to logical qubits, and how do we reduce the error rate so we actually have something that's going to be useful? But let me ask you a different question here, Alex: How big is quantum computation as compared to AI? How big is it relative to AI? Is it larger, many times larger? What are your thoughts?

I want to bisect the question into now, short-term versus long-term. At the moment, the actual applications are relatively pedestrian, prosaic, and not economically transformative. The best applications I've seen anywhere close to being useful in the short term are for quantum simulation, leveraging the fact that it's relatively straightforward, as Richard Feynman, who arguably helped create the entire field of quantum computing, pointed out. You can use one quantum system to simulate another quantum system relatively easily.

But these aren't economically transformative, not in the same way as AI, which is just turning our services economy, as we were discussing earlier, and automating it. Quantum doesn't have that capability in the short term. In the long term, I would hope quantum will enable us to build much faster AI systems.

In the long term, I'm holding out hope that quantum will, in the end—there's almost an angle, you'll forgive me for this—have a redemption arc. That's what I'm hoping for with quantum information systems, because so many of the problems AI is solving right now are grand challenges, like protein folding. Do you remember 10 or 20 years ago, when there was a sizable community that thought protein folding would require quantum computers to solve? That did not happen. We were able to solve it with just AI on top of classical computing.

There's almost a “Who Moved My Cheese?” angle to the sense that the grand challenges quantum was supposed to be the great white knight to solve for us keep getting devoured by AI instead. I'd love to see a bit of a turnaround sometime in the next 10 years.

Fascinating. My favorite science-fiction books all have digital superintelligent AIs and conscious AIs doing so on the backs of quantum clusters.

Potential advantages include energy efficiency. If we could build a fully reversible AI supercomputer, that would probably have some sort of quantum-coherent foundation that would be transformative. We wouldn't need to build all these SMRs, fission plants, and natural-gas co-location facilities if we had fully reversible, quantum-computer-based foundation models everywhere. But we're not there yet.

Nice. Dave, let's go to the next article here, and I'd love your thoughts on it: “President Trump Eyes Equity in U.S. Quantum Firms.” This is the potential beginning of a sovereign-style VC fund for the United States. He's targeted IQM, Rigetti, D-Wave, Quantum Computing Inc., and Atom Computing.

I mentioned on the last pod, when we talked about this, that I had taken D-Wave public through a SPAC—a huge, 8,000% return from the earliest, lowest point to where it is today. Dave, thoughts on this?

Yeah. Well, I love it and I hate it as a precedent, but I still love it because Alex is always pointing out that what we're doing right now is unprecedented, except maybe during the buildup to World War II. Think about 1939: We're basically flying biplanes in the U.S. Air Force. Five years later, we have jets.

It's just an incredible amount of government investment. That's what's going on right now in AI, and it's great. It's what we need. Now that's moving into quantum, too. You've made the point many times, Peter, that our economy doesn't function well in these areas that require you to think more than 5 or 10 years in the future.

China works really well at thinking 10, 20, or 30 years in the future, but we don't do that well. So the government kickstarting quantum is a great move if you believe in it 5, 10, or 15 years in the future. But as a precedent for government involvement in the economy, it's terrible, because they're going to make terrible decisions in the long run.

These are very good decisions in the short run, but that's because all this incredible talent has gone to Washington for the first time in my lifetime. But that's not sustainable.

We see the government investment triggering huge amounts of private investment that follow on, right? So after the Intel deal, Intel stock doubled, from $20 a share before to $40 a share a day or two ago.

And we’re seeing this again: a 10% to 15% increase in these quantum stocks after the story got leaked.

I wonder where they’re going to go next. I think the government is going to go into rare-earth metals; we’ve seen some of that conversation. Where else might they be making strategic investments? Well, I hope they take Alex’s World War II analogy and stay focused on the things we need in this very specific race to AGI and ASI.

Rare earths would fit for sure, and energy would fit for sure. Quantum may or may not.

I’m not going to second-guess the Commerce Department or the executive, but there is some reporting that there may have been some money left over from the CHIPS Act. Quantum firms would certainly be interested in either obtaining equity investments, or—my guess is more likely—loans, warrants, or some other financial structure.

I think the question of how strategically important quantum is as a technology, when you compare it with more obvious feedstocks like rare earths, energy, compute, or fabs, is still to be decided. I don’t know.

I’m kind of shocked that the government hasn’t made a move to get into the fusion companies or the SMR companies, really to help accelerate that, because I think that one thing would bring a lot more capital. Commonwealth Fusion Systems is probably the best-funded. I was talking to some of the fusion companies here at Visioneering and talking about Helion.

Interestingly, they said, “Helion is so close-lipped. We actually have no idea what they’re doing or how far along they are.” There’s public disclosure and some information. They’re claiming 2028 for Microsoft, but we don’t actually know. And these were top fusion experts. Commonwealth Fusion Systems is targeting 2030, but they still have a lot more development. Alex, do you have any thoughts on that?

Yeah, I’m not going to second-guess the Commerce Department or the executive, but there is some reporting that there may have been some money left over from the CHIPS Act. Quantum firms would certainly be interested in either obtaining equity investments, or—my guess is more likely—loans, warrants, or some other financial structure.

I think the question of how strategically important quantum is as a technology, when you compare it with more obvious feedstocks like rare earths, energy, compute, or fabs, is still to be decided. I don’t know.

Well, I will say I can’t add anything to Alex’s insights on this at all, but I will say I talked to Frank Wilczek about it. He’s a Nobel Prize winner in physics, famous, and spent his whole career in quantum physics, and he said almost exactly the same thing Alex said. So there are 2 data points.

Yeah, I’m not going to second-guess the Commerce Department or the executive, but there is some reporting that there may have been some money left over from the CHIPS Act. Quantum firms would certainly be interested in either obtaining equity investments, or—my guess is more likely—loans, warrants, or some other financial structure.

All right, let’s jump into energy. A few different articles here. This one’s interesting, in particular, because there’s a chart showing us the increasing price for US construction of nuclear reactors versus China. Here’s the quote:

“Construction costs for nuclear reactors in the United States have risen roughly 1,000% since the 1970s, while China’s costs have steadily declined.”

That’s not good news. Alex, do you want to weigh in on this?

[snorts]

Construction costs for nuclear reactors in the United States have risen roughly 1,000% since the 1970s, while China’s costs have steadily declined.

Yeah, I think there is an alternative history where the US basically never stopped building nuclear plants in the late 1970s. If you’re familiar with all of the microeconomics around experience curves, unit costs tend to collapse the more you make of a given item. As a country, the US basically stopped making nuclear power plants decades ago.

I think if we’re going to feed the voracious energy appetite of these AI data centers, we need, as a country, to relearn how to build lots of next-generation nuclear plants. The good news is that the demand signal is being sent by the AI data center companies. But I think there will be all of these knock-on benefits, not just for AI data centers, but for everyday life, if we live again in a truly power-rich society.

Well, Alex, it’s worse than that sounds, too, because it’s not just about unit costs. If you look at the actual construction of a nuclear facility in the US, it’s mostly overhead, regulatory, political garbage costs.

It’s regulation, it’s litigation, it’s loss of manufacturing expertise—all of these things. We’ve done it to ourselves. All right, next article here is fascinating.

The US is offering nuclear energy companies access to weapons-grade plutonium. This comes from Energy Secretary Chris Wright. The US Department of Energy will let private firms use 19 tons of plutonium from old warheads to fuel their next-generation reactors. The move is boosting domestic nuclear supply and reducing reliance on Russian uranium.

I find this fascinating. I mean, talk about removing the shackles and giving entrepreneurs access to feedstock. Who wants to take it?

Well, everybody probably knows this, but the cost of the fuel in a nuclear reactor is tiny. It’s a rounding error. So everyone’s been buying their fuel from Russia for a long time. Opening up the US supply doesn’t really change anything. It’s a rounding error in the overall costs anyway, but if you’re going to buy it from Russia anyway, what’s the harm in using our surplus plutonium? So it’s not changing the math one iota.

Alex, I’d also comment, maybe even more broadly, on nuclear engineering as a vibrant discipline. There was maybe a bit of a hot take, but there was a period of time for a few decades when nuclear engineering, unless it was for, say, some biomedical application, was positively unfashionable to study.

I don’t want to call it a nuclear winter, for obvious reasons, but I think that period of time is something we’re coming out of now. As a society—speaking particularly of the US, but the West in general—we’re entering an era when we need to refamiliarize ourselves with the nuclear fuel cycle and get comfortable with nuclear fuel cycles in general. It’s part of the future.

In particular, part of the future is fusion, and so the US has put forward a new roadmap for fusion energy. The DOE roadmap touts commercial fusion by the mid-2030s, with an actual aim for public infrastructure in the 2030s to scale up. Interestingly, this has 0 dollars of federal funding behind it and $9 billion of private investment. Alex, you found this particular timeline. Talk to us about it.

Yeah, I enjoyed reading the roadmap. I thought it was delightful in some respects. The roadmap calls for 3 stages of advancement in fusion energy in the US.

The first stage, call it the short term, over the next 2 to 3 years, calls for early-stage fusion demonstrations. So that takes us through 2027 and 2028. The second stage, the medium term, calls for early-stage fusion pilot plants between 2028 and 2030. The third, quote-unquote, long-term stage calls for actual operation and production of power-generating plants between 2030 and 2035.

So this is actually a very ambitious timeline, I think some would say, at least by historic standards, where fusion was always 30 to 50 years out. Now it’s basically in our short term. It also, I think, aligns with some of the public announcements that Helion, on the one hand, and Commonwealth Fusion Systems, on the other hand, have made regarding actual test facilities being in operation between 2028 and 2030.

So I think, in short, this roadmap is more a reflection—or at least I interpreted it as more a reflection—of some of the most ambitious private-sector players and their actual plans.

Yeah, Dave, we’re going to be having dinner with Bob Mumgaard on Wednesday night in Riyadh. We have our Abundance Dinner that we’re co-hosting with Amjad from Replit and Link Ventures. A lot of incredible people are going to be there, so I look forward to asking him more about this.

Yeah, me too. Yeah.

I mean, the head of Commonwealth Fusion Systems has done extraordinary work. I’m excited to see where they’re going to go. All right, continuing on the energy theme: Amazon bets big on next-gen nuclear.

This is about the state of SMRs—small modular reactors. This one is with X-energy. We’ve talked about X-energy before. Its initial 320-megawatt output can scale to nearly 1 gigawatt, which can power data centers, obviously carbon-free.

I love SMRs, and I love the Gen 4 nuclear reactors. Unfortunately, we’ve shut down our ability to manufacture these—we’ve talked about this—and so this has become an entrepreneurial effort. One thing that I find fascinating is that while we have the designs and the permissions, the timelines for getting these SMRs out aren’t 2026, 2027, or 2028. They’re in the 2030s.

Which is concerning. Why can’t we get these going faster, right?

No, the timelines are really interesting to track, and it’ll come up at FII next week in a big way. A gigawatt—Eric Schmidt said we need 100 gigawatts by 2030, and that’s just a fact. It can’t go up or down because that’s the number of GPUs we’ll be making. They’re going into production one way or another, so you need to find 100 gigawatts by 2030. That’s only about a 10% expansion of the US power supply, so it’s not insurmountable.

But in 2031 and 2032, the new fabs will be online, and GPU production will go way up. So then you need something massive. The 100 gigawatts is a stepping stone to something much bigger just a few years later. If fusion comes online in 2029 or 2030, it’s massively important, but if it’s just 5 years late, where’s that power going to come from? Then suddenly you’re launching them into space, and so these are completely different ideas. The modular reactors here are fission, so that’s the third option, plus renewables is a fourth. All those things are racing against this 2030 clock.

I have to imagine by 2030 we’re going to have figured out more energy-efficient compute—10× or 100× more efficient.

And you know, Alex, I'd love to hear your thoughts on that.

The intelligence of the AI between here and there is going to be like—

Yeah, but also, I have to invoke Jevons' paradox. We're presumably going to have much more demand for it as well, even though cost per compute and algorithmic advances are going to improve 5× to 10× every year. That may be, optimistically, the amount of energy reduction that we need in any given year. So I don't know when or if there will be a turning point where we need less energy.

I will point out, though, with the SMRs, I think it's striking: no cooling towers. This is a totally new form factor. Decades of acculturation have trained people to look for those iconic cylindrical cooling towers. No cooling towers.

These can be put in so many more locations. They are compact. They can be put into novel sites that otherwise might never have been on the table for some of the first-generation nuclear power sites. So even if there is a sequencing issue, and even if the first boatloads of SMRs start arriving circa 2030, I do think they're very likely to end up being an important part of the overall power mix for AI data centers and otherwise. Yeah, keep in mind, the vast majority of data centers don't need to be near population centers.

And that's a big difference. Those iconic cooling towers that Alex was mentioning—people hate them when they're on the beach in front of your house.

But these SMRs can be in Wyoming, Texas, and Nevada, in the middle of very unpopulated areas. That's a great place to put some of these really large-scale data centers. So this will happen for sure.

They look like normal buildings. That's what's most striking to me. You would never, at least with the eyes of 2025, look at the building that you're seeing and say, “Aha, that's obviously a fission site.” It looks like a normal building.

Amazing. So we're going to see a continued mix. I sure hope that the government does start backing solar, SMRs, and fusion more. We need to accelerate our energy production beyond just natural gas and coal and other areas. I'm going to end this with what I'm going to call a weird science article.

So let's end on something that doesn't normally enter our conversation in the exponential world. Alex, you found this one. It's called “Butt breathing: a real medical option.” Do you want—

Sure. Sure, Peter. I'll take the hit for ending on a low note. But in all seriousness, this is a transformative breakthrough, or at least the beginnings of a transformative breakthrough, for people suffering from severe respiratory failure who can't breathe through their lungs.

If folks have seen The Abyss, the science-fiction movie where there's a famous scene where a character is consuming an oxygen-substitute liquid—so breathing liquid, basically, deep underwater—they'll have some familiarity with novel forms of respiration and blood oxygenation. This was also the subject of last year's Ig Nobel Prize for discovering that nonhuman animals could oxygenate their blood supply by consuming oxygen via the other end, as it were. So, only so many euphemisms I can use here.

Well, the intestines are a very blood-rich, large-surface-area part of your body. And so, if you're able to put a hyperoxygenated fluid enema—let's call it that—then you can perhaps oxygenate your blood supply and get enough of your red blood cells oxygenated to get to your brain. Seriously. [laughter]

We've lost our entire audience [laughter] on this particular article. The only reason I like this story and wanted it in the podcast is because every time Salim says something in the future, we have the option to say, “Oh, he's butt breathing.”

To maybe just try to elevate a little bit, there's been interest over the decades in nanobots that would help with oxygenating the blood, so-called respirocytes. And, to the extent that it's possible, I should add parenthetically that there are also sci-fi scenarios, like enabling humans to hold their breath underwater for hours on end. So there's been persistent sci-fi pressure to discover new ways to oxygenate the blood in environments that are, call them, less than hospitable.

You're really reinforcing the theory that you're an AI. [laughter]

So all of this materializes on the backside of nanotechnology. And one of these times, I really want to dive into wet nanotechnology, where we're using DNA origami, but Drexlerian assemblers—that just opens up everything. And respirocytes are fantastic, literally BCI-enabled through nanobots in the brain. I can't wait.

So I'm going to get Ray Kurzweil on our podcast so we can have the conversations with him. Ray's been a dear friend and a mentor for so many years. At the end of the day, his prediction is nanobots by the early-to-mid-2030s, so 2033.

And that's going to unlock high-bandwidth BCI, but also basically unlock longevity escape velocity. Or I don't like using the term immortality because it sort of hits so many different negative buttons. But if you can repair, on a cellular and subcellular level, all parts of your body, that is an incredible future.

Well, if you get Ray and Alex on the same podcast, that podcast could also be immortal. That would be something I would kill to see.

Well, we'll do that for sure. And again, to all of our friends listening, I hope you've enjoyed this episode of WTF. If you're not a subscriber, please join us. We're putting out news as it breaks, so while we try and do this once a week, sometimes it comes out twice a week, and you'll get a notice of that. We hope that, other than butt breathing, this helps you understand how fast the world is changing and that we're living in this extraordinary time where we can solve any grand challenge. Congratulations to the Visioneering XPRIZE teams for winning Visioneering and to the entire XPRIZE organization for really accelerating these grand challenges.

Crazy.

It's going to be fun. I'll see you and Immod and Sem there. Alex, we will miss you.

I'll be there either digitally or in spirit, but we have quite the week lined up, meeting with the top CEOs from all the AI and tech companies. It's going to be fun. Any favorite meetings are you looking forward to, Dave?

Well, you know, you're kicking it off with the big shots. You've got Eric Schmidt, Larry Fink—just the big, big money people and the big vision people. That's going to be such a fast start.

But then backstage, it's like—God, it's just a who's who of incredible people. So I'll be backstage the whole time.

Yeah, it's going to be wild.

So, thanks all.

Yeah, no, a pleasure. I chair FII out of Saudi. It's the Future Investment Initiative, and I'm on the board there, and I chair their AI activities.

One of the things that's going to be interesting this year is we have 20-something heads of state, and I'm going to be co-chairing a conclave with Mida[?] from a16z. We'll be talking about how to use AI to accelerate governance for countries.

One of the biggest challenges we have is that the speed of change is so extraordinary and so disruptive in terms of AI, humanoid robots, and longevity that countries out there are having a difficult time trying to understand what policies they put in place. What do they do best for their nation-state and their citizenry?

So we're going to be announcing a program called the Sovereign AI Governance Engine. We'll talk about what that means, but it's really to help people around the world deal with disruptive change and disruptive opportunity at the speed of AI versus the speed of governments and PDFs.

Yeah, it's going to be good. And the reason that's coming out in Saudi Arabia, at FII in Riyadh, is because the deployment rate of ideas like that can be very, very, very fast in those countries. They make decisions in a very tight-knit, very fast-moving group.

And so that'll be a huge bellwether for Western democracies, because it'll happen there long before it happens in the U.S. and Europe.

Alex, what's the week like for you, buddy?

It's, in some sense, the same as every week for me, which is trying to accelerate and smooth out the gentle singularity.

Yes, I love that.

All right.

Sounds great.

All right.

Take care, all.