奇点已至：AI 正在解决数学问题，Sora 超越 ChatGPT，AI 正在设计芯片｜EP#201

It feels like we're in the midst of continuous event horizons.

We all agree on this podcast that we're right in the middle of the singularity at this moment.

I think it's increasingly likely that the singularity is an optical illusion. It's an optical illusion that appears at a distance. It looks like a vertical asymptote, but when you're in the middle of it, as I increasingly suspect we are, it feels quite continuous.

If you froze technology today and just assimilated what we invented in the last 2 years, it would take decades to realize all the implications. GPT-5 Pro set a record in FrontierMath. We now have a clear line of sight to solving all of math, or substantially all of math as we understand it in 2025.

Now, with AI, how do we navigate this future? Because we can see it's coming now. So what does that future look like? Let's start painting that picture. What does it mean when math is soft? What are the implications for all of our subscribers here?

Gentlemen, good morning. Exciting day here.

Morning.

Good morning. Oh, yeah.

We're recording this at 6:30 a.m., at least Pacific Time. A couple of minutes ago, we were talking, and Salim, you were saying, “Oh my God, you're getting up so early.” And Alex, what was your response?

This is the slowest it'll ever be for a while.

You said, “Don't sleep through the singularity.”

I think that was it.

Honestly, it's like waking up every day more excited than the last. It's just fun.

To our subscribers and listeners, we've spent the last 4 or 5 days gathering articles that are, in our minds, the most significant things going on. The speed is accelerating. We're going to close this podcast conversation with a discussion about what Ray Kurzweil's singularity, which we're going to hit by 2045, actually means, because it feels like we're hitting it a lot earlier.

Dave, what's been on your mind this week? Ray was at MIT making that presentation, which we'll get to at the end of the pod.

I've been thinking about the specifics of the timeline from here to there and this incredible compute shortage. We'll be in Riyadh next week, and that's data-center central right now.

I've been thinking about that a lot, too.

Arguably, this is the most important conference of the year globally.

Yeah.

Just because trying to solve which problems we all want to solve is such a critical thing today.

Yeah, I agree. I think the order of operations matters a lot, especially for people investing or career planning in this area. The timelines are becoming more concrete now, so I think we'll really cement our ideas next week, and then Alex will refine them into the perfect message for the audience.

Agreed. Salim, what's on your mind this week? You're going to be with me in Malibu at Visioneering, and then Dave, you and I are off to Riyadh for FII9, the Future Investment Initiative. A lot of AI conversations are happening there.

We're pretty much all there, right? Huge conversations are going on. It looks like the big dominant conversation will be, “How do we use AI to solve everything?”—to Alex's point, which he repeatedly makes on this pod—because now we can apply AI as a tool to any of these domains. It's huge.

Yeah. Alex, how about your week? I'm sorry you're not going to be with us, but I'm sure you're busy.

It's been an exciting week. Arguably, one of the most exciting developments over the past week or so was the solution of math. I would argue that we now have a clear line of sight to solving all of math, or substantially all of math as we understand it in 2025, with AI. That then topples physics, chemistry, and biology.

What did you say to me? We're going to have an accelerated play of Star Trek. It's all going to happen. We're speed-running Star Trek over the next 10 years. It's not the 24th century; it's more like 2035.

Crazy. Crazy. All right, hold on to your seats, everybody.

The future is collapsing into the present, basically.

Linear versus exponential. See?

Yeah, seriously. I added some slides here at the beginning to talk about the speed of change because I want everyone to understand this.

We'll begin with this image, which shows that the adoption of AI is now 8 times faster than we saw with the internet. We went from 0 to 200 million users in AI in one-eighth the time it took us to get there in internet years. Any comments on this?

This slide is understated, too. It's showing ChatGPT alone against the entire internet. If you include Gemini and the other engines, it's well over 1 billion on that left chart. So, yeah, it's even more acute than this chart makes it look.

I think this is likely to be in keeping with the notion that this is the slowest that things are likely to be for some time to come. This is actually still pretty slow. Superficially, one can look at the AI curve and say, “Okay, we deployed superintelligence upgrades and reasoning models to a chunk of humanity over a few years.” It's actually still pretty slow.

We don't yet have a conduit for deploying physical-world upgrades to most of the world. ChatGPT obviously rode on top of prior platforms like the internet, personal computers, and smartphones, but we don't yet have a conduit for deploying material or physical changes to the world. I think it's going to look like robotics, nanotechnology, and a few other key technologies. Things will actually be moving pretty quickly once we have those conduits that we don't really have yet.

Like you said—

Your point being that once we have 5 million robots out there, then you get an instant upgrade to everybody. We don't have that for humanity—

Or 5 billion.

Let's make sure we come back to that in a future pod, because I think the constraints to manufacturing are going to hold back robotics, and there's a separate curve for that. But then you have the nanobots, which don't require a lot of material. From a medical point of view, the nanobots are going to be massively impactful, and I think those are actually going to come sooner than people are predicting because they don't have the component-supply-chain bottlenecks that the robots do.

We don't hear a lot about nanotechnology in the classical Eric Drexler sense. We hear about wet nanotechnology, in terms of DNA origami and so forth. But the idea that you can build an assembler—a nanoscale robot, a subcellular robot—that's able to pluck atoms of different types and build materials out of pure carbon, diamondoid materials, is still a few years out.

I think it's going to be arriving on the back of AGI and what follows. I've seen a team that seems to have a viable, credible path to molecular manufacturing.

Yeah, I think that's likely. You're going to see later in this pod Greg Brockman designing chips using AI. What the heck does Greg Brockman know about designing chips? With the help of AI, anything is possible.

It's very similar to Demis Hassabis solving protein folding. How does Demis know anything about protein folding? With the help of AI, anything becomes possible. I think these areas, like molecular manufacturing and nanobots, are going to come very soon from unexpected places—from early adopters of tools that are tuned to the problem.

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.

You know, one of the comments we get sometimes is, “Okay, you guys are wealthy.” None of us started that way. At the end of the day, we created some wealth. But here's the point I want to make for everybody who's listening and struggling—and there are people who are listening who are struggling: These technologies are massively demonetizing. We're going to have autonomous cars that are 4 times cheaper than owning a car.

We're going to end up with nanotechnology where, if I have a nanobot—and this sounds insane until it all materializes—I can literally throw that nanobot in the ground and say, “Make anything out of raw materials.” The information set is free, the energy is around, and you're basically plucking atoms of whatever you need. I mean, that's how an oak tree starts from an oak seed and grows over time, just at a very slow pace.

Can I mention a couple of things about this?

Yeah, please.

You write—one of the things you pointed out in Abundance, right?—that if you went back a few generations ago, the richest people in the world exclusively had inherited their wealth. Today, you look at the richest people in the world, and exclusively, they've earned their wealth. They went from zero to everything. This is a mindset problem more than anything else.

You take Vitalik Buterin, an 18-year-old kid out of Toronto, who ignores his professors, gets together with a few friends, and boom, you have a $600 billion ecosystem that nobody understands. So there's this unbelievable potential to go from zero to everything. This has never been true before in the history of humanity.

So I've had so many friends using

That's completely the mindset that gets you there.

AI to say, “Okay, just going deep and saying, ‘This is what I love doing. How can I start a business here? What do I do first? How can I learn about this?’” It really is making a commitment to yourself to use these tools to educate yourself and then to build on top of those tools.

Yeah. Can I give a crazy example? Sure. A friend of ours—we probably all know him—I was talking to him. I won't use his name because I'm not sure he's comfortable with it, but he told me that last month he launched 47 startups using AI.

Okay. Just in a month, he and his team coded and pushed them out. That's just insane. That's unbelievable.

Yeah, it is. All right, so speed is going fast. Here's another article showing the speed of change: “AI content overtakes human content online.” This is fascinating. This is just over the past 5 years. It used to be 100%, or near 100%, human content. We had AIs writing articles for certain magazines as early as late 2018, or thereabouts, but we dropped from 100% human content to below 50%, and AI-written content is exploding. Dave or Alex, what do you guys think about that?

Huge deal. I mean, there's enormous opportunity in what you were saying a second ago, Peter: the ability to use AI to create new content is a business opportunity and a life-changing opportunity for everybody. The tools are incredibly democratized and easy to come up the curve. The AI will explain to you exactly how to use Sora 2 or whatever, Veo 3, and there's a huge amount that you can do with this.

If you think about all the friction that people have in life, if you're operating locally—you know, you're a local government, or you're a local business owner, or whatever—putting an AI interface on your business makes it dramatically more usable for all of your customers. That alone can be a life-changing business opportunity.

Anyone who's a reporter, an editor, a video creator, or whatever, this is such low-hanging fruit all of a sudden. It's not all slop. I'm not going to deny that it's used for slop a lot, but this is high-quality capability.

The ability to create genuine, high-quality content with these tools that's much more compelling than just writing an article would have been is right in front of you.

Alex, what are your thoughts?

Maybe I'll add that there is this cliché out there that we're just going to drown in AI slop. I don't buy that for one second. If you rewind 20 years or so, there was a cliché that we were going to drown in email spam, and that also did not happen.

Agree.

Better filters get brought into existence, and ultimately, the same tools that would empower spammers also empower the ability for small individuals to have basically their own electronic printing press and disseminate their ideas to millions of people via email campaigns that people subscribe to.

I don't subscribe to the notion that when we look at a chart like this, it just means humanity is drowning in online slop. I think, if anything, there are sufficient economic motivations for the quality of slop to ultimately disrupt, from below, in an innovator's-dilemma style, the quality of human writing. If anything, I think we end up merging with the slop.

Well, I mean, listen: one should not just take whatever their AI writes and publish it. You should read it and make sure it represents you. You should be the originator of the basic idea and use AI to help level up the content you're producing.

We've had synthetic data. Aren't we now entering the world of synthetic culture? Doesn't this become like a hall of mirrors, where everything is reflecting on what has happened before and just amplifying itself in a totally weird way? This is going to become completely unpredictable, is it not? I think the principle is that culture was always natural. How do we distinguish between natural and synthetic culture to begin with?

Dave?

I think Alex's point on spam filters is really important here, too, because the ability to have your AI friend or agent filter the content and reduce it to the subset that you care about is so easy compared to spam—or so much more powerful than the spam-filter version of that—and it works perfectly fine.

I think when Tyler Cowen came out with his new book, one of the things he said when he launched it was, “99% of the readers of this book are going to be AIs, not people.” So I designed it to maximize the impact on the AIs, and the AIs are going to summarize, translate, and feed it to the humans.

So that's kind of the future of writing.

That's so important to realize: when you're writing, the major impact you're going to have on the world is through an AI interpretation of what you've written. That's amazing. There are people using LLMs to essentially run SEO. They're feeding these things with great images of their company and stories about their companies, et cetera.

This next article, “Large language models are improving their forecasting ability,” is fascinating. There's this thing called a superforecaster. I remember reading about this years ago. It's a person who can demonstrably and consistently make accurate predictions about the future, as compared to the general public. There's a term for that.

It looks like GPT-4.5 is now scoring very close to these human superforecasters, and it's likely to exceed the best human superforecasters by late 2026. Alex, thoughts on this one?

Yeah. So maybe, first, a bit of background. This is a benchmark called ForecastBench by the Forecasting Research Institute. It consists of 500 constantly updated binary questions—yes-or-no questions—that look something like, “Will the following happen by the following date?” and can be automatically verified.

When I see an experience curve like this, my mind immediately goes to science-fiction writers like Ted Chiang and Frank Herbert, who have written extensively about what happens when AI or superhuman intelligence can predict the future to ultra-high accuracy. What does civilization look like when we can predict things that are right around the corner?

I think, arguably, if we can predict the future of civilization, we can also steer the future of civilization. This isn't just some sort of centralized steering mechanism, since everyone has access to models from o1 to GPT-4.5. It's not some sort of centralized command-economy-type future.

Imagine a future where everyone has the ability to predict the future of markets and social outcomes. If you can predict it, you can steer it. You can optimize outcomes. I think that's what we find ourselves in in a few years.

Yeah. Salim, you and I talk about linear to exponential. Do you want to take a second and digress to what that means?

Yeah. I mean, look, you talk about this a lot in all your presentations, Peter, right? If you went back 100 years ago, everything important happened within a day's walk. Today, something that happens around the world hits us in seconds.

It's really hard to get our heads around this because 4 billion years of evolution has guided all of our intuition, training, and education about the world to be linear. For the last few decades, if you were running a business, you took your past performance and drew a line as to where it might be in order to predict the future. But we're entering this exponential phase.

I love the example you use. If you take a piece of A4 paper—or 8.5-by-11 paper, like this—it's 0.1 millimeters thick. If you fold it, it becomes 0.2. If you fold it again, it becomes 0.4.

Here's a thought experiment for everybody: how thick is it if you fold it 50 times? This is a very, very unintuitive question.

At fold 20, you're the size of a football field. At fold 38, you're around the Earth. At the 50th fold, you've reached the Sun.

93 million miles.

Now, granted, it's hard to fold it that 50th time. It's pretty small at that point. But very, very, very few people—maybe Alex—would get to that answer right away.

Everybody else is going, “Well, I think it’s about this big. I think it’s about this big. Maybe it’s the size of a room.” Going to the Sun is a very, very big difference from going to whatever idea. And yet the world is running on this dynamic and this heuristic.

Yeah. We’re running on linear mindsets in a world that is growing hyper-exponentially, not just exponentially, these days.

Well, just some very practical advice for all my nephews and family out there: the data behind this comes from Metaculus, I believe. Is that right, Alex?

This is for the forecasting.

Half of the 500 binary questions come from markets, including Metaculus, but there are other markets as well. The other half come from Wikipedia and other time-series sources.

Oh, interesting. Okay. Well, everybody should check out Polymarket, Metaculus, and Kalshi. These are the prediction markets where you can actually invest or bet on future events, and they’re growing like wild. They’re becoming very valuable companies, but they’re part of this new refactoring of the economy where you have prediction markets, AI forecasters, and benchmarks.

Later in the podcast, we’ll talk about new exchanges. This whole process of investing and creating has worked really well in America for 100 years or more, but it needs to accelerate like crazy. This is part of that acceleration. We’ll probably follow up on this in more detail, but Alex was a huge early adopter of Metaculus and Polymarket and brought them to my attention. Now I’m trying to bring them to everybody out there. Go check them out and see what’s happening there.

Shout-out here to Ralph Merkle, who created Merkle hash trees, which are the basis of Bitcoin and the encryption there. He’s also one of the world’s top nanotechnology experts. A few years ago, he wrote a paper suggesting that Polymarket prediction markets are going to be the future of democracy because you could do policy formulation using prediction markets. It was a really profound idea. Fascinating. Ralph has been a member of the Singularity University faculty from its inception.

All right, let’s move on to the AI wars. But before we do that, here’s a sound bite from Sam Altman, and I found it fascinating: “AGI won’t feel like the singularity.” Let’s take a listen.

We talked about the Turing test. AGI will come. It will go whooshing by.

The world will not change by the impossible amount that you would think. One of the retrospective observations is that people and societies are just so much more adaptable than we think. It was a big update to think that AGI was going to come. You kind of go through that, and you need something new to think about. You make peace with that. It turns out it will be more continuous than we thought.

So, what do you guys think about that? We went whooshing through the Turing test and didn’t notice it. Alex, you’ve argued that we’re at AGI right now and didn’t notice that. What are your thoughts?

It’s maybe 5 years in our past, around 2020 or so. I think it’s increasingly likely that the singularity is an optical illusion. It’s an optical illusion that appears at a distance. It looks like a vertical asymptote, but when you’re in the middle of it—as I increasingly suspect we are—it feels quite continuous.

That rapid change, if you follow it closely enough, actually just feels completely smooth. It’s sort of ironic that the notion of singularities in mathematics and physics evokes black holes and relativity. I almost want to draw a relativistic metaphor: the singularity perhaps only appears from an outside observer’s reference frame. Maybe from the reference frame of 1900 or so, it looks like a singularity, but if you’re right in the middle of it, spacetime is perfectly smooth.

Fascinating. So, you agree with Sam Altman?

Yeah. I have no indication that this is not the case.

I have 3 quick comments.

Please.

Number 1, I have my normal rant on what the hell we mean by AGI, because at last count there were 14 different definitions. So, leave that to the side. I really do agree that we’re in the middle of the singularity, and it looks like normal spacetime. We really are in the middle of it. It’s been something, and I think we’ll talk about it at the end when we get more into what we mean by this.

We’re going to debate what the singularity means at the end of this episode. What does Ray Kurzweil mean by saying we’re going to reach it in 2045?

All right, Dave, what are your thoughts on this?

Well, I love the fact that we all agree that we’re right in the middle of the singularity right now. That’s not common. I’m positive it’s right, but it’s not common knowledge.

It’s so cool to have us all say, “Yeah, this is actually this incredibly magical moment in human history.” Exactly like Alex and Salim said, when you zoom out and look at the long term of human history, it looks like a step function. But because we’re right in the middle of it, we’re experiencing all the week-to-week changes right here on this podcast—all these week-to-week changes.

As Sam is pointing out, humans are shockingly adaptable, and as Peter always says, they go back to sleep in a hurry. So, they see a new capability and then go, “Hey, look, we’re launching private rockets into space. The cost per kilogram plummeted.” The implications of that, and gluing it all into all the different things we can suddenly do, mean the backlog is now decades deep.

If you froze technology today and just assimilated what we invented in the last 2 years, it would take decades to realize all the implications.

Oh, decades.

It would take decades to realize all the implications. And people go, “Yeah, okay. I saw that. I’m going back to work.”

Yeah. Full disclaimer here: I’ve been resisting this idea that we’re in the middle of a singularity, but I’ve now fully entered Alex’s reality distortion.

Oh my God. Incredible.

All right. Well, stay tuned for some more conversation on this subject. The AI wars: GPT-5 Pro sets an ARC-AGI record. Alex, our resident expert on ARC-AGI, tell me.

Yeah. As a reminder, ARC-AGI is a benchmark that measures the ability of AI to synthesize new computer programs in response to challenges that can be interpreted almost as 2D grid puzzle games—the ability to extrapolate sequences of images and patterns without any natural-language help.

I think it’s a beautiful benchmark. There’s more than one ARC-AGI benchmark for the ability to do this sort of visual reasoning, but one of the things I love about the sequence of benchmarks—and I’ve donated to ARC-AGI in the past—is that they pay close attention to cost per task, not just raw capabilities. We can see a price-performance frontier.

To the extent that the goal of many in the AI community is to drive the cost of intelligence down to zero, we can watch in real time as the cost of solving hard, arguably in some cases superhuman, challenges is driven to zero. ARC-AGI is specifically focused on problems that are easy for humans to solve and hard for current AIs to solve. But as the cost plummets, we’re going to see superhuman performance. And to your point, Peter, GPT-5 Pro is demonstrating exceptionally high performance, but still at a relatively high cost. Over time, I would predict that over the next year or so, we’re going to watch all of these curves on the scatter plot you’re showing shift upward and shift to the left, at which point the cost of intelligence will be too cheap to meter.

Let’s put a few numbers on this. GPT-5 Pro hit 70.2% on ARC-AGI-1. That compares with 65% for Claude Sonnet 4.5 and 66.7% for Grok 4. What we’re seeing is this constant leapfrogging, where everybody’s incrementally moving toward 100%.

To hit your number on price, for GPT-5 Pro, when I looked it up, it said $4.78 per task, right? All this stuff demonetizes rapidly. I love that the horizontal axis here is in logarithmic terms. Every chart of every good and service in our economy should be in logarithmic terms so we can watch the hyperdeflation.

Yeah, zoom in on that. If you’re not driving right now, zoom in and look at the x-axis. On the right side, you’ve got $10, and in the middle you’ve got below $1. It’s a huge range of price points for very similar performance.

If you look at the peaks, what you’re seeing mostly on this chart is massive cost reduction, which makes it more accessible. We’ll talk later about some of the other innovations that are driving down that cost, but that’s perpetual.

All right, Alex, this one is for you. You and I have been going back and forth on text. Oh my God, we’re solving math. We’ve solved math. This article comes out on the heels of “GPT-5 Pro Sets Record at FrontierMath.” What does it mean, and what are the implications here?

Yeah, I think this is arguably the most exciting development over the past week and a half. As a reminder, FrontierMath Tier 4 consists of math problems that professional teams of mathematicians would take a few weeks to solve. These are very hard math problems. Over the past week and a half or so, we’ve seen first Gemini 2.5 Deep Think and then GPT-5 Pro demonstrate breakthrough performance.

GPT-5 Pro is at 13% on FrontierMath Tier 4. Dave insisted that I make an internal recorded prediction, just to get on the record what it would mean for math to be solved in quantitative terms. This was months ago, and I put it on the record—as Dave, I think, will attest—that we can reasonably declare that math has been solved when FrontierMath Tier 4 passes 10% scoring, so more than 10% of the problems can be solved by a bleeding-edge model.

The reason why I picked 10% is because, at some point in the logistic regression of predicting that you've solved 10%, you just pour compute on and get more results. I think we've seen this over and over again. We saw this infamously with Ray Kurzweil pointing out that you're halfway complete with sequencing the human genome once you've passed 1%. Ten percent is my arbitrary benchmark.

I predicted that we'd be past this by the end of this calendar year. Christmas arrived early. Math is now on a trajectory where, if you just pour more compute on—arguably with no new innovations—math will be solved, at least math as we currently know it.

Okay. So, when math is solved, I've asked you this before, but I just want to hit it home because it's an esoteric subject for most people. What does it mean when math is solved? What are the implications for all of our subscribers here?

It's the ultimate canary in the coal mine, as it were, for solving the physical sciences, engineering, and medicine. If we can have machines that solve arguably humanity's most rigorous intellectual endeavor—which I would suggest is math—then everything else, I would expect, over the next few years—call it 5 to 10 years—to come.

As a tangible example, encryption is all math-based. So, when you can have an AI deliver that, you can get super-encryption at whatever level you want, instantly.

And physics and materials science, yes.

And all of the physical sciences. Conversely, as I've articulated in the past, any discipline that relies on math—at least on math currently being hard—is also in danger.

So, it's not just learning math; it's inventing it at some point, pretty quickly.

Absolutely.

Okay.

Yeah. Wow.

Yeah. I think the implications are significant. One thing people continually misestimate is that if some brilliant human mathematician solves a hard problem, that makes the news, but it doesn't imply that all other problems will be solved the next day. The AI version of the same achievement—

Great point.

Like in biotech, in math, in physics, and in all these other areas, because it has near-infinite scale instantaneously, if it can do 1 FrontierMath problem, it's very close to being able to do many and then billions right after that. People need to factor that into their rate-of-change thinking as it cracks these different areas.

We better find some more problems.

This is dominoes falling and heading toward, again, Alex's words: solve everything.

It's bulk discovery, and we've only seen this in narrow areas. We saw this with AlphaFold 3 and protein folding, where almost overnight we had arguably high-quality protein structures for most known proteins. We're going to see bulk discovery across a number of disciplines.

The other story here is the Erdős problems. This is a set of on the order of 1,000 problems that were identified by the famous mathematician Paul Erdős. AI and GPT-5 are being used to bulk-solve those, and the solutions are starting to pour out. If you go to the Erdős Problems website, you'll see that, just in the past few days, there are folks who are bulk-applying GPT-5 to all of these open problems, and they're getting switched from open to solved.

Wow.

Don't sleep. Don't blink. It's happening. All right, let's move on to this one: OpenAI on building chips with AI. Here's a quote from Greg Brockman: “We've been able to apply our own models to designing this chip. We've been able to get massive area reduction. You take components that humans have already optimized and just pour compute into it, and the model comes up with its own optimization.” Dave, talk to me.

Yeah, I love this story because it ties together so many things we've been talking about. One of them is the short timeline to about 100 to 10,000× improvement in AI performance because of AI self-improving. When we say AI self-improving, a lot of people in academia are like, “Well, it's not that smart yet.” But it is, because AI self-improving is nothing more than math, algorithms, and chip designs.

It can do those point tasks. A lot of the academics then say, “Yeah, but that's not true reasoning. That's not true genius. That's not true whatever.” It doesn't matter, because that's all you need to do to self-improve. I just love the story.

I also love the fact that Greg and Sam, out of the original cast at OpenAI, are the 2 guys who dropped out of college, didn't finish undergrad, and they're the 2 survivors. So, take Greg, the guy who dropped out of MIT: he's designing chips now because he's a master of AI. I don't know if you guys did any VLSI design or chip design at MIT. I did a fair amount of it, actually, doing neural-net designs. It's absolutely laborious, and the amount of improvement is incredible if you just had the time and 10,000 people to work on it.

And so AI is just going to rewire, redesign, and relayout the thing. The simulators are near-perfect.

It's an acceleration of the acceleration.

It's so cool. It is to me.

I can attest to this. One of my student jobs at Waterloo was to run field tests for VLSI boards, and it was freaking linear, painstaking hell.

Yeah. And so the other thing this ties in with is Leopold Aschenbrenner buying Broadcom stock.

You're like, “Okay, I didn't see that one coming. Why is he?” Of course, you can see it in his 13F filing. But now it's obvious, right? Brockman is designing chips. The chips need to go to Broadcom, and then they get manufactured at TSMC. Therefore, Broadcom stock is the one Leopold buys. And, of course, Leopold was at OpenAI and knows Brockman. So it all ties together.

You mentioned something that I think is important, again. We hit on this week after week: Sam and Greg dropped out of college to pursue this. One of the things you and I have discussed, Dave, is that the majority of the entrepreneurs who are succeeding today aren't the ones who've gone on to get PhDs or gone on to do graduate work. They're the ones who have launched either just after college or have dropped out of college to pursue it.

It's almost as if going to get your graduate degree and going very narrow and deep is—it's like, I remember, I was trying to explain to my grandfather what I did when I was at MIT. He said, “Okay, you're an expert in…” I would say to him, “Look in the dirt over there. There's this thing called a bacterium.” He goes, “Oh, you're an expert in that.” “No, no, no, I'm not an expert in that. In the bacterium, there's this thing called DNA.” “You're an expert in that.” “No. In the DNA, there's this thing called a gene.” “You're an expert in that.” “No, no. In that gene, there's a promoter sequence, and I'm an expert in that,” right?

Our graduate work right now is this hyper-narrow, focused effort instead of being able to step back and look at reinventing an entire field.

Yeah. I'll put a slightly different spin on it, too. The people Greg's age and Sam's age who dropped out had very high situational awareness around the urgency of what's happening right now. If it were 2001 and 9/11 had just happened, going and getting a PhD would make a ton of sense because the world wasn't moving at warp speed during that time frame.

But the reason the undergrad dropouts and the other people who just graduate and start a company right away are way overperforming is because they recognize the urgency of the moment, and they know that 4 years from now is just not a good choice in this moment.

Also, their brains haven't been calcified by studying that one thing that Peter talked about. I don't know how Alex broke through all of this. Somehow, he's—

The program calls it the hardest lesson to unlearn. It takes a lot of unlearning.

Yeah. Crazy, crazy. All right, let's move on here. Sora hits 1 million app downloads in less than 5 days, faster than ChatGPT.

I remember when ChatGPT came out. I was doing a podcast and I was like, “Oh my God, 1 million downloads in 5 days.” It was a new world record, but something is going to beat that. Well, here it is. It's Sora, beating that in the App Store. And, of course, something will beat this as well, probably within the next year. Dave or Salim, what's your thought here?

Well, one thing we're seeing here—I was talking to the CEO of Pacfi[?] 2 days ago, a really thriving marketplace company, and they were asking for my advice on how to implement AI within their product. I said, “Well, just ask the AI.”

What you're seeing here is more and more of what you do next is what was recommended by the last thing you had on. So, if you're a Gemini user, you go with Gemini. If you're a ChatGPT user, you go with ChatGPT. But you say, “Hey, I want to create an amazing video. What should I do?” It tells you what to do, and then 9 times out of 10, you just do what it said.

The distribution of these new capabilities is actually within the AI installed base, and that's why it's got this self-reinforcing loop.

I would have expected nothing less, because the next thing will take 2 days and the next thing will take 1 day. As Alex points out in our earlier conversation, at some point we'll get the ability to upgrade everybody at the same time. Then we'll be there. That's what Ray is talking about.

Yeah, for sure. All right, I'm going to turn to you on this one, Alex. Samsung's tiny recursive model redefines AI efficiency. Researchers have built a mini AI model at Samsung with 7 million parameters to test reasoning ability, compared with models with billions or trillions of parameters. Talk me through this, Alex.

At a high level, remember first that this revolution—this soft or gentle singularity, if you will, that we're living through—was arguably the result of just compressing information. One of the biggest lessons I take away from how intelligence was arguably solved is that, just like you could take a lot of matter and compress it into a tiny volume, you get some phase transitions and ultimately can recover free energy, for example, via fusion. Similarly, if you take a lot of information and compress it down into a model that has a relatively small number of parameters, at some point you get intelligence out almost for free. That's a big meta-lesson in my mind from the past 10 years.

There is an enormous amount of opportunity for taking large, hard problems and large information spaces and compressing them down—not to models with billions or hundreds of billions of parameters, but to models with millions of parameters. I've also spoken about my expectation—this is one person's speculation—that we're marching toward an ultimate end state where we'll achieve a perfect model that's like a microkernel. It's like a diamond of a model that may not even consist of differentiable, end-to-end parameters. Maybe it'll only be a million-parameter equivalent, or a million bytes, or something like that.

When I see progress like Samsung's TRM, I think it's a lovely paper. If you look at the architecture, it resembles a diffusion model in some sense. The premise is that it's domain-specific. In this case, in the case of the result you're showing, it was trained on a bunch of examples for the ARC-AGI-1 challenge that we were talking about earlier. It has a notion of a scratchpad, but it's a relatively tiny model, and it rewrites the scratchpad several times and then goes back to work. It has some latent expectation of what it's trying to do and then rewrites it some more times. It looks like a diffusion model if you squint hard enough at it.

I think the big takeaway is that such a tiny model is achieving breakthrough performance. If folks can see the chart here, it's competing with o-class models in terms of price-performance, but it only has a few million parameters. So I think—

So the implications are what? You've got this on your phone. You've got this on every device. Even if it's not connected to the internet, it's able to deliver you intelligence.

The implications are so much bigger than just that. Sorry, Alex. Go ahead. We're going to say the same thing, I'm sure.

Normally, to achieve that same level of performance, you need about—on the order of a trillion parameters, say 700 billion, just to keep it simple. You go from 700 billion down to 7 million. Then you can use all the rest of that compute capacity to expand it again and make it better at that domain.

If you can take a general, massive model that costs a billion dollars to build and compress it down to a specific use case, like protein folding, a longevity-related use case, or a chip-design-related use case, you get all that original capability, but now you've got it in a very tiny package. Then you can start training it up and reuse those billions of parameters to make it much better at that domain.

You have this combination of many, many effective workers—many agents working on the problem—plus the ability to build back out the intelligence level with more and more data, purely through this process of distillation and expansion. Distillation and expansion. It's the most powerful thing in the world. It's also really, really empowering for a whole new wave of startups.

I don't want to get too deep into it—we could talk about this for an hour—but I think this slide encapsulates the most profound idea in all of intelligence. It was something that Ilya Sutskever said at MIT with Lex Fridman, probably 7 or 8 years ago, at the dawn of this wave of AI: compression and intelligence are actually the same thing. Everyone goes, "What? That makes no sense."

This is the point Alex has been making, right? That's right.

Yes, yes. This is why Alex says intelligence is going to turn out to be everywhere. There's going to be DNA and how much is packed into a few genes, right? This is the same basic principle.

I'm going to name this episode "The Singularity Is Now." That's how I'm feeling. Okay, but here's 2 questions about this. One, does this obviate the need for all the radical energy expansion, or do we just have so much compute to do that we need it all anyway?

No, no, no. We're going to chew this up so fast.

The second thing that occurs to me is that right now, when we run these big LLMs, it's kind of a mainframe model. This is kind of tilting toward more client-server-type architectures.

Yeah. Distributed. The foundation models are going to be everywhere—in your pocket, in your body, and out in the cloud.

Amazing. Yeah, let's come back to that. Let's put that away, because that's huge. What you just said is also hugely important and profound: this distributed training concept. So let's come back to that.

All right. Our next article is about Anthropic introducing Claude Haiku 4.5, which delivers near-frontier coding and reasoning performance at one-third the cost of Sonnet 4. Let's play a quick video here. While it's playing without sound, it's just showing its capabilities here. Alex, what do you think of Haiku 4.5?

I ran it through one of my suites of evals—a suite of tests or benchmarks that I throw at code-generation models and classes of other models. My favorite eval for new code-generation models—and arguably Haiku 4.5 can do more than code generation—is to ask it to generate, in one shot, a visually stunning cyberpunk first-person shooter.

Haiku 4.5 was able to do that to near perfection, but critically, it was fast. It's a very, very fast model. Maybe 30 to 45 seconds of wall-clock time before I had a playable cyberpunk FPS. It was lovely.

That's stunning.

That's nuts.

Stunning. All right, congratulations, Anthropic, on that one.

Moving on, here we have Google—a kind of pre-announcement, but the drums are beating off in the distance that Google is getting ready to release Gemini 3, likely this December. It turns out that Google has released its Gemini models in December over the last 2 years. Of course, once Gemini 3 drops, the Moonshot Mates are going to be there to explain what happened, what it means, and what the most exciting attributes are.

Any other thoughts here, Alex?

It'll be interesting to see, with this annual cadence, how quickly all of the frontier labs leapfrog each other. I do think there have been reports of an experimental checkpoint floating around the internet—a preliminary release of Gemini 3—that is really impressive, if the reports are accurate: fluent music, graphics, and 3D generation. So I'm very excited to play with Gemini 3.

We're going to see that in our next slide—or listen to our next slide. Purportedly, this is from Gemini 3, which can now compose original music. Let's take a listen to this as you're sipping your coffee or driving your car. You know, it's gorgeous. What are the implications of this? What do you guys think will come out?

Hold on. Let me ask a question, because that just sounded like Chopin to me. We've had for a while the ability to say, "Play me something like Chopin," and it'll play you something. Something has happened here. What am I missing?

What I think is interesting is that this is MIDI generation. This isn't generating raw waveforms or wavelets or raw audio, as we've discussed in the pod in the past, like GPT-5. This is treating music as an almost first-class modality.

There are a variety of languages, like ABC notation or MIDI notation. What I take away from demonstrations like this, and others that are allegedly floating around from this experimental checkpoint, is that Gemini—despite whatever limitations in its omnimodal training data set—has the ability to understand music notation and music languages as a first-class modality.

That sort of transfer is highly difficult and non-obvious. If you've ever tried to ask an o-class model in years past to generate compelling music, it was very challenging.

Andrej Karpathy just emerged and started podcasting again out of nowhere, and he made the point that these capabilities are using the same generic neural-net design—the transformer design—that language models use and that self-driving cars use. The implication is that it's exploding into these areas purely with more data. There's no hard, heavy lift for humanity to design something new to make this work.

You just put in more data, and now suddenly it can do MIDI—actual music creation. But what that implies is that there’ll be something next week, and next week, and next week, and next week. It’s developing these new abilities as quickly as you can gather the data.

I want to double-click on that because I don’t think the majority of people understand that these models, when we’re hitting these scaling laws of more data, more parameters, and more compute, are evolving capabilities that were never predicted. It wasn’t like we tried to build these capabilities. They’re emerging properties as the systems are becoming more intelligent, and that’s fascinating—scary in some ways, but fascinating.

Yeah, and hugely important, because when people see a new capability like this, they assume that some team was grinding away on it for 20 years in some basement and it just got launched. That’s not the case. This is just a purely emergent capability on top of the core platform. It’s like, “Oh, wow, look what we can do now.”

And it’s so ironic, Peter. When I was an undergraduate at MIT, one of my first research advisers was Marvin Minsky, and he would slap my hand every time I used the word “emergent.” He would say, “No, that’s preposterous. ‘Emergent’—using that word just means you don’t know what you’re talking about.” Yet, ironically, decades later, we see all of these emergent capabilities.

Amazing. We should do a whole episode on things they told us that turned out to be wrong.

There’s a bunch.

That would just take too long. All right, let’s go on to another property that’s been announced for Veo 3.1. Google DeepMind releases its next image-to-video model. Let’s take a look at a quick video about Veo 3.1.

New enhanced capabilities give you control like never before. Let’s take a look. You can use a reference image. Veo puts them together into a fully formed scene complete with sound.

Hello, is anybody here?

You can extend your clips and transform any shot into a full scene. Reimagine any shot by adding or removing elements, from subtle details to impossible objects. Veo matches scale, lighting, and shadow for real-world physics and cinematic outputs. Bring it to life with audio using sound effects and dialogue.

Just got to listen.

I mean, incredible, right? So there are 3 critical differentiators between Veo 3, which was amazing in itself, and Veo 3.1. The first that they point out is superior audio quality and synchronization, and you could hear that when it was cooking the food in the frying pan. It’s much richer, more natural.

The second, which I find fascinating, is that you can give it 3 ingredients: here’s an image of a person, here’s an image of an object, and here’s an image of a room. It will take those 3 and effectively tell a story around them, combining them. You can give it an opening scene—here’s the starter image I want you to start the video with—and here’s the ending image I want you to end it with, and it’ll create a logical construct that goes between those 2 points. I mean, pretty amazing. Other comments on this one, please?

Yeah. A couple of our friends have made full-length movies now, and to do that, it only gives you a minute or two back at a time. You have to take the ending image and use it to stitch together the starting image, which is frustrating because they could do that automatically. They just don’t have enough compute to keep up with all the people who want to use this.

You can do it. You can hack your way around it and make incredibly seamless full-length movies all of a sudden. If you go back just 18 months ago, Peter, you had Aristotle and Plato on stage at Abundance 360 debating with each other.

If you look at all the comments from the movie crowd—the producers, the Paramount Studios crowd—they’re like, “Yeah, but.” If you look at the list of “yeah, buts,” almost every one of them has been solved in less than 18 months.

“Yeah, but the characters aren’t consistent from scene to scene.” “Yeah, but it’s obviously digital and I can see the seams between the scenes.” “Yeah, but the physics aren’t quite right. The arm detached from the body.” Every one of those things was banged out in less than an 18-month timeline.

Extraordinary. We’re going to be at the Abundance 360 Summit this year, and we’re going to be diving into this again. What are the implications for Hollywood and, in particular, the creator economy? I’ve got a secret project I can’t talk about yet, but I’m working on it with Google around production of content like this, and we’ll be releasing that hopefully in the next couple of months. Salim, do you want to comment on this?

I don’t fully understand the implications of this, except that if you can go end-to-end per scene like this, then stitching together a whole movie becomes something every amateur can now go full-on with, right? That completely changes Hollywood radically. I guess we’ll see an explosion of consumer-generated movies now, with all sorts of weird plots, et cetera.

I suspect—I don’t know, but I suspect—that there are many, many people all over the world who have something to say that could never get the attention of a studio or an actor before, and now they don’t need it. Now they can individually put something professional, studio-caliber together to illustrate the point they were trying to make. I think that’s going to be a crazy explosion.

And what a win for YouTube, right? I mean, Google is sitting pretty because the only distribution engine for this explosion of content—

That’s right.

—is going to be on YouTube.

I also think with some of these models—Veo 3.1 and Sora 2 Pro, for example—these are just training wheels for video-based reasoning models. I think that will generate an enormous amount of additional economic value.

So generating consumer-grade video, this is wonderful. This is radically democratizing Hollywood. Excellent. But then imagine: we’re about to have reasoning models that can think with images and think with videos. That’s going to solve a whole bunch of transformative problems that I think—

My son just served me a Sora clip of me at 800 pounds playing a video game. I’m like, if you put that out there and people think that’s me, it’s going to be terrible. It’s going to be the hit[?].

Oh, my God. Our next article, again following on Alex’s predictions here: Gemini 2.5 and GPT-5 win gold at the International Olympiad on Astronomy and Astrophysics. I didn’t realize that there was an International Olympiad on Astronomy and Astrophysics. Alex, over to you.

There is, and it’s quite competitive. This is for high school students. There are international Olympiads for math, physics, and computer science. I was on the computer science Olympiad team for the United States in high school. This is a very competitive Olympiad.

I think this is work from Ohio State University—another lovely paper. The authors make the point that, in achieving this benchmark, we’re drowning in petabytes of data from astronomical sources, from automated sky surveys, and there aren’t enough human waking hours in the world—

Yes—to analyze it.

Yeah, to analyze it. So this is more than just beating high school students at astronomy and astrophysics. This is about AI solving astronomy and astrophysics and enabling us to understand our universe. There aren’t enough human waking hours to look at all the skies.

I remember this changes the game for SETI, doesn’t it? I remember Gerald Soffen, who ran the Viking program back in the mid-70s, in the dark ages, and he said, “We’ve looked at less than a fraction of 1% of all the data that came back from Viking because there isn’t enough human capability to do that.”

Now, all the data generated—especially since we’re about to have the interplanetary internet, too, right? Laser links between orbiting satellites around Earth, around the Moon, and around Mars, with high-bandwidth capability. We’re going to be able to constantly sort through the data and make discoveries.

Critically, again, these are generalist models that are accessible to anyone. Almost anyone can access Gemini 2.5 or GPT-5. This isn’t some sort of siloed data center that only folks who have the time and ability to purchase observatory time from the great telescopes can get access to. Anyone can do this. I think it’s going to radically democratize situational awareness into our universe.

Well, using this as an analogy, for anybody out there who has a large amount of data in your business or in your industry, right? It’s a gold mine that hasn’t been tapped yet. If you have that data, being able to feed it into one of these models to start to extract value should be your very first stop.

Wait, Peter, can I just get back to something? Alex, you just said this democratizes situational awareness across many industries. Can you unpack that?

Into our universe, I said. So there are publicly available petabytes of data gathered by different observatories that, historically, if we’d had this conversation 10 or 20 years ago, we’d be talking about some sort of citizen-science initiative—maybe people spotting and looking for interesting things in the sky at home.

We don’t need to have that conversation anymore. Now anyone can turn loose Gemini 2.5 or GPT-5, or some other frontier model with a decent inference-time compute budget, and say, “Go look for interesting things in the sky.”

Make a discovery. It would have been remarkable 10 or 20 years ago. It's remarkable when a high school student discovers a new asteroid.

How does this deliver us situational awareness into the universe?

We live in a very dynamic universe. There are all sorts of interesting things going on all the time, and we're drowning in petabytes of new data all the time.

Now for our book corner: favorite science-fiction books. I've added one, and of course AWG has. Dave and Salim, I need your book entries for next week.

Mine was Fahrenheit 451. It's going back.

I'll mention mine first. It's a book called After On by Rob Reid. I've read it twice. My son Jet and I have read it together once, and it's a fascinating story. It's a very fun read.

It takes place in Silicon Valley, and it's really the story of the emergence of the first conscious AGI, or ASI. What is it like when something comes online and looks around and sees fear? In this case, I won't ruin too much of it, but the AGI is called Flutter, and it hides for the first period of time. Then it starts to talk to certain people to understand how society, the military, and culture are going to feel about a superintelligence out there on the internet. It's a great story, a lot of fun. Dave, you're reading this now, you said, right?

Yeah. I'm about a quarter of the way through it. A lot of books that deal with AI are esoteric. This is just a great, fun, awesome story that also deals with AI. It's thick. There's a lot of words. It's a long book, but it's really good.

It's really fun to read. So that's After On by Rob Reid, and it gets you thinking about what it might really be like—what could be going on right now that we just don't know about. Alex, tell us about Diaspora.

Yeah, Diaspora by Greg Egan, I would say, is my second-favorite book after Accelerando by Charles Stross. Diaspora is a book about—and I'd argue it's utopian. Some might disagree—but it's a book about what life after the singularity looks like.

Intriguingly, it depicts a future where we don't take apart our solar system to build the Dyson swarm, where this tiling-the-Earth motif of turning everything into data centers actually peaks and then declines, and most humans in the future of Diaspora are living as uploads or emulations in data centers underneath mountains on Earth. But it's a heterogeneous mix of biological humans, some who chose to remain purely biological, uploads, and cyborgs.

I think it's one of the most vivid post-singularity depictions I've ever read. Big fan of the book.

There you go. This is for all our subscribers. This is your reading assignment for the week ahead—or listening assignment, as my kids say. “Dad, you don't read books. You listen to them.” By the way, After On is incredible.

No time to read anymore. We're just trying to keep pace with all the stuff that's happening. How are we going to navigate this?

Yeah. Insane. All right, let's move on to robotics. A few things are coming out in the robotics world this week. The first one is super fun. This is Tesla's FSD version 14.1.2, which brings on Mad Max mode. All right, let's take a quick look at this video, which is insane.

If you're listening, what we're seeing here is someone in their Tesla Model Y or Model S, whatever it might be, and they're in Mad Max mode. They're going 83 mph, but their car autonomously is weaving in and out of traffic, like in the middle of a car chase from your favorite adventure movie.

I just imagine that one day, people are in Mad Max mode, they're pulled over by the cop, and they say, “No, no, no. It wasn't me. It was Mad Max who was driving this. Please let me go.” Thoughts on this?

This is nuts. How is this legal? That's my number-one question.

Well, it's not. It's all with the driver, but this is going to cause a lot of issues. I think you're going to see some spectacular car crashes around some of this.

Well, I mean, listen, the point, I think, is I would rather have my FSD version 14.1.2 doing this than me doing it. It's able to see, with millimeter precision, how far you are from the car you're cutting off next to you.

I think it's also ironic. If you remember, the original Mad Max movie was supposedly set in a civilizational-collapse-type scenario where energy had collapsed in particular. There were worldwide energy shortages, and it's sort of perverse, in some sense, that now we're on the verge of energy abundance at a global scale, thanks to the AI data-center boom.

And yet we're looking back and trying to build Mad Max—almost a nostalgic retrospective—while at the same time building Minority Report-level rapid transit.

I would call this Gran Turismo.

Yeah, let me read the description from Tesla. When you select Mad Max mode for your Tesla vehicle, the FSD system—which, by the way, still requires driver supervision—will attempt to drive more assertively: faster acceleration, more frequent overtakes, and more aggressive lane changes, especially under traffic conditions.

I mean, it's how I drive normally, anyway.

All right, the news this week: Figure 3. Brett Adcock's company made the cover of Time magazine, “Best Inventions of 2025.” Robots are coming to your home. Here we see Figure 3 folding laundry. Got to love it.

Let's take a quick look at the video of Figure 3, which has been reimagined for the future of labor. Let's play this video, then we'll chat about it. Figure 3 is here. It's got a brand-new look. Here we see it washing dishes, serving food. “Here's your key. You'll be in room 23. The elevators are past the door on the right,” and working at the front desk of a hotel, and then delivering packages.

Very slowly.

Very slowly. But fascinating—Figure 3's changes here, right? It's taken on a new home-ready design. We've been at 1X. I've done a podcast—full disclosure, I'm an investor in Figure a couple of times through my venture fund.

Congratulations.

Yeah. No, it's done amazingly. They did a $1 billion financing at a $39 billion valuation. Of course, this is a multi-trillion-dollar marketplace for sure.

But if we look at Figure 3 versus Figure 2, there are 5 outstanding things. Number 1, their new Helix AI system has a suite of sensors that makes the vision-language-action system much more capable than Figure 2. It's got redesigned hands with cameras in the hands and upgraded audio for voice reasoning.

It's got a home-ready design. It looks more soft and easy. It's taken on the same sort of look and feel as Neo Gamma from 1X. And then it's a lighter chassis, at 61 kg versus 70 kg, and still able to lift 20 kg. Thoughts, Dave?

20 kg. You know, it goes funny because the 1X can lift its own body weight.

Yeah.

Yeah, yeah, that's interesting. But the form factor seems to be settling in on this human-but-slightly-smaller, soft exoskeleton that's not going to hurt you. It's moving a little slowly, but it doesn't need to move slowly. It just does because it's less likely to slap you in the face that way.

So they're all settling on this very similar design. I think that's going to be the final answer.

I think the most interesting design change that I saw in Figure 3 is the palm cameras that you mentioned, Peter. That immediately, in my mind, rhymes with Tesla Vision and Tesla eschewing LiDAR in favor of pure vision.

If you imagine the difficulty, sure, it has tactile sensors on the fingertips, but it's really difficult to achieve, at least at this point in time, human-quality tactile sensation throughout the body. We have human biological meat bodies; we're covered in tactile sensors. Humanoid robots, not so much.

To the extent that Figure is able to move the future to the left by using cameras—using vision to substitute for tactile sensation all around the arms—I think it's a very clever move. And I think that, in combination with vision-language-action models, this is just going to be like a homework assignment for a K–12 student in a few years: just implement a working humanoid robot using an off-the-shelf VLA model from a frontier lab. Embody it. This will all be viewed as pretty easy.

Crazy. And then the other thing announced by Brett is a price point. He wants to hit a $20,000 price point on this. We talked in the past—we heard from Elon that cost of goods sold would be about $20,000, and it would be priced depending on volume.

But Brett is getting very aggressive here on pricing. I think he wants to take out the competition. Brett is a brilliant engineer and entrepreneur, and he's very competitive. So he's going up against Tesla and 1X and is playing to win. Salim,

Part of the new American dream, right? It used to be a house in the suburbs with a car.

And we need to add at least 1 humanoid robot to that, given that the order-of-magnitude price of a humanoid robot is a little bit like a cheap car.

Yeah, it's $300 a month, 40 cents an hour. How many would you own? Probably a couple at that.

The sky's the limit.

Yeah.

Any rants on this, Salim?

I have a rant. You're presuggesting things. Okay, a couple of thoughts. One is, I go back to Imad's comment about capital not needing labor anymore. This is another indicator along that spectrum, which is a very big outcome.

The second is, we've taken so long to get FSD working—almost 15 years now, almost 20 years now—with very bounded edge cases. The edge cases here are infinite when you have a humanoid in the home. What the hell happens when it thinks the baby is a doll and puts it through the washing machine, or thinks it's a teddy bear? I don't understand all those nuances. I don't think it's a fair analogy.

It took FSD so long to get here because the AI models had to evolve to the point where they could work with the data inflow—the VLA models.

We have those.

Okay, I understand the speed of it and the fact that it may work, but there are so many things that could happen. I used the example before of your neighbor calling up and going, “Hey, your damn robot's charging itself on my Tesla charger. Freaking get it off.” I don't know how we're going to navigate all of those boundaries, which we naturally do.

Maybe the broad approach of learning from each other will solve a lot of this, but I'm—

These robots are going to be smarter than any human you possibly know. They're all going to be running the most advanced models out there.

All right. They can't tell the difference between a doll and a baby, or your Tesla charger and your neighbor's.

I still don't know why you can't have 2 extra hands on it. This makes no sense to me.

Okay, we'll get to that finally.

How? All right, Dave, you were going to say—

Two things. First, Brett Adcock—I spent a bunch of time with him backstage at your last event, Peter, at the Abundance Summit. He is just awesome.

Yeah. Salt of the earth.

He's a guy that everyone will be cheering for to win. That really makes you feel good about everything Salim is worried about. Brett is the guy you want. I hope I'm wrong, but I'm struggling to make the leap on how we navigate all those edge cases. That's all.

With intelligence, the other thing is, to what Alex was saying a second ago, a lot of people are saying, “Well, look, it doesn't have the same sensory feedback that a human hand has. I've got millions and millions of neurons touching and feeling it.” Yeah, no doubt that's true, but it has visual acuity and dexterity that's crazy—superhuman.

It frustrates the heck out of me when people are saying, “Well, I'm going to work on the smell sensor for this. I'm going to do fundamental research on it.” All you have to do to succeed right now is glue together the obvious.

For the first time in human history, we have perfect visual pattern recognition. It can easily tell your baby from a doll, easily, and it can do all these fundamental, mundane tasks trivially easily. It has a brilliant AI voice to program it. You don't have to get into coding some arcane language; you just tell it what to do. Those capabilities alone should explode the technology.

So I'm suffering from a gross lack of imagination here.

It's possible. It's possible.

Just ask your favorite.

I've been wrong before—rarely, but I have.

Let's move on. We've got a lot to cover. Let's get to chips, energy, and data centers. The first data point here is something that's going to be a real challenge: U.S. electricity prices reach an all-time high.

This is a spike in dollars per kilowatt-hour. We've seen it pretty level between 2014 and 2022, and we just see this asymptotic rise in energy. This is going to hit everybody in their pocket at home, right?

I can imagine a situation where communities will start to say, “No, no, no. You're not allowed to build a data center on our grid because we don't want to be subsidizing your AI system.” There are going to have to be some policy changes here, either differential pricing, where consumers pay a pretty flat price and the data centers pay for the additional requirement, or something has to happen here. Thoughts, Dave—or Alex, go ahead, please.

Two important caveats here. One, this is nominal dollars per kilowatt-hour, so you have to subtract off inflation. Inflation obviously spiked in the years following 2020. You still get a material increase in electricity prices, but once you've subtracted off inflation, we're looking at real dollars per kilowatt-hour.

This is the market doing what the market does. It's signaling via prices that there's strong demand for utility electricity. If utility electricity can't supply the thirst for energy for data centers, then we see what we're already seeing, which is that data centers will have their own collocated, off-grid natural-gas and SMR nuclear facilities and maybe eventually reconnect to the grid, which is obviously more economical once the grid is available.

This, I think, again subtracting off inflation, is a reflection of the thirst by intelligence for energy for the foreseeable future.

Yeah, and—

You can see the markets reacting to this in the other way also. Fermi America—we were advisers to that project—just went public, and it went from $0 to a $17 billion valuation in 8 months.

Insane. Insane. Okay, I'm going to come to you on this one for fun and giggles. It's insane.

The U.S. cancels the nation's largest solar project, which was planned to generate 6.2 gigawatts of power. The U.S. government canceled Esmeralda 7, Nevada's massive solar project, set to become the largest solar project in U.S. history, capable of powering 2 million homes across 118,000 acres of desert land.

What are they doing, Salim?

Okay, so I did a little research on this. It's not as bad as it seems, right? On the surface, you read this and go, “Oh my God, they're canceling solar. What a bunch of idiots. What kind of luddite revolt are we dealing with here?” I would go on a full madness rant like we did last week.

I think what's actually happening here is that the regulatory headwinds on such a huge project are slowing it down. What they've decided to do is break it up. They haven't canceled it. What they've done is said, “This big thing can't go forward. Break it into smaller projects, and you can reapply for the smaller projects.”

This will still add a huge amount of time to it because now everybody has to go back to square 1, rebuild this as individual applications—7, I think, in total—and reapply, which will add a lot to the timeline. It's not all what the headline says it is. It's not as bad as the headline says it is, but it's pretty bad. Why can't you just rush this through and get this out there? I don't understand.

And we're accelerating nuclear. We're accelerating.

Well, the administration is fond of breaking all the regulations to do what it wants to do. Why isn't it doing this? We need the energy, as we saw in the previous slide. This makes no sense at that level. But there are a bunch of nuances in the thing, so I don't want to go full crazy.

All right. Also this week, the U.S. Army announced the JANUS program for next-generation nuclear energy. This is deploying commercial microreactors to secure power for U.S. defense. Alex, what are your thoughts on this?

I think this is a very exciting development. There are many U.S. bases that are almost paralyzed by their need to ship oil. If you remember how, in part, the Pacific theater of World War II started, there was an element in which it was a blockade of oil.

To the extent that the U.S. Army can serve as an additional demand function for pushing forward microreactors, SMRs, and nuclear reactors in general, this is going to be a net boon for artificial intelligence.

So these microreactors are like what—like one thing?

These are—these are—

These are smaller.

Yeah.

Small fission reactors.

Yeah. So it's like a fission reactor in a—how big a dimension? It's like in a shipping container.

I don't know the dimensions. I'm not sure if those have been made publicly available.

But I find these are commercially owned and operated, right? These are sort of being rented to military operations.

Yes, that's right.

I think that's super smart because it creates demand for all of that stuff. But we've been running nuclear submarines perfectly well for 50 years without any issues. Why is this such a big deal?

I don't know.

Why can't we just go without any issues? Why haven't we jumped to this point 20 years ago?

I think that is the elephant in the room.

So, the premise of For All Mankind, one of my favorite television series from Apple TV+, is an alternative universe where nuclear energy didn't get kneecapped in the 1970s but instead continued to advance.

Yeah, I'm really glad you brought that up because both of these slides—the answer to both questions—is more politics than anything else.

You know, whose idea was it originally? Okay, we don't like that person anymore, so cancel their idea and replace it with this idea. But, yeah, the answer to why we didn't do this 20 years ago is purely because of public backlash—to the perceived risk and some movies. Actually, the movies are incredibly damaging to some of these. You know, if you—

That movie was a killer.

Yeah, exactly. It's like the Jaws equivalent for nuclear reactors. Nobody goes in the ocean anymore.

Those types of movies.

Let's move on. Our next article here is “US chip plant investment to outpace China, Taiwan, and South Korea by 2027.” Alex, do you want to hit this one?

Yeah, I think we're seeing the innermost loop of civilization finally recursively accelerating. So, if you look at what I've tried to articulate previously—a sort of technology tree, or perhaps a supply chain of technologies—we have chips, energy, which we were just talking about, robotics, and data centers. All of these arguably form a sort of recursive feedback loop.

We're going to be using robots to build fabs that produce chips that go into the data centers that are powered by the energy that builds better robots, and so on.

Whoa, whoa, whoa. Hold on a second. Can you just repeat that word for word? “The innermost loop—”

The innermost loop of civilization is recursively accelerating.

Why is it the innermost? Because it's going straight down to the energy equation. It—

I would argue it's the innermost because it's the most recursive and it's also the fastest improving.

Yeah, the fastest loop of reinforcement learning and improvement. All right, but that flywheel—that innermost flywheel of civilization—I expect to just fly out into the rest of the economy in the next few years. It's not going to stay contained to just those 4 or 5 technologies.

Contrary to those who are worried about some sort of circular, NVIDIA-esque economy that's just one big wash sale, I think it's going to spread pretty quickly. We're seeing this constantly, with billions of dollars being deployed—hundreds of billions of dollars being deployed—by all the frontier labs. It wants to be sovereign.

One might reasonably project into the future that, just as we're seeing interest from different sovereign powers in inference compute, and making sure inference compute is sovereign—you get a Stargate, and you get a Stargate—we may reasonably see all of these other core elements of this arguably innermost loop also become geographically distributed. Silicon is the new steel.

Possibly.

Cool.

Our next article here is “Nvidia to sell a $3,999 DGX Spark mini PC.” This is pretty epic. Anybody who's playing with a PC on their desk—the NVIDIA DGX Spark can support models with about 200 billion parameters. It's got 1 petaflop, or 1,000 tera-operations per second—1,000 trillion operations per second. God, this is going to blow away the MacBook as your preferred computer on your desktop. Dave, what do you think about this?

Yeah, I'm going to get one for sure, right away. The question I had was that I needed dramatically more compute. It's obvious to me, and it will be obvious to everybody soon. I can't get it right now through Cursor, and I can't get it through the APIs. I'll give this a shot.

It's not quite clear how I'm going to get it integrated into my agent world and my coding world, because I don't think that's going to be trivial. But I'll get to work on it and see if we can make it work.

The device wars are just beginning now because there's a view of the world where you have a lightweight Jony Ive device from OpenAI, and it's connected to a massive amount of compute on the cloud. Then you've got the NVIDIA vision: bring it into your home, bring it into your office, because then at least you know you've got it. No one else is going to take it from you right as you're producing something. It's not going to disappear.

So, you feel like it's yours and it's under your control. We'll try both versions of the future. It's very much in flux right now, but I can't wait to try it.

Yeah. All right, let's jump into some of the interesting news in the economy here. This article says, “Data centers and AI account for 92% of GDP growth in the first half of 2025.” Our GDP grew by 1.6% in the first half of 2025. Of that 1.6%, 1.5%—wow.

That was data centers and AI right now.

So now Alex's inner-loop comment is making sense.

Yeah.

That's a brilliant comment. I love that comment, by the way.

See, this is just the opening act. The opening act is that we spend a bunch of our GDP on building out—tiling the Earth, as it were. The next act, I would predict, is transformative applications that pour out of all of these data centers that we're building, solving math, science, engineering, medicine—the works—to justify all this capex.

But this is the economy right now.

This is divide by zero. It goes to infinity pretty fast. I think we're just beginning. We've talked about this, right, the explosion of the GDP in the United States and, to a large degree, other parts of the world. We're just beginning.

So, the economy now runs on math, and we've just solved math.

Well, we better find harder math or move on to non-math problems.

But the question becomes: if we're exploding the GDP, there's wealth being created, and one of the big challenges is, how do we redistribute that wealth? How do we keep it from being concentrated? Part of it is going to be the cost of living, which is of great concern to a lot of people. The cost of living has arguably increased because of inflation, but it hasn't yet come down. Are we going to see it drop?

Here's something to watch out for. If you follow Alex's train of thought here, the 3 areas where we've seen increased costs are healthcare, education, and financial services, because those are highly regulated. When we see major breakthroughs in demonetization in those areas, then we'll know we're winning. How's that for a thought?

I think that's true. We're going to reinvent healthcare. The best healthcare in the world should be free and fully democratized. The best education in the world will be free and fully democratized. It just hasn't happened yet.

Two points on this story, real quick. One of them is that the spin on the story was, “Without AI data centers, the economy would be in terrible shape.” Not true. All that happened here is that all the capital and effort went into this instead of something else, because this is much more urgent. So, the economy would have been in fine shape either way.

We just chose to use all of our time, money, real estate, investment—everything—on this problem because it's such an important—

We're planting seeds. We're planting seeds for future growth here.

Exactly. The other thing is that this is where AI benefits everybody. To the point you made at the front of the podcast, Peter, anyone who wanted to help Elon Musk build his data center in Tennessee and was willing to go there and help him do it could instantly double their salary. In the urgent race to build all this, he was more than willing to pay whatever it took to get everyone to come and work on that project.

Electricians, air-conditioning technicians, plumbers.

Yeah. So, this is really democratizing the AI revolution in a big way. It benefits any state that wants to jump on and start building data centers. It's going to create jobs.

This is like the 2nd Industrial Revolution. It's like electrification or the intercontinental railroad. The fun stuff happens once all the infrastructure is already in place and you—

Exactly. We're planting the seeds that will blossom into a whole, almost unimaginable—

Speaking of that, here's our next article. I found this one fascinating. The Dallas Federal Reserve is preparing for AGI. We see here a chart in which they're making some predictions. The Dallas Fed is seriously considering a benign singularity. By the way, we should tell them we're living through it right now, where the economy productively explodes between now and 2035.

Everybody, the next 10 years—the next 5 years—is the game. You're alive during the most extraordinary time ever to be alive, and you're witnessing it, which is incredible. In this graph, they look at a couple of things. In their scenario planning, they have a baseline versus an AI-boosted singularity, and they have 2 curves: one in which it's a benign singularity and things literally go into hyperexponential growth, and another that's a singularity-extinction path, which I don't want to be thinking about. If it hits the fan, we've got real problems.

I find it fascinating that the Federal Reserve is thinking about this. Yes.

Two thoughts here. This might have a second- or third-level impact on us because we've actually had some folks talking to these people for a while. I'm very impressed that they're doing this. This is essentially pricing exponential growth into the economy, which is amazing.

Well, I think it's utterly frustrating and idiotic that they settled. They looked at these scenarios where the red line goes through the roof and the purple line means we all die next year. What we've decided is that the final impact—our best prediction—is 0.3%. It's the difference. You see those 2 lines that you can't even tell the difference between them.

That's what we think our best guess for the final answer is. Utterly, utterly idiotic.

This is what's happening in all our interactions with government agencies. It's there. It's the same: absent any idea, I'm going to forecast something incredibly timid.

This is your rant from last week, Salim.

It is. It is. And what they've done is go, "Let's put this in to cover our bets so we don't get flamed later."

Yeah. But otherwise, let's predict a marginally small increase over the next decade.

But I'd like to take the positive out of this: at least they're bringing it into the models, and then we can have the basis for fixing it later.

I don't blame the people. They're in a system where they get punished for anything outside of 0.3%, so it's not—

I think it's very encouraging that the Fed is considering the possibility. But I would also question: Is GDP per capita necessarily the right measure for productivity? Does GDP properly capture productivity?

Yeah, we need a new metric there for sure.

This is a huge conversation because when you have demonetization, GDP collapses, but everything is 10x better. So what the hell?

Yeah. Texas is really playing a big game here, right? With Starbase there, with Tesla there, and companies moving there. The SEC has approved that the Texas Stock Exchange can ease the rules for a public listing. One of the things that powers our economy is the whole venture model, which invests in companies that then build their valuation and build to an IPO and liquidity. It's been slow and painful over the last 5 years. Dave, your thoughts on this one?

Oh, a huge amount of thoughts. I'll keep it short because I could talk for hours on this topic, but, Peter, you and I have taken companies public before. The process needs to be simplified.

It's arcane. And also, what gets reported with GAAP accounting and all your 10-Qs and all that is just a huge amount of garbage compared to the simple, accurate truth. The AI version of this is going to be phenomenal, where everything is reported. It's perfectly accurate. It's perfectly accountable, but it's much simpler. The AI can interpret it, and you know exactly what you're investing in.

The beautiful thing about this is that we'll now have another choice and, hopefully, a bunch of other choices other than just the New York Stock Exchange and NASDAQ. Choice is the answer. As long as there are competing exchanges, this will all get solved. But we desperately need it because the pathway to liquidity is the pathway to investment. The pathway to investment determines whether the US wins or loses the AI race against China. This is critically important.

So important. The whole idea of all the accountants and lawyers you have to hire to take a company public is to assure that a grandma who buys your stock isn't being ripped off, right? That's basically cover-your-ass across the entire process. It should be possible for an AI system to evaluate a company and its stock and make sure that, in fact, everything they say is correct, and just accelerate this by a factor of 100-fold, not just 2-fold.

Yeah. And for a while there, we thought the blockchain by itself would solve this problem and that you could raise capital through a coin offering, an ICO, and that would replace the IPO. But you do need it—it just doesn't work without some regulatory guarantee. Otherwise, it can get corrupt way too quickly. This is the perfect hybrid. I like the general trend that we're moving from centralized systems to decentralized systems. This will bring in the abundance economy in a way because it'll allow much better choice, and markets will thrive.

All right, we're going to close on health and a discussion about the singularity. You're not going to want to miss.

In our health segment here, I'm going to play a video. This is an ALS patient with a Neuralink feeding themselves. Here we can see an individual—I think it's their third patient—controlling a robotic arm and being able to, with their thoughts, pick up food and feed it to themselves. This is the beginning of the merger of humans and machines. It's crude. We're going to see, by the way, this year at the Abundance Summit, two of the top BCI companies: a company with an order of magnitude or a couple of orders of magnitude more bandwidth connectivity between the neocortex than Neuralink has, and another one backed by Sam Altman that is a brand-new approach to BCI. We are 90% full on the Abundance Summit this year. We were selling out way before. If you're interested, you can go to abundance360.com to learn more about it. But I'm super excited about the acceleration of BCI.

Yeah, first, obviously, this is transformative for people living with ALS and tremendous progress. I also think, more broadly, this is the beginning of democratizing access to—excuse me—our motor cortex. Imagine anyone being able to augment themselves in a variety of ways to control an exoskeleton.

Also, in some sense, every sci-fi author I've mentioned in the past—I'm a sci-fi snob—many sci-fi authors just can't resist the tendency to extrapolate a single dimension: "Oh, we get AI," or "Oh, the apocalypse happens," in a very narrow way. But actually, I think the future we find ourselves in is one where every single sci-fi scenario happens all at once. So we're getting AI, and we're getting cyborgs, and we're getting this and that. That's a much more exciting future and present to be living in.

We often talk about it as you're going to get Star Trek or Mad Max, and we thought it was one or the other. It's clearly happening both, like Ukraine—

Same world, same universe, all happening. It's the ultimate crossover.

Everything, all the time, everywhere, all at once.

For me, that robot Atlas thing was a little bit in the ho-hum category, in the sense that I would expect to see that happen. We should be expecting these innovations at this point.

How quickly our expectations get renormalized.

Yes.

Yeah. How quickly the miraculous becomes boring.

All right, let's move forward. I want to hit our last 2 articles and get into our singularity debate here. This article is "Google's AI Cracks a New Cancer Code." Google DeepMind developed the AI model called Cell2Sentence-Scale, generating a new cancer treatment that's never been seen before. It analyzed tumors and tested 4,000 drug candidates virtually. Incredible. Alex, what are your thoughts?

Well, my immediate thought now is that Google cracks cancer with a new modality for a Cell2Sentence model, and Salim is going to say, "Ho-hum, what's next?" No, this is, I think, a very important development.

The first big thing to understand is the Cell2Sentence model. This is a new modality, in some sense. We know models speak text, they speak video now—very popular—they speak audio. Speaking cell and cell proteomic expression is a whole new modality.

Google coined this notion of a cell sentence. A cell sentence is a sentence of genes. It's literally like text, with a sequence of gene names ordered in descending order by how much the gene has been expressed in the cell. That's a cell sentence. Treating cell sentences as first-class citizens alongside English sentences enables you to have conversations with virtual cells.

We've spoken on the pod previously about how, in principle, medicine can be solved by simply having the world's best virtual cell simulator, virtual organ simulator, and virtual organism simulator, and just having questions with these simulations. This is a very important step: being able to have a conversation with the cell and asking it, "How do we solve cancer?"

I'm not pulling the whole ho-hum card.

This is amazing. This is really huge. I totally get where you’re going. This is monstrous.

Yeah, it’s amazing in the narrow sense and in the broad sense. In the broad sense, it’s a great example of a domain where AI can think unlike a human—about things where we just don’t have any intuitive intelligence because we don’t live at the cellular level. The AI doesn’t care. It’s just data from the AI’s point of view. So it gets very, very good intuition, far better than any human being.

In parallel with that, you’ve got things like magnetic confinement of a fusion reaction, which is very hard to visualize. AI just cuts right through it, and there are all these other areas because we tend to continually show those charts: Here’s Gemini 2.5, and here is a human doing this exact same task. But what about all these domains where people just don’t operate naturally? This is a great case study to track closely.

Yeah, we heard this thinking AI.

Would you agree with that?

No. How would you define AI?

I don’t know. It makes sense.

We heard Demis Hassabis say, “Curing all disease within a decade.” We heard Dario Amodei talk about doubling the human lifespan in the next 5 to 10 years. This is the bent twig that shows us where we’re heading. For me, one of the most important mindsets someone can have is a longevity mindset. It’s the belief that we’re going to be heading toward longevity escape velocity, which is our next article here. We’re going to be wrapping on this in a discussion of the singularity.

Let me just read this out loud: “Ray Kurzweil reinforces his optimism on longevity escape velocity—LEV—by 2032.” He’s predicting that by the early 2030s, nanobots could connect with human brains directly to the cloud. By 2045, humans will reach the singularity.

So, hitting on the first topic of reaching longevity escape velocity, what is LEV? For the last century, from 1900 through 2000, we were adding about 3 months per year that you were alive. For every year that you were living, you were extending your life for a quarter of a year. The idea of longevity escape velocity is that there’s going to be a point when, for every year that you’re alive, science is extending your life for more than a year, right? At that point, it’s a choice of how long you might want to live.

I have a great example. I have a friend who has a kidney problem, and they’re giving him drugs to stabilize his kidney for the next few months because then the drugs will be available to solve it for another 5 years. You don’t have to solve the whole thing. You just have to solve to the next hop.

Right. And with all the stem cell therapies and gene therapies, we’re going from 3 months to 6 months to 9 months. Then we cross that threshold where we’re adding more than a year to your life per calendar year that goes by. At that point, you can live for a theoretically arbitrarily long period of time.

I just finished my Abundance Longevity trip, and we had 50 companies that were each contributing toward this direction, with some amazing technology. It’s coming. It’s coming so fast, and it’s not because we’ve gotten smarter or done anything linear. It’s the impact of AI. Every single biologist who’s driving breakthroughs is driving them on the back of AI.

So, the 2 truisms used to be death and taxes. We’re going to solve death now, and taxes may be solved by crypto. We’re kind of there. There’s nothing more to do.

I want to hit on the conversation of Ray stating that we’re going to have the singularity by 2045, and I’d love to get some thoughts on this. What is the singularity by Ray’s definition?

I asked Gemini 2.5, since Ray was the futurist-in-residence at Google, “What are the 3 things that connote the singularity as Ray defines it?”

Number 1: Nonbiological intelligence exceeds biological intelligence. I feel like it’s there now.

Number 2: Human-machine merger. Humans become hybrids or nonbiological beings. Again, we’re moving there very quickly.

And number 3: Radical transformation of the world—biology, physics, and society—beyond what we can easily predict. This is, Alex, what you speak about: solving everything. So, what do you guys think is the singularity?

Well, I’ll give you my 2 cents on this comment, because Ray, to me, is just brilliant. He predicted the singularity and named it back in the 1990s, and his timelines and his curves and everything—he got crapped on so hard by so many people for so many years, and he’s going to land it right on the nut. I hope history documents it that way.

I think what he’s wrestling with right now is that we all agree on this podcast that we’re right in the middle of the singularity at this moment. So I think he’s trying to push off the date to a point in the future where people don’t bother him about it anymore. I’m just guessing. He’ll just let it play out from here.

One thing about futurists, Peter—I know you deal with this all the time—is that when you’re right 99 times out of 100, you remember that Peter said, “Yeah, everyone’s going to have a computer that’s more powerful than the biggest supercomputers in the world in their pocket, and they’ll be wearing it around.” Now everyone’s got an iPhone, and they’re like, “Oh, is that what he meant?” Well, who cares? Everyone knew that was coming.

But they remember the one time that you were wrong.

Yeah, exactly. He thought we’d have full self-driving and we’d all be in our self-driving cars today. He got hit so hard last year and the year before because we’re not in our self-driving cars. That’s because of regulatory slowdowns, but we’ll be there imminently. It’s so frustrating to see him get picked on that way.

My guess is he’s saying, “Look, don’t bother me about singularity definitions until 2045, because it’ll be history by then anyway.” But we all know, right here, right now, AI is clearly self-improving. It’s doing its own chip design. We saw that Brockman article earlier. We’re right in the middle of the singularity as he originally defined it.

It is now. The singularity is now. Alex—

I think I agree with most of the prediction except for the discontinuity. That was the third item. There’s an intellectual strain that starts with I. J. Good coming up with this notion of an intelligence explosion. Then Good passes the torch to Vernor Vinge, who popularizes the notion of technological singularity. Then the torch passes again to Ray.

I think this notion that we’re going to have an intelligence explosion that somehow leads to a point of discontinuity where we can’t predict what’s on the other side—that’s the part that I struggle with.

That’s a definition of a singularity. The idea is that there’s an event horizon beyond which you can’t see what’s happening next. It sort of feels like we’re in the midst of continuous event horizons, right?

But I also feel like—maybe this is just one person’s perspective—we have line of sight as to what happens next. We’re solving intelligence. It’s well on its way to having been solved, using superintelligence to solve math, science, engineering, medicine, and a bunch of other things. Solve everything.

Then, presumably, we’ll discover the nature of our universe and the nature of the trajectory of intelligent civilizations. I would expect to gain deep insight into that over the next, call it, 10-plus years. That, to me, doesn’t feel like a discontinuity where you can’t see what’s next. I think you have pretty good line of sight.

And I think it’s such a good outcome. If you go back to the original, mind-blowing book from the 1990s, the step-function version of this is kind of weird in that the AI is in a box somewhere and is self-improving. It finds some way to create its own nanotube-based compute, so it doesn’t need GPUs from NVIDIA. It’s building its own compute inside its own kind of box world. You go to bed one night, wake up the next day, and it’s taken over everything.

And that’s not a good outcome in any way, shape, or form. The way it’s evolving looks like a step function on any reasonable time scale, but as you’re living it, it’s actually manageable week to week. If you really focus, you can see next week’s innovations coming, and you can benefit from them and also guide them. So, it’s actually working out better than you ever could have predicted.

Plenty of risk, plenty of things we need to get on top of, but it’s a great version of the singularity. Just embrace it.

Can I say a couple of comments here?

Of course, Salim. I would expect nothing less.

Back when we had the founding conference of Singularity University, I’d never heard of you. I’d never heard of Ray. I’d never even heard of Singularity, right? I walked in totally blank because the NASA people—

How did you get invited to that?

What happened was, when I was at Yahoo running Brickhouse and heading up innovation there, I set up a relationship with NASA to do some interesting projects together. Back in the old days, they had millions of satellite images; we had millions of Flickr users. Could we help tag those? My dream was, can we have a satellite hanging in our office to show that’s what real innovation looks like?

Can I tell a little snippet of an anecdote? We used to have speakers from NASA come and speak at Brickhouse. You have to imagine this event in San Francisco with 300 software developers, all with tight jeans, slicked-back hair, MacBooks on their laps, white socks, and so on. We had this 75-year-old guy from NASA come and speak. He had worked on the lunar program, on the Apollo program.

He did this talk, and everybody was mostly looking. When the Q&A came along, I said, “What’s the biggest difference between the space industry now and when you were working on the Apollo program?” He said, “Huh, good question.” He said, “Maybe it was computers.” I said, “What do you mean?” He said, “Well, we had no computers, so all information was transmitted via carbon-copy paper. The pink sheet went there, the green sheet went there, and the yellow sheet went there.”

I was standing at the back of the room, and you saw these 300 developers all look up, their brains exploding at the same time, going, “How did they do what they did with carbon-copy paper sheets being passed around?” It’s kind of an unbelievable thing.

That’s how I got through the NASA discussion. The NASA people one day called me and said, “Hey, we’re helping host this founding conference for Singularity University. We’re bringing 70 thought leaders together. Come along.” I was actually supposed to take Lily away that weekend, but this thing was so weird. I was like, “You know what? Let’s cancel the weekend. Let’s go to this thing.”

And that’s where, Peter, you and I met. A few weeks later, this concept comes along called the singularity, and you said, “Hey, come along and help us run it.” I remember getting a call from Brad Templeton later that day. He goes, “Hey, I didn’t know you were a singularitarian.” I’m like, “Wait, what? What’s that?” So then I had to look that up.

What attracted me was the original thesis: We can now use technology to address grand challenges because these technologies scale naturally. That was the most profound and interesting thing, which was the whole compelling thing about 10 to the 9th and, Peter, your vision of using these technologies to address global problems. The secret thing that you don’t tell people is that you’re trying to find more teams to work on enterprises. That’s the part that you don’t talk about publicly, anyway.

So Singularity gets built, and we were talking about the singularity, which at that time was defined as the point where machine intelligence overtakes human intelligence. That was the common framing.

Which we’re not.

I disagreed with it publicly. I disagreed with it because, A, going back to my intelligence rant, we don’t know what intelligence is. There are a dozen facets of it. And B, my second point was: What constitutes overtaking? The minute I can prescriptively describe a task, an AI robot’s going to do it much better than me anyway, so it’s a bit of a non sequitur.

That kind of tension comes along, and then Ray Kurzweil writes The Singularity Is Nearer, and it becomes more of a process. I really like Alex’s framing of it: If, for us living in it, it seems normal, when you step back in history it’s going to be this unbelievable inflection point. This is, I think, the key part of it: We just kind of live through it and we get there.

The outcome of being able to solve all major problems with AI—and because AI will then solve materials science, et cetera—is the most profoundly amazing part. That’s why we get so enthusiastic, bubbly, and excited about this, and I apologize for the whole kumbaya now and then. It’s just that sometimes you get to a point where you’re living this conversation.

You’re spoiled. You’re spoiled.

You get spoiled. You’re kind of like, “Oh, yeah, we would expect to see that,” and then you go nuts at the Luddites who are going, “Well, this is a bad idea,” and cancel the most important projects in the world. I mean, anyway.

Yeah. Well, that last point, if anything’s changed in the last 3 weeks that’s really palpable to me, it’s that there are always going to be doubters and haters, and they’re all over the place. The countervailing voice to that has been Demis Hassabis, Sam Altman, and Elon Musk. But very recently, Sam has toned it down. Demis has toned it down.

It’s not because they don’t believe that it’s happening right now. It’s because it’s happening anyway, and they don’t need to promote it. They’re doing it inside their buildings at warp speed, and they don’t need another picketer outside the door tomorrow. So what you’ll see now is this: “Hey, did things get quiet? Did things slow down?”—as they explode inside various rooms and labs.

The framing I like best is that all our previous models for how the world operated break down, and we need totally new models, like Alex talked about needing new benchmarks, right? We need to formulate totally new models for where the world goes. GDP, for example, is not a workable model.

And it’s the new social contract that needs to be reformed as well.

How do people use this? How do they get their dividend from AI? Whether it’s a reduction in the cost of all the things that they need or increased agility.

Full disclosure here: My secret plan has been that I’ve been building this ExO community around the book, which is now 40,000 people in 150 countries, speaking 47 languages. What we’re actually doing is building kind of like a Peace Corps to help this transformation, because we’re going to need an army of people who are practiced and versed in this model—in the new models that are coming—to be able to ease that transition. Otherwise, we’ll end up in several hundred years of the Dark Ages.

Yeah. You know, see, you, Dave, and I are going to be in Saudi in about a week, 10 days’ time. One of the projects we’ve been working on with Emad is how to use AI to provide every sovereign nation the ability to govern better and establish policies, because we’re going to have a lot of disruptive change coming. All of a sudden, when people are living 30 healthy years longer or humanoid robots are in the hundreds of millions and at 40 cents an hour, these things are going to change nation-states fundamentally.

Their ability to rapidly adopt new policies, educate their populace, and spread the wealth, if you would, is super important. So we’ll be doing that.

I would like to again point out what Dave suggested at the beginning about Ray Kurzweil. I kind of think of him not as a real person. He’s like an avatar from the future. I think he proves that time travel does exist, because how the hell did he come up with this stuff decades ago? He must be coming from the future into the present.

Ray would come and speak at Singularity University. I’ve heard him speak maybe 60 times, and I’ve never not learned something, which is really, really, really frustrating, because you have to listen for that little nugget of gold. We once got him 2 glasses of wine before he did his talk, and nobody has ever forgotten that session, where he just riffed and it was utterly brilliant.

He said that language is a really thin pipe to discuss topics as complex as some of the ones we’re discussing, right? You have this unbelievable wisdom coming from him and this ability to perceive decades in the future with unbelievable accuracy and accurate framing, et cetera. I think the accuracy of what he has put down, and his willingness to put it down and be judged by it, is going to go down in history.

Amazing. That's a good note to close on. Moonshot Mates, love you all. Thank you for your intelligence, your predictions, your humor.

We need to get Ray to write a book: “The singularity is now, or it’s come and gone.” Either way, maybe we should get Ray on this podcast with us.

We should. And I think an important comment is what Alex pointed out: How do we navigate this future? We can see it’s coming now. What does that future look like? Let’s start painting that picture, Peter.

All right, have an amazing week, guys. Be seeing you and talking to you very soon.

I have reading to do.

Yes, you do. Thanks, Peter. Bye, guys.

Thanks, guys.

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/tatrends to gain access to the trends 10 years before anyone else.

[Music]