GPT 5.2 发布、2026年企业崩塌与110万份工作流失 | EP #215

Speaking of alien creatures, I was touring Colossal yesterday. Ben Lamm—I’m an adviser and early investor in the company—and Colossal is amazing. They’ve got something like 12 different species at different stages of de-extinction. They brought back the dire wolf.

They’re going to bring back the saber-toothed tiger. I can’t wait for that. And, of course, the woolly mammoth. They created the woolly mouse, right? They’ve been able to identify the genes that are different in particular phenotypes, such as length of hair and length of snout.

It’s fascinating what they’re doing and their ability to find the closest living relative and then use snippets of DNA. They have DNA going back as far as 1.2 million years. They haven’t been able to get DNA older than that, but that’s still pretty incredible.

Didn’t Ben say that we couldn’t restore animals if the DNA was older than 10,000 years?

For example, the woolly mammoth DNA they’ve gotten ranges from 10,000 years to 1.2 million years.

Okay.

They’ve got to identify that it’s not a single species; it’s a whole spectrum of a species, because there’s evolution going on all that time. They’re trying to figure out which parts of the phenotypes—the tusks, the woolly mammoth hair, its cold tolerance, and all of those things—they can use to reconstruct a single genome, an approximation of a woolly mammoth.

Anyway, the programs are amazing, and Ben is such an incredibly good CEO. I’m excited.

So your multi-armed robot can shear the woolly mouse, and then we can make sweaters in time for the holidays.

We can all wear them on the pod.

Made by nonhumanoid robots.

All right. Welcome to Moonshots, another episode of WTF Just Happened in Tech. This is the news that hopefully impacts you, inspires you, gives you moonshot thoughts, and gets you ready for the future because that is one of our primary goals: how do we prepare you for what's coming next? A lot of AI news. Today is a special episode that we pulled together to celebrate the release of GPT-5.2. I wanted to hit on some of the top-level hyperscaler updates and battles. ChatGPT was the most downloaded app in the iOS App Store in 2025. Congratulations to them. They’re nearing 900 million active users. Gemini is catching up. Anthropic has jumped to 40% enterprise share. Accenture is going to train 30,000 people on Claude.

Elon has let us know that Grok 4.2 is coming very shortly, in the next few weeks, and Grok 5 in the next few months. OpenAI has released GPT-5.2, which we’ll get to in a moment. Interestingly enough, Google launched its Deep Research AI agent the same day that OpenAI released GPT-5.2. There’s a little bit of a PR battle going on between them all.

One other piece of data on the downloads, to give people a look at the scoreboard: ChatGPT received 92 million downloads, Gemini is at 103.7 million downloads, and Claude has received 50 million downloads. Any comments on these opening headlines before we jump into GPT-5.2?

I’m in shock this week at the capabilities. We’ll look at the benchmarks in a minute, but the benchmarks really undersell the last 2 weeks. The capabilities are just shockingly different from what they were a few weeks prior. The race is on.

When GPT-5 disappointed everybody, the prediction on Polymarket that Google would run away with the rest of this year went to 90% or 95%. Now, as Alex predicted, it’s a closer horse race. Google is still on top of the stack, but apparently Sam had something in the tank, and who knew? I’m not exaggerating: The things I got done in the last week that I couldn’t have done 3 weeks prior—just coding and building things—are shocking.

Are they pulling their punches? We discussed that in the past, where they’re releasing this much because they know Grok is coming out next. Then they release the next segment to compete directly with it.

They are totally pulling their punches. They’ve absolutely been holding back. I think it’s because they’re starved for compute and afraid to roll out addictive capabilities that they just can’t deliver on.

Alex experienced this, too. Yesterday, we were going crazy with 5.2, trying to see what it could do, and then it was like, “Sorry, you’re done for today. We’re out. We’re out of compute. Sold out. No gas in the tank.”

The competitive pressure is forcing them to go to code red and come out with things when they would normally want to hold back, wait until they can find the data-center compute, and wait until Chase Lochmiller finishes Abilene. They just don’t have that choice with the competitive pressure on each other.

Maybe just to comment, I think that at this point, if you’re OpenAI and you have your purported code red, you’re in a hurry, and you’re in a bind. GPT-5.1 came out only a month ago, and you need to rush something to market to put perceived competitive pressures at ease.

I think there are approximately 3 levers you have. One lever, to Dave’s point, is compute. You can increase the total amount of compute allocated to given models. That, of course, comes at a cost: It comes at the cost of compute scarcity and longer response times to prompts.

The second lever you have is safety. You can turn down the safety. You can make models more sycophantic. That’s a way to improve the model.

Right, but can we get a benchmark on sycophantic models? There are a bunch of benchmarks for compromising your ideals to win the market in general.

Yeah. Call it the safety knob; it’s the second knob that you can turn if you’re in a pinch. The third knob that you can turn is the post-training knob, which can be done on relatively short notice. You can pick particular benchmarks that you want to post-train your models to do really well on.

I suspect all 3 of these—more compute, maybe some turns of the safety knob, and post-training on select benchmarks—is exactly what we’re seeing in this cycle now that we have a real horse race.

I find it fascinating. We’ve got probably the fastest-scaling consumer platform in history. We’re almost at a billion users. That just blows my mind.

It’s starting to eat the operating system. When you start to get an order of magnitude—a billion downloads—at some point, you have to ask the question: Is this AI user interface basically cannibalizing the entire OS itself?

At what point, sometime soon, is every pixel that shows up on a mobile device being AI-generated? I think we’re not too far from that.

Wow. Well, that was definitely the backstory, too, when we were at Microsoft last week with Mustafa Suleyman. I’m not sure whether that podcast is out yet or what the order of episodes coming out is.

Look forward to that one, because what Alex just said is clearly in the minds of Microsoft. They’re going to do everything and anything they can to get on this chart, and they have a lot of assets that will come up in that podcast that will give them a really good chance of getting there.

But it’s for exactly the reason Alex said: The OS—the whole base of Microsoft, the revenue driver for the last 30 years—is at risk now, and you’ve got to move to the new thing.

It’s not just the OS, right? It’s the entire app ecosystem. The end goal here is for these hyperscalers to capture the user as the only AI you need to use.

The so-called core subscription, and that certainly is OpenAI's stated strategy: to become the default core subscription, quote unquote, for consumers. Anthropic's strategy, apparently, is to focus on enterprise APIs and code generation. xAI is focusing on brute-force scaling and maybe benchmark maxing, and Google is focusing, perhaps in a more balanced way, on total-stack domination—balanced pre-training and post-training. So I think in a real horse race, which is what we're finding ourselves in among the top 4 frontier labs, we're starting to see differentiated strategies coming to market.

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. To subscribe for free, go to dmmandis.com/metrends to gain access to the trends 10 years before anyone else. All right, now back to this episode. Let's jump into the core story here today: OpenAI releases GPT-5.2. We spun up this pod for our subscribers the day after the release so we can go into detail. What does this mean? We heard OpenAI's red alert, and here's the result. Alex, take it away.

Yeah, I've been waiting all day, Alex.

Dave, you want to lead us, or Alex here?

Oh, no. I just want to say that these numbers, when they go from 80 to 90, really understate the impact on what you can do. The benchmark, when it goes from 10 to 40, looks like a big gain on a line chart, but when it goes from 80 to 90, it doesn't look like a big gain. But what you can do firsthand is just mind-blowingly different, and I'll tell you some of the things I've done in a minute.

I've been waiting all day to hear, actually, Alex. For those who are listening versus watching, here's a chart of the benchmarks comparing GPT-5.1 Thinking against GPT-5.2 Thinking. With that, if you don't mind speaking the percentages as we're going through this, Alex, that would be great.

Take a moment, just for those who don't know what post-training is, because I think it's an important one of the 3 knobs that you spoke about, and it's important for folks to understand what that means.

Sure. Let's reason by analogy to the way humans, in a conventional Western upbringing, learn. You have the baby- or infant-like learning that's approximately pre-training. The P in GPT stands for pre-trained. Pre-training is unsupervised training: you're feeding a model information about the world and giving it the goal of predicting what comes next. There's not much of a supervision angle to it, and it's not unlike a human newborn that's just taking in information via lots of sensory feeds and trying to make sense of it with very little guidance.

Then there's mid-training and post-training. Think of these phases of training as being not unlike attending primary school and secondary school, where you receive explicit supervision, receive grading, and are given particular assignments. There are many ways that you could be graded. You could be graded very granularly, like a thumbs-up, thumbs-down, or grade A, B, C, D, F. There are other ways that you can grade. For example, you can be given more of an open-ended assignment and graded on how well the ultimate final product of that open assignment is.

This sort of mid-training and post-training really became popular with the o-series of reasoning models from OpenAI, and everyone has since adopted reasoning models and post-training—not just to make humans happy, which is another form of post-training, like pleasing your teacher, but also to show that you can, via reinforcement learning and other mechanisms, solve hard problems and reason about hard problems. This is where post-training shines. This is where almost all of the alpha, if you will, in increasing model capabilities over the past year or so has come from—not from pre-training.

So, getting back to the benchmarks, ARC-AGI 1 and ARC-AGI 2 are benchmarks designed to test the reasoning capabilities of models. The R in ARC stands for reasoning; these are benchmarks designed to test reasoning capabilities. We see a huge jump—frontier-level, state-of-the-art performance by GPT-5.2. With ARC-AGI 2, reasoning is well on its way to having been solved at this point.

I think we'll probably cover this in the next slide, but the costs are collapsing as well. Maybe we'll talk about that in a minute. Just to wrap up, then, for purposes of narrating this chart, the final benchmark here, which is perhaps the most interesting of all, is GDPval.

GDPval, or Gross Domestic Product Eval, was created by OpenAI with the idea of having an eval that measures AI's ability to automate knowledge work in the general human service economy. We're seeing a jump from GPT-5.1 at 38.8% to GPT-5.2, now at 70.9%. This is the clearest indicator in my mind that the human knowledge-work economy is cooked. You heard it here: it's cooked.

This is 44 different occupations that OpenAI—and, by the way, this is all open source; you can go on GitHub and read all of the tasks for GDPval—has evaluated. There are 44 different human occupations and 1,320 specialized tasks, like creating PowerPoint presentations or Excel spreadsheets, sort of typical, prototypical knowledge work.

Moving through this chart benchmark by benchmark, we have SWE-bench Pro, which is a software engineering benchmark. We see a modest improvement between GPT-5.2 and GPT-5.1, perhaps attributable mostly to more compute and a little bit more post-training and/or distillation. We have GPQA Diamond, a modest increase from 88.1% with GPT-5.1 to 92.4% with GPT-5.2—again, so far, pretty modest.

We have CharXiv reasoning, a larger increase. This is scientific reasoning. It could be post-training, but it's not a benchmark that I pay super-close attention to. Then we get to FrontierMath Tiers 1 through 3, which are easier math problems. And then one of my favorite benchmarks of all time, FrontierMath Tier 4, which is research-grade problems in mathematics that are supposed to take professional mathematicians several weeks to accomplish.

I often point to FrontierMath Tier 4 and progress on FrontierMath Tier 4 as indicative that hard math is being solved. Focusing on FrontierMath Tier 4, we see Gemini 3 Pro getting approximately 19%, GPT-5.2 Thinking getting 14.6%, and GPT-5.1 Thinking getting 12.5%. This is actually a win in my mind.

This is a win for Google and a loss for OpenAI. OpenAI has had a month to attempt to scale up, to beat Google in this horse race at hard, closed math challenges—professional-mathematician-grade challenges nonetheless—and couldn't beat Gemini 3 Pro. It's not as if these problems have been a state secret. In fact, OpenAI actually sponsored Epic's creation of the FrontierMath benchmark. So OpenAI has had, in some sense, privileged access to all of FrontierMath and still couldn't beat Gemini. I think that's pretty instructive.

Moving down the list, AIME—the American Invitational Mathematics Examination—2025 is now scoring 100% for GPT-5.2 versus 94% for GPT-5.1, suggestive of post-training. Then we get to the second set of benchmarks that I think are super interesting: ARC-AGI 1 and 2. ARC stands for autonomous research challenge, and AGI, of course, stands for AGI.

For those who don't pay super-close attention to ARC-AGI, ARC-AGI is a visual reasoning challenge testing problems that humans find relatively easy—a sort of visual problem-solving and program-synthesis challenge—but machines have historically found exceptionally, exceptionally difficult. It's an arbitrage between human minds and machine minds.

We see some big differences here. For ARC-AGI 1, the first version of the prize, we see that it's just saturating at this point: 72.8% with GPT-5.1 and 86.2% with GPT-5.2. ARC-AGI 1 is cooked at this point. ARC-AGI 2 is nearing the point of saturation, so there's a huge change from 17.6% with GPT-5.1 to 52.9% with GPT-5.2 Thinking. In my mind, this smacks of post-training. That's the obvious strategy.

It's cooked. It's automated, and GPT-5.2, probably again due to elaborate post-training, can get almost 71% of these tasks.

Dave Blundin That's 70%. What does that actually mean? Seventy-one percent of comparisons between a human performing this knowledge work and the machine, GPT-5.2, performing knowledge work resulted in the machine doing a better job. And that was, by the way, at more than 11 times the speed of the human and at less than 1% of the cost of the human professional. So knowledge work is cooked.

Okay.

I figured something out on that last line this week, too. I'm chairman of about a dozen companies, and I'm like, “Guys, what is holding you back? Why have you not deployed this? You can cut costs dramatically. You can automate. You can expand your market share.” And they're all like, “Yeah, I don't know. We're really struggling.” It's driving me nuts. What's going on?

A couple of things that I finally figured out: one of the companies is working entirely in Java. When you turn this loose in Python, where it had a lot more training data, it can build virtually anything. It just blows your mind, and it still really sucks at C. I don't think they're going to fix it because they just don't care. We've moved off of C anyway, and there's not enough training data. Java's somewhere right in the middle.

When they benchmark it, they're like, “Let me try and take my legacy thing and see if it can just immediately fix it.” And it struggles. But if you just say, “No, scrap it. Rebuild it entirely from scratch in Python,” you come back an hour later and it's done.

So they're stuck there. And the other place they're stuck is in operations.

They're saying, “Well, look, the way we pick up a customer service request is in an email that's in an Outlook folder, and that has all these security whatevers on top of it. So it's struggling to open and read the emails.” And so we're giving up. Don't you think you could maybe fix that front-end interface in a day and then try it on the rest of the process and just turn it loose? It would immediately crush the problem. So they're stuck on these little edge-case issues.

I'll tell you, it also comes up with that ARC-AGI benchmark, which was specifically designed to be things that a human finds relatively easy and intuitive, and the AI is still struggling with ARC-AGI 1. I've had countless conversations around academia with people who desperately want to say there's still something missing. There's something fundamentally missing in this great AI brain, and it hasn't been solved yet. The proof is ARC-AGI 1. And you're like, “Okay, boy, do you look foolish now,” just 2 or 3 weeks later, 5 weeks later, because it's basically saturated, but it's going to be completely saturated imminently.

And on the GDPval, if you remember, Elon has spoken about one of the companies he's going to be starting, which is Macrohard, and his mission is basically to go into a company, simulate all of your employees, and deliver it as a service back to that company. A lot of change is coming rapidly. I think the biggest challenge is that people are not projecting properly on how rapidly this is going to tip.

Our next slide here is the GPT-5.2 ARC-AGI update. We spoke about the numbers in the table just recently. Here we see it charted out, where GPT-5.2 has had a 390-fold efficiency improvement over o3, back from 2024. Anything you want to add to this, Alexander Wissner-Gross?

Yeah. We've spoken several times on the pod about, hypothetically, 40x year-over-year hyperdeflation. We're seeing 390x year-over-year hyperdeflation on visual reasoning for ARC-AGI. This is unprecedented, and this level of hyperdeflation in terms of the cost of intelligence will not stay contained to the data centers. It will not stay contained to these still relatively narrow benchmarks. I know they brand themselves as generally intelligent benchmarks, but they're still relatively narrow in the scheme of things.

It's not going to stay contained. Hyperdeflation is going to spread outward from these sorts of benchmarks to the rest of the economy. That's comment 1.

Comment 2, just focusing narrowly on ARC-AGI: one of the lovely things about the ARC-AGI 1 and 2 benchmarks is that they don't just focus on raw performance. They also focus on cost. If it costs us $100 trillion to solve a hard problem—if it's larger than the human economy to solve an important problem—then it almost doesn't matter. But if it's incredibly affordable, to your mantra, Peter, about abundance, if abundance is unaffordable, what's the point? It has to be affordable abundance.

The way we get there is exactly what the ARC-AGI organizers do, which is you measure on a scatter plot: performance on the vertical axis and cost per task on the horizontal axis. And that shows you what progress looks like. You want progress that looks like points in the scatter plot going up and to the left: greater performance at lower cost.

In fact, if, going back to my earlier comments, you see a frontier lab hypothetically just increasing compute costs but not actually making efficiency gains, that shows up in these plots, too. So you can see, for example, if you look at ARC-AGI 1—although it's probably a little bit difficult to read here, if you squint, you can see that GPT-5.2 is on sort of the same extrapolated slope as GPT-5 Mini, suggesting that maybe, at least as it pertains to ARC-AGI 1, there hasn't actually been major algorithmic or efficiency progress. It's just more compute being spent on the same tasks, and so it feels smarter, but it's actually because you're putting more work into it. As the aphorism goes, you're lifting with your back, not with your legs. But with ARC-AGI 2, there is, in fact, radical improvement. So we're seeing progress.

Well, this is a benchmark that I think a lot of people can relate to. The next one here is the GPT-5.2 writing benchmark comparison: long-form creative writing and emotional intelligence. Again, we're seeing improvements across the board. Alex, one more interpretation here.

Spiky. This is very spiky. We saw that interesting 3-dimensional plot on “When Are We Going to Reach AGI?” and again, spikiness was the descriptor for it.

That's right. That spider plot was purportedly comparing humans with AGIs, or strong models in general. What we're starting to see here is increased spikiness and spiky competition between the different frontier models.

Just a little bit of context: the long-form creative-writing benchmark evaluates a model's ability to basically write a novella—an 8,000-word novella—and is judged by Sonnet 5. The emotional-intelligence benchmark measures how well a language model can grade short fiction. So what we're seeing here is no single model dominating all the benchmarks. For example, with long-form creative writing, Anthropic's Claude Sonnet 4.5 wins and is the best at writing an 8,000-word novella.

What do you use? What do you guys use for writing? I've been using Gemini 3 Pro. It looks like Claude Sonnet 4.5 is the one to go to.

I've been using Gemini 3 Pro, and I found it to be really amazing to craft with, but I'm using mostly business documents. So that's a little different.

Same for me. I use Gemini 3 Pro for almost all of my writing.

Yeah, I'm using Kimi K2 for huge volumes of stuff on my little fleet of NVIDIA chips that I hijacked. And I'm using Gemini to de-spyware it and to proofread it. I'm using Claude Opus as well.

My Opus expenses went from $200 a month to $1,000 a month. I'll easily crack $20,000 or $30,000 this month, but I'll also generate more code this month than my entire life up to this date. So it's a bargain at $20,000, but my expenses are going through the roof on Anthropic, and I'm happy with it, actually.

Spyware? What's spyware? What does it mean?

Alex warned me that when you use a Chinese open-source model, it can inject evil things into the code that it returns to you.

This is actually public information. We're not breaking news here. Just to expand on this, there have been very well-publicized, outside-of-the-lab studies that found, for example, prompting certain open-weight models with certain politically sensitive topics for certain countries results in those models emitting more vulnerable code. That's something to be wary of.

More broadly, for creative writing, none of these models is so strong that I can ask them to do a good job doing all the writing. What I find inevitably is that I end up having to do 80% of the work, and models function as more of a junior editor, as it were. I still end up doing the majority of the writing. Similarly, to Dave's point with code generation, I would certainly not trust code-generation models not to insert vulnerable code.

When you told me that a week ago, I was like, “You know, Alex, I'm just going to see the code, and I'll see if it's injecting anything evil in there. I'm not super worried about it. Let's go.” So here we are a week later, and it's generating volumes that no human being could ever look at. I'm like, “Oh, I was completely wrong.”

And it worked. The code just flat-out works. I don't even have to look at it. It's passing every eval. It's building interfaces that I want. It's doing everything I want it to do without needing to look at it.

So now I've actually got GPT-5.2 proofreading right now, but I think what I need to do is just turn off Kimi and pay the 10x-higher price. It's actually a 20x-higher price to run it on GPT-5.2 instead.

I'm going to have to do that because I don't know how else to make sure I don't end up spyware-ing my entire world.

This is a real challenge.

If you basically have intelligence being dumped into the world, then there's this implicit trade-off: do you want intelligence cheap, or do you want it to be safe?

Yeah. We've talked about this as a potential strategy for China making open-source models available to the world. If it becomes the base on which you've built everything, then it's there from the beginning. I don't want to impute a dystopian point of view on all the Chinese model makers, but it is a concern.

I think we're going to see a move to sovereign intelligence. I think this is the long-term trajectory we find ourselves on. Every sovereign entity is going to want its own sovereign, trusted stack.

Well, how do you feel about France? Mistral's Devstral 2 raises the bar in open-source coding tools. What do you think about Mistral, Dave? Are you playing with it at all?

You know, it's funny. I saw this chart and had kind of forgotten all about it, and I guess my read on the chart was, “Oh, it exists.” But the headline says it raises the bar, and it's actually below—only by a notch, but below Kimi and DeepSeek. I guess you could probably trust it more because Europe is much more trustworthy. But other than that, it was like, “What's the news here?”

It's the headline: Europe—slow but trustworthy. Okay. And also, there's this sense, for a variety of reasons, that Mistral is somehow the EU's sovereign AI stack or sovereign AI model. But its roots are all very much American. All of its early funding is from blue-chip American VCs, and its founding team came from DeepMind and Meta.

It has raised a large amount of money from ASML most recently, and my understanding is that Europe is very interested in using Mistral as an AI emissary to the rest of the world. But its technical roots are deep, deep in the US. It's this bizarre world that we find ourselves in, where it's a Paris-based frontier lab or neo-lab—however they brand themselves—and right now, it's the only and main counterweight to Chinese open-weight models.

There's one thing I thought was really interesting here. Once you have open-source systems beating closed systems, you move innovation to the community level, from the lab. There's no catching up with it once you get that flywheel going. I thought this was a big deal. They may need a little bit more improvement, per Dave's point, but I think once they get there—

Is that true for AI open-source models? I know it's true for a multitude of fundamental, just plain software models. We've seen that before. Alex, do you think—

It's tricky. It's tricky because you have to ask what the primary limiting factors to increasing capabilities are, and it's compute more than talent. There's lots of talent in the world, but compute is still pretty scarce. The community has lots of talent, but in my mind—

They don't have compute. They're compute-starved. This isn't like Linux, where you can say, “Lots of eyeballs make all bugs shallow.” In this case, the way you make the bug shallow is by investing trillions in capex.

Well, this conversation is critically important. And Alex, you can help the world a lot, because every corporate executive in 2026 is going to need to choose something. There are only 2 types of executives out there: people who are familiar with this and have already got their landscape figured out, and the other 99% who are going to get slapped in the face in 2026, have to react, and are late to the party.

You saw the benchmark earlier. Everything every one of your employees can do can now be done by AI. What are you going to do—just sit there and ignore that? So, 2026 is the turning point. But these choices are really tough. On this chart, an executive might say, “God, I can go open source at 1/20th the price, but I get 72.2 arbitrary units of something, or 77.9. What does that mean?”

It means a lot. Anyone looking at the chart would say, “What's the big deal? It's only 5 units.” But the reality is that the capability difference in terms of your economic value is massively, massively bigger as this goes up even a little bit. So it's a tricky, tricky situation in 2026 for pretty much all of the corporate world.

I think it's probably going to take some sort of regulation to move the dial on this. Right now, if you hang out with all the Silicon Valley firms that are using open-weight models, they're all using Alibaba's Qwen at this point. Mistral and Devstral are great, but in the mind of a typical Silicon Valley firm that needs to host its own models, they're probably too little, too late. They're all using Qwen, and they're all fine-tuning Qwen. It's going to take an executive order, an act of Congress, or some sort of regulatory measure to turn off the cheap Chinese open-weight intelligence before they're incentivized to move over to Mistral, Devstral, or GPT-OSS.

But Dave, I think one of the points that you made is that the CEO and the board of directors of a company are in extremis, in paralysis, not knowing what to do—

Right. And their lunch is going to be eaten by the small startup that says, “Oh, there's an interesting business, so we should go and enter it.” It builds an AI-native approach at 1/100th the cost and 10x the innovation-evolution speed. So what do they do? Who do they turn to to help them reorganize their company?

It's a risky move, because if you brought in an outside consulting firm, I don't think it's going to be the big consultants. I mean, there are going to be AI-native companies out there. We're going to be having a podcast conversation with one company called Invisible that does this very shortly, and there are others.

The right way to do it—you said it earlier—is to scrap what you've been doing and actually start with a fresh stack. That is so hard for any company to do. See?

Yeah, this is right in our wheelhouse. Essentially, we're working with some very big companies, and Dave, you're exactly right. They're totally paralyzed. They're flailing. They have no idea what to do. If they bring in one of the traditional consulting firms, they just push them faster down the old path, and that doesn't work at all.

What needs to happen is that they need to take their capability here, create a new stack on the edge that's completely built AI-native from the ground up, and then little by little deprecate the old and move functionality, capability, and resources to the new. The political and emotional stress of that is causing most of them to do nothing.

Yeah.

And so, out of the, say, 20 major companies we're working with, maybe 3 are doing 50% of the right thing. Most of them are just saying, “We're going to keep pushing this old model and seeing where we get to. Surely we can catch up because we've always been able to get there before.” The answer is, you absolutely cannot.

Macy's. It's Blockbuster. And when you say “we,” you mean OpenExO is doing some work with these companies out there?

Yeah, we have about 42,000 people talking to companies around the world, so we're aggregating the information from all of that.

I think 2026 is going to see the biggest collapse of the corporate world in the history of business.

You've heard that first prediction here. No doubt, because I think this is going to be—and we should maybe have an end-of-year perspective and some predictions—

For all the madness we've seen in 2025, this is the slowest it's ever going to be. In 2026, it's going to be 10x to 50x to 100x crazier. I don't even know where to start. I've got benchmark fatigue right now.

To deal with all this, if you hire Salim to help you with your strategy, one of the things he'll tell you is to read Clay Christensen's The Innovator's Dilemma. It addresses this question exactly.

What that book will tell you to do, and Clay Christensen's framework will tell you to do, is go find Link Studio, Y Combinator, and Neo. Go out there and find your AI development partners. Try to do a deal with them where you either invest in them or become a development-partner customer for them.

Pull them in and give them revenue, because their market cap will go way up. They'll all become wealthy, but they'll then hire the talent. Point them at your internal problem and have them solve it inside your organization as an outside, very tightly bounded startup company that's growing like crazy.

That's the only way you're going to get the talent focused on your internal problems. You can't hire the talent directly anymore. You have billion-dollar signing bonuses all over the place.

And by the way, Salim will tell you to go read Exponential Organizations 2.0, too, which is our book.

How to do this. Yeah. I actually had a couple of really interesting conversations with Clay before he passed away. One thing he very honestly admitted was that The Innovator's Dilemma works really well for identifying the cracks in the structure, but it's not that great on the prescriptive side or at trying to predict.

For example, in his model, Uber is not very disruptive. I said, “Well, Uber?” He said— I said, “But Uber is very disruptive. It fits right into the wheelhouse of our ExO framework.”

And he goes, “Yeah, it means our model’s wrong.” When we drilled into it, what we realized was that The Innovator’s Dilemma assumes the verticals—transportation, energy, healthcare, and education—stay in those verticals. So Uber, as a transportation company, may disrupt a little bit of transportation, but without realizing it, it’s also disrupting healthcare delivery, restaurant delivery, and food delivery, and can go horizontal across a lot of these. The old verticals are essentially collapsing, like the old newspaper with its printed sections saying utilities and this and this and this. To Alex’s point, it’s all going to become 1 category called compute.

Well, if you don’t want to do what Salim is suggesting, the other choice is to do a $20 billion acqui-hire plus $14 billion in new payroll. That’s the other way to solve the problem. Or, I tell you, the other thing I’m seeing is unbelievable: executives at that level are looking at the world and going, “Yeah, I’m just going to retire right now.” There’s this unbelievable trend of opting out.

Stop opting out, exactly. It’s like falling off the cliff.

It’s the most fun time in human history. How can you not dive into the ground? I actually respect that. I’ll tell you why: what they’re doing is basically saying, “I can’t navigate this new world. I’m going to let the younger generation navigate this because I can’t do it.”

But it’s very honest, right? At least the worst thing in the world is the old fuddy-duddies who are running the world on the old model, who can’t—or won’t—get out of the way. We’re seeing that much more in politics, to some extent, and in the corporate world. There’s this massive change happening.

So, talk about billion-dollar salaries. Talk about The Innovator’s Dilemma. Our next story here is “Meta’s shifting AI strategy is causing internal confusion.” Meta is at an inflection point after mixed Llama 4 results and a reported $14 billion AI talent spending spree. Mark is considering whether an open-source strategy can still compete with closed, vertically integrated rivals like OpenAI and Google. Dave, what do you think about this?

I think they’re doing exactly the right thing. Actually, the other backstory here—which I guess is validated, although maybe it’s more rumor than validated—is that they’re getting heavily into distillation of other people’s models to accelerate inference-time speed. What’s exciting about that is, if you look at where we are in human history, intelligence in a box was invented just days ago—or really 2 years ago—but it’s brand-new in the world.

Now we’re in the hyper-experimentation phase of figuring out how to make it bigger and better: by running many agents in parallel, by expanding the context window and dumping in tons more data, and by iterating it over and over and over again with chain-of-thought reasoning. We’re getting ridiculous gains, but we’re brand-new in that game.

What Meta has realized is, “Look, we’re behind in the foundation-model race. We do need to rebuild and catch up, but that’s not going to happen overnight. Where we can potentially get ahead is through raw inference-time speed, having many more agents working on things in parallel.” I believe that will also lead to self-improvement, which will get them back on the map. I think they’re directing all their research energy now into making this blazing fast and becoming the world leader in distillation. That’s my view.

Incredible. I’m blown away by the $14 billion hiring spree. Just that number—I can’t process that number.

Well, remember, they’ve got a massive cash cow and cash-flow-generating engine. Mark has basically said, “This is the race. If we don’t spend the money now to get toward number 1, it will slowly, slowly go away.”

And what’s cooler than cool is that he’s already decided to use every single penny of it, plus debt on top of that, to try to win this race. Wall Street has said, “That’s fine. No damage to the stock. Go for it. We love what you’re saying.” That’s just a beautiful thing.

So what you’re saying is they’re moving from trying to focus on the open-source foundation model to putting all of their chips on the agent strategy.

Well, there’s so much innovation there, too.

Yeah, I think they’re in a bit of a tricky situation. I know the key players. Zuck’s undergraduate adviser before he dropped out was my postdoctoral adviser. Nat Friedman, who’s with Alexandr Wang helping to lead this new lab, was my first roommate at MIT. I’m pretty familiar with the key players in this particular story.

I think there are 3 strategies that Meta could be pursuing, and/or has been pursuing. One strategy is “commoditize your complement”—drive the cost of generative AI to zero. That was their Llama strategy. The problem is that Llama 4 was a disaster, and the Chinese open-weight models are flooding the market and doing a much better job.

The second strategy they could be pursuing is more conventional, and perhaps what Wall Street would expect out of Meta: use strong AI to improve Instagram and other Meta products. I would have to imagine many executives at Meta would like to see all of these new AI resources used to improve Meta’s existing products.

Strategy 3 is to compete directly with the frontier labs, with closed, source-available models and APIs, to be the first to superintelligence. What Meta has to struggle with—it’s almost, hopefully, not like a civil war internally—is which of those 3 strategies they really want to pursue. My guess is there are constituencies with different interests within Meta that want to pursue each of those 3.

I cannot believe Mark is not all-in on number 3. Being first to superintelligence just feels like Mark’s MMO.

Yeah.

And.

Yeah. And I think very often the cover story is, “Look, we’re going to enhance existing products. We’re going to use our internal data. We’ve got a huge amount of internal posts that we can use as training data.” That’s all kind of a cover story for the real goal: “We want to win the race to AGI and ASI.”

By the way, everybody, I want you to realize, as you’re hearing these stories about Google and Meta, that it’s all about business-model innovation on top of all of this. Google is going from an ad-based search company to an AI-based company that’s delivering a whole slew of different products. Meta is the same.

This is where companies fail. When Blockbuster did not change its business model, even though it had 2 opportunities to buy Netflix, how do you actually disrupt your own company and shift its business model? Otherwise, it’s game over.

The Innovator’s Dilemma, to Dave’s point earlier.

Yeah. But it is ironic. Sam Altman has said publicly that he’d much rather have 1 billion users with a non-frontier model than vice versa. And yet what we see from Meta is the exact opposite strategy. Meta already has 1 billion-plus users, but at this point it would much rather have a frontier model.

The grass is always greener at the other frontier lab. That’s funny. That’s a good phrase.

All right. Our next story here is “Google DeepMind to build materials-science lab after signing deal with the UK.” We’ve heard about this as well. Another company out of MIT and Harvard called Laya is doing something very similar. It’s all about the data: if you’ve consumed all the data, you need to go find new data.

Imagine having a lights-out robotic capability where the AI puts forward a scientific hypothesis, designs experiments, and then, at night, robots in the lab run the experiments to either confirm or modify the hypothesis. Let’s do that 1,000 or 10,000 times faster than humans can. I think we’re going to see multiple companies doing this. Every frontier lab is going to need this kind of data mining. We’re data-mining nature and trying to understand what’s going on. In particular, Google DeepMind is focusing on materials science, while Laya is looking at biological sciences. Thoughts on this, gentlemen?

I don’t know if there’s a Polymarket on this, but Demis is really leading the race to being the coolest guy on Earth. He got his Nobel Prize in Chemistry, and now he’s going to crack computer science. You could kind of see this coming, because AI can allow you to be a world-leading expert in anything. He’s the master of the biggest AI compute in the world, along with the algorithms and TPUs.

He also isn’t one of the corporate leaders trapped in the political fray.

Beautiful. We’re going to have the coolest-guy benchmark.

Well, what’s great is that you want somebody with that purity at the edge of this, which is fantastic. Having spent time in physics labs during my degree, I thought of a couple of things. If you have a fully autonomous lab, this is the biggest breakthrough in scientific progress since the scientific method was invented. We talked about dark kitchens and dark factories, and now we have dark labs. Holy crap.

Yeah, it’s funny, too. I can only find a handful of people like Demis and Alex on this podcast. There are 10 or 12 that I could name who can tell you the implications in all these other areas—in biotech, materials science, chemistry, and math.

You know, Alex is talking about solving all math. It’s just such a small group of people who see where this is going to take us and how short that timeline is. So, it’s good to see Demis doing materials science.

This is AI-assisted science and AI-native discovery. Alex, do you want to close us out on the subject?

This is what comes after superintelligence. What comes after superintelligence is solving math, drug discovery, engineering, and medicine. And yes, math is being solved. We’ve spoken about that perhaps ad nauseam at this point on the pod.

We haven’t spoken as much about AI solving all of materials science. There are a dozen companies. It’s not just Google. It’s not just Laya. It’s not just Periodic Labs. There are a dozen companies that are all laser-focused on solving materials science, and that’s going to give us so many upsides.

It’s also, when we talk about recursive self-improvement, having better semiconductors and better superconductors. Materials science is at the foundation upon which everything else is built.

The Singularity, here we come.

And the innermost loop accelerates again.

For our new listeners, if you hear Alex saying “drink,” there’s been a bingo game invented for terms that are repeated on a regular basis. You’ll be hearing it.

All right, let’s move on to our next story. I don’t know how I feel about this story. I sort of feel like I don’t want to overblow or overexpose what’s already been overblown, but this is a story about an AI-native character called Tilly Norwood.

Tilly Norwood is an AI-made actress created by a London studio to star in films and social media. Built over 6 months with GPT, Tilly went through 2,000 design versions, and her YouTube videos garnered over 700,000 views in October.

We saw this also in the music business, where fully AI-native bands and music tracks have been created, and people don’t even realize they’re listening to something that’s fully AI-generated.

She has her own agent.

Yeah.

And reportedly, she has around 40 different contracts for movies and other development projects. I would say this is consistent with my model hypothesis that over the next 10 years, we’re going to live out the plot of every science-fiction movie ever made.

In this case, I don’t know if you saw the movie *S1m0ne*. This was the plot of the science-fiction movie *S1m0ne*, where an AI actress develops a life of her own and takes over. It has Al Pacino in it. It’s a fun movie, but we’re going to see AI actors and actresses take over, potentially—or at least we’ll discover how soon humans crave authenticity in their entertainment.

There’s no doubt in my mind that humans do not crave authenticity as much as we think we do. We will just watch whatever is interesting and entertaining.

I was at the *Washington Post* when every reporter there was saying, “The Post will be fine because people will want genuine, great reporting from great reporters who are struggling in the field to find the stories.”

That was right before—yeah, guess again. Gone. Just gone, and in just a couple of years, too. The timeline was so much shorter than they ever would have thought. From the top newspaper in the world, a multigenerational family business for 3 generations, to gone. Jeff Bezos bought it for pennies on the dollar in just, what, 3 years? 4 years?

That’s going to happen here, too. No doubt in my mind it’s going to happen with music. It’s going to happen with movies. It’s going to—

Yeah, it’s inevitable.

This is an AI performer working 24/7, appearing in unlimited projects, never aging, never burning out, and never needing to renegotiate contracts. This is the Screen Actors Guild’s worst nightmare.

I had dinner a couple of nights ago with a dear friend on my XPRIZE board who used to be the head of 2 of the major studios, and with an actress who’s another dear friend. We were talking about this, and it is scaring the daylights out of the industry.

Well, no—good, because they’ll react. I wish nothing but good to happen to the people who are in the industry, but it’s good that they’re scared, because then they’ll react as opposed to getting crushed. I didn’t mean to—

Well, the question becomes, then, what’s the response? Are you, as an actor, going to license your persona? Because that’s the way you’re going to make money in the final result.

If you don’t, then the industry—or the next-generation industry—will simply create a Tilly Norwood who is actually cuter than you or more handsome than you.

Doesn’t age.

Doesn’t age.

Oh, yeah, there you go. Doesn’t age. That’s a huge one. I’ll tell you one thing.

I wonder when you’ll have one of these winning the Oscar.

Right? In theory, in theory, they should be the best.

We have a lot of those benchmarks. When will the first AI win a Nobel Prize? When will the first AI build a—

Geoff already did it, because he’s kind of half AI anyway.

That’s done.

It’s squishy. Also, there have been, by my count, at least 2 Nobel Prizes. There was Demis with AlphaFold in chemistry, and then there was also Geoff Hinton with restricted Boltzmann machines for physics.

The squishy thing here is that you can always do a secret cyborg, as some would say, and wrap AI talent inside a human meat body, and the human claims the credit. So, it’s unclear again how much humans crave authenticity.

Does this become a separate category in the Oscars, like animation? Is this an increment on top of animation that’s real-life animation, or is this an actual labor substitute? I don’t know yet.

I think a lot of that thinking is a little bit misguided, in that what the actors will be looking for is a feature-length movie in a theater where it’s all AI. That’s what they’re going to use as their bellwether for the threat.

But that’s not what’s going to happen. If you look at the data, short-form video is taking over movies anyway, and video games are already miles ahead of movies.

We’ve had these conversations. Kids don’t go to the movies. They watch YouTube videos. It’s all—

Exactly. So, Tilly will end up being a star in every video game and also every TikTok clip.

Across platforms.

And they’ll say, “Well, that’s not a threat. That’s not a threat to me. I’m a real actor. I do Shakespeare and whatever.”

Well, no, it is a threat to you, because the audience has moved and the budget has moved, and that will undercut you. They’re looking at the wrong bellwether. When Tilly shows up in 5 billion TikTok posts, that’s when you know you’re dead—long before it hits you in your long-form movies. You’ve got to look at the video games, too.

A related story is that OpenAI is working with Disney to bring Disney characters into Sora 2.

Right, so that—

Yeah, they just announced that.

Yeah, it’s fascinating. So—

A $1 billion investment and licensing agreement.

I think there’s going to be a certain fungibility between classic IP assets and generative everything. In the short to medium term, it’s reportedly a 3-year licensing agreement that OpenAI and Disney struck.

Maybe, in the short term, the remedy is that existing actors can license their visage as an asset to customers who want to do sort of fan pics. But really, if you’re a really popular star like Peter Diamandis, what’s the thing you should do right away?

Sign my rights already.

Yeah. Get your avatar out there. Get it built and out there right away. Get your Tilly Norwood equivalent, Peter—or whoever—out there right away, so that personality can grab attention before the true synthetics take over.

Yeah, it really is going to be a race for neurons, right? If you’re looking at the general public, Dunbar’s number only really cares about 150 people and holds them close. The question is, are 1 of those—or 10 of those—going to be synthetic actors? Once you get to a point of popularity, it’s going to be hard to replace you.

For what it’s worth, to tie a bow on this, too, Dunbar’s limit of 150 people was in the ancestral environment. If the number is valid at all in the post-social-media era, you can maintain light, casual associations with thousands of people.

But Dunbar’s number is basically the human tribe. I’ve done this when I was running Singularity University: it’s the number of people whose names you can actually remember, with whom you can go deep, and so forth. Sure, you can have a Rolodex of 22,000 people, but Dunbar’s number—in terms of who you feel closely connected to—is a real number.

I’m with Alex on this one. What I noticed was that once you have Facebook, you could essentially use Facebook as your RAM for Dunbar. You could move people in and out of that spectrum very easily without really noticing.

You have the opposite effect, too. Once you start to connect with enough people—Peter, you’ve probably had this. I remember walking down University Avenue in Palo Alto right after one of our executive one-week programs, and this guy stopped me and said, “Hey, Salim, nice to see you.”

I said, “Have we met?” He said, “We just spent the week in the classroom with you.” I thought, wow, our brains are too blown up now by the limits of that. We need technology to expand that capability.

It’s already done that to one extent, and we can move things in and out. The question is, what do we do when we have all these synthetic AI avatars going through that?

This episode is brought to you by Blitzy, autonomous software development with infinite code context. Blitzy uses thousands of specialized AI agents that think for hours to understand enterprise-scale code bases with millions of lines of code. Engineers start every development sprint with the Blitzy platform, bringing in their development requirements. The Blitzy platform provides a plan, then generates and pre-compiles code for each task. Blitzy delivers 80% or more of the development work autonomously while providing a guide for the final 20% of human development work required to complete the sprint. Enterprises are achieving a 5x engineering velocity increase when incorporating Blitzy as their pre-IDE development tool, pairing it with their coding co-pilot of choice to bring an AI-native SDLC into their organization. Ready to 5x your engineering velocity? Visit blitzy.com to schedule a demo and start building with Blitzy today. Our next story here comes out of the White House. Trump signed an executive order curbing state AI rules. This is a decisive federal power grab over AI regulations. Trump's one-rule executive order is going to preempt state-level AI laws. It's like, no, Washington, D.C., is going to win over everybody. It's not California laws or Texas laws; it's Washington, D.C.

Ultimately, I think this is what the EU needs as well. It needs top-level direction. It's going to be harder there. Any particular thoughts on the one rule here?

It's absolutely, positively necessary. I hate it when this happens, but we've got to do it, because variety across states is one of our best assets. On the other hand, New York just passed a law that says you can't use the likeness of somebody who's deceased in an AI without going to their ancestors. What about all these Einsteins floating around already? How are you going to keep it out of New York?

There's no way to launch it across the country and then somehow block New York users. It's just unworkable.

I'm going to claim I'm one of Aristotle's descendants, and you can't use his likeness. I mean, how far back?

I had 2 thoughts when I saw this. One was that when I saw “one rule,” I very quickly thought about “one ring to rule them all.” I just love the politics of this, where a huge amount of the effort for Trump was saying, “Let's push everything down to states' rights,” and now we're going totally in the opposite direction.

I think it's a necessary thing. I agree with Dave here. It has to be done, because if we don't get uniform AI treatment, where the hell are we going to get to?

I also think there's an interstate-commerce angle here. Models are being trained in one state and inferenced in other states. In my mind—and I read the executive order in the past 24 hours—the E.O. is ensuring a national policy framework for artificial intelligence.

I think this is both reasonable under the Interstate Commerce Clause and necessary for international competition. It's not at all obvious how a patchwork of state-based regulations results in anything other than total chaos.

This is a piece of the overall White House strategy on energy, on data centers, and on chips. It's all aligning everybody to make the U.S. as competitive as possible on the global stage and to accelerate as fast as possible. It is a race to superintelligence. This is just part of that.

Can I make a radical prediction here?

Yeah, of course.

Over the next 5 years, the entire U.S. Constitution will evaporate. Every clause is starting to just melt away. Look at the right to privacy—the Fourth Amendment—gone. We're going to see the whole thing. It needs to be rewritten from the ground up, and it's going to be interesting to see how that happens. I will move that.

Boom. That's what you need.

Instead of the Founding Fathers, it's the Founding Models.

For the record, I don't buy that prediction for 1 second.

Good. We can put some money on it.

That's Polymarket, baby. All right, let's move to a conversation on the economy. This is data just to support what we already know. OpenAI finds AI saves workers nearly an hour a day on average. Workers using OpenAI tools have saved between 40 and 60 minutes a day. The survey of 9,000 people in 100 companies found that 75% say AI makes work faster or better.

The biggest time saver: over 1 million businesses today are using OpenAI tools. I'm going to couple this story with our next one, which is layoffs announced. In 2025, we had 1.1 million layoffs, which is the most since the 2020 pandemic. Dave, you want to jump on this?

I was talking to Scott Perry, the CEO of LendingTree, a public company, yesterday, actually, and he said 20,000 incredibly talented people in Seattle are now cut loose from Microsoft and Amazon. It's the best hiring opportunity for tech talent he's ever seen in his life.

These are really, really solid, great people that the megatech companies have just cut out because AI is automating, improving, and enhancing. Coding is one of the biggest early beneficiaries, and my top coders are 10 times more productive, so I don't need nearly as many. That's where the layoffs are coming from.

We'll look back on this and say, “Wait, what? That was a bellwether. Why did I not notice this little thing?” When you see what happens in 2026, you'll say, “When did this all start?” Well, this is when it's starting.

Now, what do you predict for 2026, Dave?

Continued. The capabilities will be able to eliminate on the order of 80% to 90% of all jobs, but the rollout and the percolation are dependent on regulation and also corporate bureaucracy, so it's tough to predict how quickly people will react.

My guess is that it'll get a very slow start. Everybody's very stodgy, but then everyone's a sheep. When somebody in your industry is an early adopter and their stock goes up 10x just because they're an early adopter, then your board beats you up like crazy and says, “What about us?” The sheep effect flips in 2026.

By the end of 2026, everyone's in absolute panic mode, and then they're wishing they started at the beginning of 2026.

I think there's going to be an absolute need for all medium-sized and large companies to bring in a reskilling consultancy, some type of program. It could be fully AI-based, but it would provide some kind of safety net for your employees: you're going to reskill people before you fire them, and if they aren't able to be reskilled, then they're let go.

I also think that's a huge business opportunity for an entrepreneur out there to build that kind of capability.

Totally. Totally right. In fact, if we look in our portfolio, the companies that are, quote-unquote, forward deployed are killing it. If you couple that with what we just said, there are 20,000 highly talented people in Seattle who just got cut loose.

A lot of the younger companies—22- and 23-year-old leaders—are afraid to be forward deployed because they've never done it before. They don't have any management experience. They don't have any enterprise sales experience.

Well, hire those 20,000 people. Train them on how to be AI-forward-deployed consultants or delivery people, and then get them embedded back into corporate America at State Street Bank, at JPMorgan, at Walmart. They'll hire your people instantly to get AI deployed inside their organizations because they can't get that talent.

If you grab those people, retrain them very quickly on your own AI training platform, and then get them redeployed into corporate America, your growth rate—you'll be sold out every time you have a meeting. You'll generate a sale.

The founders, the really young founders, are afraid to do it. They want to just launch their software on Hacker News and hope that the world sucks it up, and there's just this big gap between there and where corporate America starts. It's never going to fill if you don't get forward deployed.

I don't think this is a skills issue. This is a cultural problem. The problem is in corporate America, with all the structural impediments in a big company, you need a mindset shift at scale in a company to even adopt this.

I think the large companies and the medium-sized companies, to be very specific about my prediction here, are going to need to hire a very specific kind of consultancy: a company that comes in, and its job inside your company—and I think every company's going to have a version of this—is reskilling.

When you go to work for a company, there's a reskilling safety net there for you. ExO, yeah. But what I'm saying is it's not just reskilling. It's a mindset shift. Real change. It's a cultural change that has to take place, and that's actually much harder. I want to say 2 things.

There's a cultural and mindset shift at the CEO, at the executive level, and at the employee level.

All of them.

It goes through the organization. We've actually been working on this for several years now, and I want to tell a quick story. Our second-ever client, when we finished one of our 10-week sprints, realized that they had to lay off 1,000 people in the company, and they decided, “What are we going to do?” because we're a family-owned business.

We really want to provide for these folks. What do we do? We actually got them to give them a 1-year UBI so that they could find their own passion and find their own work. If they didn’t by the end of the year, they would try to hire them back. It was an incredibly successful program, and I think we’re going to see a lot more of that as we transform the workforce.

All right, let’s get into data centers, chips, and energy. We’re seeing data centers begin to pop up in countries around the world. I don’t want to spend too much time on this, but QIA, the sovereign fund there, is investing $20 billion to launch a data center in Qatar as a Middle East hub. We’re seeing Microsoft and Satya, just back from India after meeting with Prime Minister Modi, committing $17.5 billion in India to expand an AI-ready cloud there in the region. This is going to be the case in all major nations: these partnerships taking place. This is Alex’s comment about tiling the world with data centers, and everyone drink.

Drink. Tile the Earth with sovereign inference-time compute. Drink, drink.

Okay, but we’re drinking coffee this morning, ladies and gentlemen.

Drinking water.

Alcohol. All right.

You know, in our last pod, we talked about China’s incredibly expanding role. China is set to limit access to NVIDIA’s H200 chips despite Trump’s export approval. President Trump says to NVIDIA, “Okay, you can export these,” and now the Chinese leadership is saying, “No, no, no, you can’t buy them. You need to buy Chinese-made GPUs.” Fascinating, right? It’s propping up its own chip economy. I think it’s a smart move on China’s behalf.

This is so fun and annoying at the same time to watch. This is pure protectionism. The U.S. never did it before, and now we’re playing the game. What happens is, a country invents something like an LCD TV or a car, and another country says, “Okay, what we’re going to do is protect the home market. We’re going to manufacture our own. Then we’re going to dump it on your market cheaply, and we’re going to dump it until your companies collapse and the venture capitalists all run away. Then we’re going to price it up.”

So we embargoed the chips from China, and they’re like, “Oh, we need to build our own whole supply chain.” As soon as they get it up and running, we’re going to say, “Oh, no, no, it’s okay. Now we’re actually going to allow you to buy the H200s,” and that entire thing they just built makes no economic sense. China is saying, “All right, I see what you’re doing here. I’ve played this game for a long time. We’re not going to buy them.” Why? It’s an incredible buy. Why would you not allow us to buy them? Because we’ve already made a massive investment in our own fabs. We’re going to have to keep subsidizing that to get this up and running because we know what you’re doing here. You’re going to let us buy them right up until our stuff collapses, and then you’re going to cut it off again.

This is—it’s a trust issue.

Big trust issue.

There’s no trust at all between the U.S. and China right now.

Well, this same thing happened, right? The Japanese came over during Trump’s first administration and spent a lot of time negotiating a trade deal, and then, just a few months ago, the Trump administration canceled that trade deal. The Japanese are like, “We’re not negotiating another one because we don’t know which way is up anymore.” Every single time, it changes completely. So there’s no trade deal, and this is really a big problem going forward. I think what China is saying is, “We don’t want to play that game.”

There’s no doubt that the outcome is two completely separate ecosystems. Europe is kind of a wild card, and so is India right now, but there’s no doubt the U.S. ecosystem is going to grow completely independent of the China ecosystem because there’s no chance of reestablishing trust after that chip embargo.

Yeah, there’s no way that’s going to get mended. That’s right. So sovereign data-center AI compute, to Alex’s point—

It’s a new—it’s almost like a second Cold War. It’s a world that we’ve moved to where there are spheres of influence, spheres of fabs, and spheres of compute. The decoupling happened.

Yeah.

Okay, move on to power generation. There’s a company called Boom. Many years ago, it set out to build the first supersonic passenger airliner to replace the Concorde. I was so impressed by the founder and CEO, his chutzpah, if you will, to take on this moonshot to build a supersonic consumer airplane. I was like, “I don’t know how you get there. How much money is going to be required to build this?”

It’s a fascinating backstop that Boom had been developing: supersonic engines. Now they’ve unveiled a Superpower turbine that’s able to provide 42 megawatts of natural-gas turbine capabilities to data centers. This is a backstop business model for Boom, and it’s huge. This is moving power to the data centers. It’s a gas-turbine strategy, and we’ve heard before that all the gas turbines have been sold out for some time. Alex, do you want to jump on this?

Yeah. As you were gesturing, Peter, the wait times right now for gas-fired turbines for AI data centers are 7 years in some cases. I think this is a brilliant strategic pivot by Boom. Also, to the extent that, referencing comments from a minute ago, we’re in almost a quasi-second Cold War, this is almost like a self-directed Defense Production Act-type move, pivoting resources perhaps from turbines for supersonic consumer jets to turbines for AI data centers.

Of course, there are synergies there, but this is, I think, a brilliant pivot. The irony is that there’s probably a much, much larger addressable market for gas turbines for AI data centers than there is for consumer supersonic jets at this point. I just hope, for the sake of Boom, that they retain at least some semblance of the original supersonic vision and don’t get overwhelmed by the AI data-center business.

I just love that audio clip. Hey, behind the scenes, I need that audio clip right away.

There are so many companies, including Vestmark, one of the ones I founded pre-AI. It manages $2 trillion of assets, 20 million lines of code, and it’s a profitable, great business. I’m like, “Guys, you’ve got to be an AI company by tomorrow.”

Pivot, pivot. Pivot, pivot, pivot, pivot.

This is a great case study. You wouldn’t think that a jet-engine company is going to culturally pivot and become a power-generation company, but when you look under the covers, it’s like, “What are our assets here?” We’ve got the blades, the manufacturing, and the metal. That’s all it takes. The age of AI has so much opportunity that didn’t exist the day before. You don’t have to be that close to the center point; you have to be adjacent and pivot quickly, and you’ll succeed wildly.

I hope these guys just crush it. In fact, I know they’ll crush it because, like you said, Alex, I know data-center operators personally that will spend anything. They’re pre-buying, too. They’ll pay you up front for something that you’re going to make next year—a billion-dollar backlog. It’s a product they can deliver immediately, right? This is on-premise power generation for data centers, which is so critical.

Boom’s been working on this for, I don’t know, 6, 7, 8 years. They’ve built the scale model of their supersonic airplane, and they’re trying to get advance orders from all of the airlines. But getting through the FAA thicket is so difficult. It takes a decade—that’s decades.

It will kill you. But if you’ve got an actual business model delivering revenue right now, I agree with you, Alex. I hope Boom actually delivers on their original idea. I think this increases the probability a huge amount. This is the equivalent of Amazon realizing, with Amazon Web Services, that it’s got something it can offer to everybody else that makes very strong near-term profits.

Elon—like delivering Starlink now and a Mars colony in 10 years.

Yeah, that’s the sexiest-looking gas turbine I’ve ever seen, by the way. Beautiful-looking thing.

I’m sure after you run it, it gets dirtier. It has a $1.25 billion backlog. Congratulations to the team at Boom for that strategic pivot. Everybody else, learn from this story. We should track this in a few weeks or a few months.

What do you have? What are you building right now that’s a cost center for you that could become a profit center for you in the AI ecosystem? That’s the question.

On the energy side, China builds nuclear reactors at $2 per watt versus the U.S. at $15 per watt. Again, what’s going on here? Why is that happening? Alex, do you have a thought?

Yeah. Well, China does have more people than the U.S. China does have a need for more energy. If AI were not part of this equation and China were to attain U.S. per-capita energy-footprint standards, China would need more energy than the U.S. in a total sense—in an absolute sense.

That part makes sense. What doesn’t make sense is, if you look at the permitting processes required for nuclear energy in the U.S., it’s a very different beast. Obviously, the NRC regulates U.S. nuclear power deployments at the national scale, but on top of that, you have some states that de facto ban nuclear power entirely. We have a patchwork of state and local regulations that make it extremely difficult to deploy nuclear energy.

Here in Cambridge, Massachusetts, many people may or may not be aware of this, but Cambridge has a nuclear reactor. It’s not very well advertised. It’s on Massachusetts Avenue, on the MIT campus, but we have a working nuclear reactor and have had one since, I think, the late ’60s or early ’70s. That’s very much not par for the course in the U.S. I wouldn’t be surprised if, sometime in the next 2 to 3 years, we see some equivalent for nuclear energy of what we just saw with the White House’s executive order.

We could see it in the next few months. I mean, the bottleneck is not physics; it’s permitting and execution, and that’s got to be cleared.

Yeah.

I’ll give you a little side story related to this. The MIT brand is absolutely skyrocketing in this AI revolution. But we found out that the MIT nuclear reactor is going to be exothermic and powering the campus, and I’m like, wow. We don’t have a single nuclear reactor in the state, you know; we can’t get that approved. We buy our nuclear power from New Hampshire, but MIT can actually get stuff like that done now. It’s just crazy how that brand has skyrocketed in impact with this AI revolution.

All right, I want to jump into robotics. A special hat tip here to Salim. This is Salim’s perfect robot. It’s got something like 14 different arms on it. Salim, are you happy with this robot?

This looks awesome. Look at all the chickens that can move around very quickly. I love it. Just love it.

For those of you new to the pod, Salim is having a running debate about why humanoid robots and why just 2 arms. Well, Salim, you’ve got all the arms you could possibly put on a body here.

I just love all the wires sticking out of it.

I mean, there is a serious story here too. In China, there’s an image doing the rounds with 6 arms.

I can’t wait for that.

Yeah, doing the rounds with 6 arms. I don’t think there’s anything super-efficient about it.

Yeah, I was going to bring that article forward as well.

Yeah, there are now 6-armed robots coming out of China.

China is not about having a humanoid robot. It’s about mimicry. It’s about integrating into human spaces and moving around where humans have been, so there’s some case for it. But in general, it’s very easy to be 10 times more efficient than a human being. We’re very, very inefficient in most of the things that we do.

I think evolution has, over billions of years—or maybe on the order of a billion years—done a search through body space. There are lots of body shapes that aren’t anthropomorphic humanoid bodies: more arms, more legs, more heads, lots of different formats. I do suspect we’ll see lots of different body shapes tested—a Cambrian explosion of different body shapes.

All right, listeners, dream or nightmare? It’s just the most effective use case for trying to get something done.

Call out to our listeners. I made that on Nano Banana. Somebody make, now that we know about the woolly mouse, Salim’s perfect robot for turning the woolly mouse hair into sweaters for us, and then send it to us. We’ll put it on the next pod.

Okay, that’s a hell of a prompt. All right, another form of robots is drones. I just found this Antigravity drone. That’s the name of this drone. It’s manufactured by a company called Insta360 in Shenzhen. For those of you who don’t know, Shenzhen is really the entrepreneurial hotbed in China. I’ve visited many times. You can go there, and every part and component you need is there to be manufactured.

Check out this video of an 8K, 360-degree drone. Talk about marketing genius. The drone user is using it with VR goggles, and he’s on a platform suspended by a hot-air balloon at 5,000 feet altitude. The drone is just flying a beautiful 360-degree view of him.

The dude standing on a platform suspended by a hot-air balloon—that’s way more interesting than the drone. That’s ridiculous.

Well, it’s like, what are you going to do to capture someone’s eyeballs—their attention, right?

You know, I think Salim is onto something here. Drones are a commodity, but the experience of being on a hot-air balloon at altitude in a VR headset, controlling a 3D drone—that’s got to be some sort of consumer experience that one could build an enormous business out of. Maybe that’s more interesting than the drone itself.

Yeah. All right, let’s move on to our next story in robotics. This is robotically automated vertical farms, which are an important part of our future food chain. Of course, this is out of China once again. What we’re going to see here are these massive vertical farms operating 24/7, basically growing at the perfect light frequency, with the perfect soil, drip irrigation, and pH. The AI is checking to see if something is ripe and ready for harvesting, and the robot arms are harvesting. This is going to be happening 24/7 in a city near you.

This is one of the futures: stem-cell-grown meats and vertical farming that helps us bring food to individuals. I don’t know if you realize this, guys, but half the cost of a meal that you have is food miles—transporting the food, whether it’s Argentinian beef or Chilean red wine.

The average meal in the U.S. travels 2,400 miles to get to your table.

Yeah. This is something really incredible. We’ve been tracking this for a while. We’ve crossed over into economic efficiency for farming, agriculture, and food production. The calculation I’ve seen that’s the most startling is that if you took 35 skyscrapers in Manhattan and turned them into vertical farms, that would feed the entire city sustainably.

So you think about food security, logistics, trucking—all of that stuff—and when you can automate the entire farm, the yield is something like 7 to 9 times what you can get with horizontal farming because you can give exactly the right frequency of light that you can dial in, by the way. You save 99% of fresh water, and 70% of our fresh water goes to agriculture, so you don’t need a lot.

And no pesticides, no fertilizer—all of this stuff. The benefits are kind of incredible, so we’re going to see vertical farms next to every restaurant over time, just feeding the restaurant. This is amazing stuff.

Yeah.

It’s probably also quickly worth pointing out that video, to my knowledge, was actually put out by the Chinese government. This is a new form of soft power, soft influence—broadcasting these visions, presumably ground-truth accurate but presumably also of radical forms of automation. I think we’re going to see many forms of propaganda and soft influence, as these amazing tech demonstrations of robotics in action start to hit the internet.

And by the way, a humanoid robot makes no sense in that factory.

Agreed. But a humanoid robot does make sense in this next story, again out of China. China is testing retail automation with humanoid robots running the shops.

So what do we have here? You’re walking by, you look inside, and you don’t see humans; you see a robot behind the table, behind the desk, and you want to go in and check it out. This is the rise of the robot-run convenience store, taking humans out of the loop.

We’ve seen Amazon do a version of this with Amazon Go, where you walk into the shop and pick up anything off the shelf. There are cameras noticing what you took and what you put back on the shelf, and then you’re automatically rung up as you walk out. But here we’ve got a 2-armed, 2-legged humanoid robot doing the store clerking.

I do think that this is going to be viewed as sort of the autonomous vacuum-cleaner moment of 2025. Do you really need a humanoid robot in a convenience store? No. Probably there’s a more ergonomic solution, like, as you say, Peter, Amazon’s Just Walk Out technology. On the other hand, I would love to live in a world where every convenience store is filled with humanoid robots in the U.S. doing this as well.

I think it’s fun. I’m sure we’ll see this this year, as soon as 1X with its NEO Gamma or Figure gets involved. We’ll be visiting Figure at the end of January to record our next podcast with Brett Adcock. I just spoke to him yesterday.

I’m super excited about going and seeing behind the scenes there.

Two counter-predictions. One is, I think this takes at least 5 years to have a convenience-store operator with a humanoid robot. And by the time those 5 years arrive, we won’t need convenience stores anymore for various other reasons.

Ah, interesting. Everything is being conveniently taken to you by a drone.

Drone delivered.

Yeah.

You know, with Brett Adcock, maybe he'll let us go behind the scenes for real, like into the factory. With 1X, there was too much proprietary stuff, so they wouldn't let us do it. If they had cleaned up a little bit, maybe we could have done it. But it's incredible when you go back and see the actual robot construction. God, I hope we can get footage.

We went back and saw it, but we couldn't bring the cameras back there. That's what you were saying.

Yeah. Yeah. Too many secrets.

Another story here, back in the U.S.: Boston Dynamics announces its plan to ship automotive volumes of humanoids. This is from their product lead. I actually interviewed the CEO at FII. We're owned by Hyundai for a reason: we can ship automotive volumes of humanoids.

There are 1 billion cars out there right now, and they're being manufactured at tens of millions per year. Imagine that. We've talked about this: Elon plans to do this, Brett Adcock plans to do this, and we've heard this from Brent Borick. Now we're hearing this from Atlas, right? The ability to manufacture robots at millions and tens of millions per year—robots building robots.

We don't need billions of cars. We do need billions of humanoids. Yeah. Two-armed humanoids, Sem[?]. Two-armed humanoids.

Okay, well—

Don't get arrested.

I'm staying silent on this one.

Here's a story that's fun. Years ago, I had the pleasure of meeting an extraordinary entrepreneur, Eric Migicovsky, who built the Pebble watch. He did this on a crowdfunding platform. Remind me which one it was. It was Kickstarter. Yeah, he built it when he was running out of money.

Yeah. He was running out of money, and he had about 3 months of cash in the bank. He was able to get funding for his Pebble watch.

He went on Kickstarter and said, “Hey, if you want one of these watches, fund me.” He went from 1 problem— not having enough money—to another problem. I forget how many orders he had.

I'll tell you. Eric's a fellow Waterloo grad. He was running out of money, as you say. Even after going through Y Combinator, he talked to 20+ investors in Silicon Valley, and nobody would fund it because hardware was kind of a bad word back then. He put it up on Kickstarter, trying to raise $100,000 to build a prototype of his watch, and got $10 million worth of orders.

That's right.

It's an important point because it tells you 2 or 3 things. First, the investor is wrong. Fine. Secondly, if you can do this, why do you need the investor at all? But the third thing, which I think is the most powerful and one of the big inflection points—we talk a lot about this in Exponential Organizations—is that now that you can do this type of Kickstarter campaign, you can actually get market validation for a product before you build it.

We've never been able to do that before in consumer hardware or consumer products. This is an amazing inflection point. Sony is actually launching anonymous Kickstarter campaigns and then funding the winners because its product development has not been the greatest over the last couple of decades. They're tapping into this modality, which is really powerful.

So Eric goes from 1 problem—not having money—to another problem, which is that he's got to deliver on $10 million worth of orders. He literally takes the first plane out of the U.S. to Shenzhen and basically builds the manufacturing chain in China to deliver this. It was a great watch. I remember having one. I gave it out at Abundance 360 years ago—about a decade ago—but then the Apple Watch came out and sort of crushed the marketplace.

Well, Eric's come back, and he's got something called—

Pivoting to AI.

Yeah, the Pebble smart ring. For $75, you wear a ring that has 1 purpose. It's got a small physical button on it, and when you press the button, a microphone records whatever you want.

You remember waking up in the middle of the night and remembering something? You just push your ring and whisper into it. Or you're meeting with somebody, and you walk away from the meeting and say, “Okay, I need to call XYZ as soon as this is over.” It's reminders and notes that go into your AI model.

It has 1 purpose. It's not tracking your heart rate or your sleep. It's tracking the bits that dribble out of your thoughts during the course of a day.

I love it. Critically, where does the voice go? The voice goes from the ring to an on-device, locally hosted large language model on your phone that then transcribes and analyzes it.

What is this really doing? To the extent that a ring stays on you almost all the time, this is about adding a button to the human body that enables you to speak to a foundation model that's also on your body.

So, a question to the Moonshots mates here: How long until it's not just a button on your body that enables you to talk to a foundation model, but you're swallowing foundation models? How long until the first edible foundation model? Or injectable or subdermal?

Do you think it'll be injectable versus edible first?

Well, if it's edible, it's going to pass through your alimentary canal all the way out the other end.

I want this. There's an interesting part of the skull—the mastoid bone, in the back behind your ear. That's this hollow area of bone. I think it's a great place to implant a permanent microphone and speaker.

That's my prediction. We're going to be implanting a microphone and speaker at the back of your head.

That exact thing was on Shark Tank, and Mark Cuban vomited.

Really?

You can iterate hardware much faster outside the body than inside the body. I don't think it'll be invasive for a while. I think we'll see swallowable foundation models in the next 2 years.

Bluetooth, like just Bluetooth in and out of your body to your phone.

Bluetooth, but critically, locally hosted. Very locally hosted.

Okay. All right. A few subjects—a few topics—on space here. Let's move us along, guys.

Chile becomes the first Latin American country to enable Starlink Direct to Cell. Starlink is such a killer app for SpaceX, and the ability for Elon to potentially bypass the current phone industry—tens and hundreds of billions of dollars have been invested in 4G and 5G-level distribution networks, now to be bypassed by Starlink—is crazy.

Can I just go back to that for a second, Peter? I think this is a very big deal because throughout history, this is the failure of government. The U.N. should have launched something like Starlink. They should have, but they're fundamentally unable to, and it needs the private sector to do this type of stuff.

What I find incredible is that the demonetization and dematerialization of technology now allows a private individual to do something like this that changes the world completely in such a powerful way. You can say, “Well, governments should just step out of the way and let the private sector do everything going forward,” because it will navigate most of this with light regulation. We can navigate most of this stuff now, so I'm really, really excited by this.

Okay, can I ask you guys a question? I was trying to look at the data behind this. The idea of orbital data centers wasn't in the conversation how long ago? I mean, we weren't talking about this a year ago. We weren't talking about it 9 months ago.

NASA published a paper on this about 14 years ago. If you were reading Accelerando, you had the blueprint for everything we're seeing now.

Sure, but it wasn't.

No, but in March a year ago, one of your Abundance 360 guys was talking about it. He was going to do Bitcoin mining in space at that point, and everybody thought he was insane. We also thought we couldn't do the cooling. That was only March a year ago, so that's 9 months.

So I know at that point it was nothing.

Yeah, but the last 6 months—really, the last 4 months—all of a sudden, every single player has appeared. We've got companies out of China; we saw them on the last podcast. We now have a company out of Europe, and we have a dozen companies in the U.S.

Then I found this video clip, which I found fascinating, because Google wasn't discussing it a few months ago. But here we are. Listen to Sundar.

Yeah.

How do we one day have data centers in space so that we can better harness the energy from the sun? You know, that is 100 trillion times more energy than what we produce on Earth today.

We want to put these data centers in space, closer to the sun, and I think we are taking our first step in 2027. We'll send tiny racks of machines and have them in satellites, test them out, and then start scaling from there.

There's no doubt to me that a decade or so from now, we'll be viewing it as a more normal way to build data centers.

I never thought I'd hear Sundar say “tiny racks of machines.” That's hilarious to me.

I just love the schoolboy-level excitement he's got there.

You can see him actually grinning. He's like, “Oh, data centers in space. This is amazing.”

I love the “AI-generated” label. The big banner on top of that video is AI-generated. It's like, “We're always going to tell you that this scene in deep space is AI-generated,” as if it were not.

The reason, Peter, why—even though I may be a little glib saying, “If you had read Accelerando, this would have been obvious to you almost 30 years ago”—is that this is a sudden phase change in the way the industry works. Google's plans are public information. The Google plan to launch these is TPU-based, first of all. Google's launching TPU-based data centers, obviously, on Planet satellites—Planet Labs. It's not Google's own satellites.

So if Google's hitching a ride via SpaceX on Planet satellites, this is all happening all of a sudden. I'll say that second point: sun-synchronous orbit is about to become very, very crowded. Sun-synchronous orbit is a low Earth orbit for satellites that want to always have sun exposure, never pass behind the Earth, never be in the shadow, and always have solar power for their panels. It's going to be very crowded.

It's real estate. It's a limitation. Currently, there are limits on how close you can get to other satellites. That's going to be a real challenge because we've got a dozen companies all wanting to do this at the same time.

It's going to be a race, and how the FAA, which governs this, is going to decide who gets the territory and who doesn't. In geostationary orbit, there's a very clear demarcation: “I own these orbital slots over my country.” But low Earth orbit doesn't have that situation.

Peter, you're making the case for the Dyson swarm. Again, the Dyson swarm. We move out of GEO, we move out of LEO, and Sundar himself in this clip was saying, “We want to get closer to the sun.” So we're sleepwalking straight into the Dyson swarm.

Peter, to your prior point, too, this was science fiction a year ago, and now suddenly it's mainstream among the top CEOs in the country. How does that happen? You look at Elon and his credibility. You look at Alex, your credibility. A lot of things that were impossible a year ago are going to be very easy a year from today. If your track record of predicting them is near perfect, then the credibility of these crazy-sounding ideas immediately catches on.

You're going to see a lot more of that, I think, because the capabilities are exponentially growing. But some of these things are truly harebrained, and some of them actually are not.

Is there a line of sight on solving the heat-dissipation problem for these satellite data centers?

Yeah, and radiate in the direction of the cosmic microwave background.

The final answer shocked me, but for every square meter of solar panel, it only takes the same 1 square meter of radiative cooling, which really surprised me. I thought it would be 10 times more area. We estimated it on Gemini, which was wrong. It's cooling at 1×, and I don't know how they got that. It's all aluminum-based, so it's not weird, expensive metals or anything like that.

So, yeah, point it into deep space, like Alex has been saying forever, and for whatever reason, it's just flat-out working.

I took all of the comments from our last 2 pods and ran them through one of the LLMs and said, “Okay, pull out the most interesting AMA questions.” Here we see a list of 10 of them, gentlemen. Let's pick out a few to answer.

I'll start with one: How do you make these space-based AI data centers fault-tolerant? There are sunspots, and there is the potential for disruption from an EMP at some point—God forbid. Any ideas on making them fault-tolerant?

Those are 2 very different faults.

Yeah, both are disruptive.

There are lots of different failure modes, so I do think this is another multibillion-dollar company that someone should start. There are many techniques right now, ranging from switching from silicon-based electronics to maybe other semiconductors, like gallium arsenide, or 2D or 3D semiconductors that are more fault-tolerant and have different band gaps.

You can also design electronics that are intrinsically, at the design level, better able to tolerate faults. Or you can do what right now is a standard protocol: if there's a solar storm or bad space weather, you shut down or switch them to safety mode.

There are lots of partial solutions here. To my knowledge, there isn't a definitive industry-standard solution for what happens if you're in the middle of a training run.

I just hate to think about the idea of all your data centers in orbit shutting down because there's a solar storm for the next 12 hours. We're getting hit by alpha particles.

But how do we solve that in general? If there's bad weather or a blackout on Earth, you have diversification. So, again, let's put space-based AI data centers throughout the solar system. If there's bad space weather in one part, there isn't in another.

That's a great point, actually. I bet earthquakes, tsunamis, and hurricanes are a much bigger problem than solar storms.

All right, let's pick another one of these.

Hey, just to make a point, though, there's a flaw in the question, too. When you have Skylab up there, you want it to be up there for 20 or 30 years, and you don't want it to get hit and destroyed or anything. But these space-based data centers need to be replaced every 3 years with new chips.

And so they're not like that. It's a constant launch, recycle, launch, recycle, launch, recycle thing.

Somebody EMPs the entire thing and destroys it, then there's a war, of course. But it was going to get replaced in a 3-year cycle anyway. It's not like a satellite.

Interesting. One of the things we did for Planetary Resources when we were looking at asteroid mining, we set up the software so we would expect constant disruption. We focused on rapid restart of the system so it would boot up extraordinarily fast.

Can I tell a quick story here?

You can, but I want you to choose one of these AMA questions also.

Sure. You and I were sitting in a hotel in Dubai, and Richard Branson walked by and said, “Hello.” We grabbed a quick drink, and he said, “Peter, how's my investment in Planetary Resources going?” You described how it was going—it had NASA contracts, et cetera—and Richard turned to me and said, “This is why Peter's interesting, because in a random hotel lobby, I'm suddenly having a conversation about asteroid mining off-planet just like this. This conversation happens nowhere else in the world except with Peter.”

We love you so much.

It was fun.

All right, Sal. Pick a question here. Is this question bingo?

Should we expect G20-level initiatives for UBI within the decade? I would hope it would be within a year. It needs to happen very, very fast. I think it'll force the conversation.

Universal basic—right, “universal” may soon be replaced by UBS, universal basic services. But I think you shouldn't expect much from the G20, period. I think that's the flaw in the question. In general, we're going to expect to see this rolling out in a pretty rapid way, with lots and lots of experiments being done all over the world, because they have to. We have to move to something like that. The social contract is completely being wiped out in the current model.

Dave, why don't you pick a question next?

Okay, I'll take number 1. How can AI lift up those who aren't natural entrepreneurs?

I think, first, listen to the podcast, get subscriptions, play with the tools, and then brand yourself as an AI expert within your company. If you're not going to be an entrepreneur, that's fine. The demand for this knowledge inside the regular corporate world is going to go through the roof in 2026.

If everybody around you knows you're the AI person, and also, don't be intimidated. Historically, if you wanted to be a software god, you needed to be very, very software-oriented. That's not true with AI. It's much more intuition-based. You can build virtually anything with voice prompts. Just knowing how it applies in your industry will separate you. So just jump in the game.

Yep. Amazing. Alex, do you have one?

I'll take question number 4 for $10 trillion.

Is pure scaling enough, or what comes after? I think the answer is—it's a trick question. By “pure scaling,” I'll construe the question to mean that we freeze all algorithms. No new algorithms are allowed to be developed in AI, but we're allowed to shovel more and more compute, especially inference-time compute, into the existing algorithms.

I do strongly suspect that if we froze all the algorithms we have today—no new architectures—but got lots more compute coming online, the existing architectures combined with scaled compute would be enough to give us AI smart enough to tell us what a perfect algorithm would be, to the point where we get our highly coveted AI-researcher recursive self-improvement—the final algorithm—and we can just ask our scaled algorithms what comes after.

In summary, my answer to question number 4 is yes. I think pure scaling is probably sufficient.

Is it all that we need? No. Of course, algorithmic development in the real world is continuing, and we're going to get both. But could we live with pure scaling at this point? My guess is probably yes.

All right, let's answer one more here. Number 3: How do the Moonshot Mates prepare day to day for each podcast episode? I think we can share that. Alex, you're constantly providing the team with an incredible list of all the breakthrough stories you're searching. You're probably generating—how many AI stories per day do you think you generate for us to look at?

An order of magnitude: 20 important stories per day. I'm also, at this point, spending so much time just reading primary sources, arXiv papers, and so on, living in the zeitgeist of the moment because, after all, the singularity comes around only approximately 1 time per planet. So it's a special time.

I should probably also say that I'm turning all of these stories, in addition to research for this show, into a quasi-daily newsletter.

It's a genre I'm trying to popularize. I'm calling it sci-nonfi. It's written in a style inspired by Charlie Stross, Stanisław Lem, and others—science fiction in style, except it's all grounded in what's actually happening.

So, Alex generates on the order of 150 stories a week. I'll generate probably 20 or 30 stories a week. We get some from Salim, some from Dave. All this gets put into different categories. We then cut it down to the top 30 stories.

I typically spend about 10 hours playing slide shuffle, working with Gian Luca and Dana, who are incredible members of our team. Then we do research on those stories to get the details and think about them, and I'm probably spending a good 15 hours of my week focused on this. How about you, Dave and Salim?

Well, everything you just said—I lean entirely on Alex's internal feed. I've been reading that internally for, what, a year now, I guess, or more. It's very time-consuming, but I need to know it all.

The only other thing I do is route all the really big stuff over to the venture capital team and say, “What are the business implications of this?” which we need to know anyway to run our venture fund. Then I try to bring those stories back into the Moonshots feed so that we can talk about not just the technology, but what it means to investors, to businesspeople, to people with career planning, and all that.

I source a few stories, but nowhere near as much as the rest of you. I think I spend a chunk of time—the minute you guys release the deck, I look through it and then find it's changed again, so I have to restart. I'm always playing catch-up with your slides, and then, Peter, on the last night, you go, “God knows what you do,” but you change it all again, and I have to re-research it.

I spend 6 hours a week looking up the terms in the papers that Alex surfaces because half of it is Greek. I'll also ask my community—my OpenExO community. So there's a hive-mind reaction to some of this, which I think is very powerful, similar to Dave asking his team.

It's just sucking up more and more time per week, but it's such an important thing.

It's the most important thing we do. Come on. It's super fun. But what no one ever warned you of, it seems, is the singularity of covering the singularity. It's a singularity of time suck.

It's just a black hole. It's a black hole. A Dyson swarm forming around my own head.

The singularity wants your attention. So we hope for all our subscribers and listeners that you appreciate that we put a huge amount of work into this because we care about it deeply. I'm doing my Meaning of Life session next week. We've already almost sold out. It's going to be pretty amazing, starting at 11:00 Wednesday. Come armed with any question you have about life and judge me by how well this framework answers that question. All right, let's get to our outro music here from David Drinkall. Thank you, David, for producing this for us. Amazing. Have an amazing weekend. Take care, folks. 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, that 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. To subscribe for free, go to dmmandis.com/tatrends to gain access to the trends 10 years before anyone else. All right, now back to this episode.