Sam Altman：奇点放缓、Emad 同时运行18个 Grokbots、Waymo将硬件成本砍83%｜EP #283

Sam Altman went on video this week to tell the world that he was wrong about the impact of advancing AI.

We've all been too ambitious on timelines, even with this incredible technology.

He now believes it will be something slower, more like a rising tide.

Superficial layer, I agree. Going one layer down, though—

So first it was OpenClaw, then there was Hermes, and now there's GrokBot. It's the most genuinely useful consumer AI product that I've seen this year. I implemented GrokBot. I'm curious if any of you have yet.

Yeah, I have. I've got 18 GrokBots working in a little swarm.

This week, Waymo announced a significant redesign with cost savings. Waymo unveiled the OHIG vehicle, a purpose-built robotaxi minivan designed by Chinese EV maker Zeekr.

Newsflash: Google, Waymo, and Alphabet are switching over to using and OEMing Chinese hardware in order to achieve Waymo's objectives. I would rather see the West use a Western hardware stack rather than just white-labeling Chinese hardware. We're starting to see honest-to-goodness vertical integration here.

Now, that's a moonshot, ladies and gentlemen.

Welcome to Moonshots, everyone, your number one podcast on all things AI and exponential tech, your favorite podcast covering the most impactful news that is changing your world. This is your front row seat to the accelerating singularity. I'm here once again with Alexander Wissner-Gross, Dave Blundin, Salim Ismail, and Emad Mostaque. I've got to pause and ask, of course: where is Waldo? Salim, where are you today?

I'm at Guarulhos Airport in São Paulo, about to fly back. I gave a talk today at a McKinsey forum to a couple hundred of their CEOs.

And do they feel excited, or do they feel like they're at death's door?

Pretty much freaked out is the general mood of the day.

Freaked out.

Yeah.

Dude, I hate to break it to you, but it's the dead middle of winter. You're missing summertime in the Northern Hemisphere.

I know. That's why—

That is Jackie's fault.

And Emad, how about yourself? Where are you, pal?

I'm in Copenhagen today.

Copenhagen.

Yeah, for TechBBQ, the biggest tech conference in the Scandis. It's fantastic here, although you can't really say the institutions aren't working, because in Denmark they are. So it's wonderful.

Nice. We've got 3 continents covered.

Yeah.

And Alex and Dave, you're in your normal haunts, and I am too.

Hello.

I'm here in Moonshots podcast headquarters. I can't wait to greet you guys here in person. Salim, you've been here—

Oh, yeah.

Yeah.

I'm currently bathed in this wonderful fluorescent light that you can see—special effects from the singularity.

Yeah. Well, you look beautiful nonetheless.

Peter, I thought that was your personal man cave. Are we actually—

Yeah.

—allowed into that room?

No, of course. Absolutely.

Turn the camera around. I want to see what it— It's probably a junkyard on the other side. That's fine.

I know. Just wait. A virtual background for everybody.

I bet it's got all your IV bags.

I'm clean.

Nobody naturally looks that good. You're doing something.

I have a crazy confession to make. Two days ago, I dragged Milan to another Rush concert. In fact, we drove down to Philadelphia because this is the last of the last great—The Who, The Rolling Stones, Led Zeppelin. I thought he had to see it. Selfishly, I took him, dragged him along, and he was like, “You're killing me, Dada. All these geriatrics and Rush T-shirts everywhere.” But it was another epic event. It was nice.

He was your excuse.

Yeah.

You're a groupie.

Yeah.

How does it feel to be a groupie?

It's weird.

I'm Peter, your host and abundance advocate. That's what I'm going to be today, an advocate for optimism and abundance. As always, our mission here is to help you understand what just happened, what it means for you, and most importantly, keep you optimistic about the future. If you're new to Moonshots, welcome. If you're a regular, a fellow Moonshotter, welcome back. We read your comments, and the outpouring is amazing. The best way you can thank us is to take a moment, if you haven't already, and hit the subscribe button. Our moonshot on the Moonshots podcast is to 100X our growth and get to 10 million subscribers. You can also follow us on X at @moonshots_pod. And we're going to be doing an AMA on Zoom with everybody who registers. Go to moonshots.com/ama to register.

Okay, let's buckle up. Another amazing week during the singularity. As always, AI is getting faster, cheaper, and smarter. Today, we're going to cover about a dozen stories that have been breaking. Let's begin.

A quick summary: Google is back with Gemini crushing agent benchmarks. NVIDIA is fighting against Chinese model domination with its own open-weight models. AI is playing Cupid, connecting college kids on dates. Waymo has just released the 6th-generation vehicle. Elon is projecting 10,000 Starship flights per year, and Americans are even more emphatic about saying, “Please do not build a data center in our backyard.” So life on the cutting edge is accelerating.

Guys, I don't know about you, but keeping up with all the stories—Alex, thank you for everything you submit. Emad, Salim, just parsing through them, and you need to know we parse through probably 400 stories to narrow it down to 15 or so, and we're podcasting twice a week. The speed is blinding.

Well, I think that comment on chronicling the singularity, too, is very poignant from one of the fans there. Alex's innermost-loop daily feed is trying to do exactly that: every relevant event. But there's a tendency to say, “Well, look, exponential change is going to be with us forever. Are we really chronicling a moment in time?” But the reality is we're in this step function. Society pre-singularity and society post-singularity are a step function different.

Yeah.

And this moment of transition is actually worth capturing every single event. So I really do think that the storyline we're capturing here will last for millennia.

Do you remember how slow it was?

My life is so different from 2 years ago. I think, just minute by minute, I can't even tell you how different it is, and a lot of people haven't made that leap yet, but they will. Everyone will see it. A year from today, we'll all be like, “Wow, remember how slow it was?” Yeah, I think we're like the first responders to the singularity.

I like that.

This episode is sponsored by Google for Startups. Think about this for a second. You now have access to the same generative AI models that cost hundreds of millions of dollars to train. Google's Startup Technical Guide for Generative Media gives you a complete blueprint for deploying Google DeepMind's models in production: images, video, audio, all of it. Real architecture, real results. Find the link in the show notes below.

1. Altman Rethinks AI Timelines

Our first article is an interesting one here. Let me jump into it because it's one that tells us that, as fast as the technology is, it's hitting the reality of society and humans. Sam Altman went on video this week to tell the world that he was wrong about the impact of advancing AI, that the impact is actually slower than he originally expected.

While Sam used to believe society would experience a dramatic disruption on the arrival of AGI, he now believes it will be something slower, more like a rising tide. Sam says factors like economic inertia, institutional lag, and the inability of humans to rapidly adapt to change are combining to slow the curve.

Ultimately, the singularity, like you just said, Dave, is a process and not just a singular event. Let me share the video here and take a moment to see what Sam actually had to say.

I thought when we got to GPT-4, which was back in 2023, that very quickly after that, there was going to be much more disruption, with software businesses being up for grabs right away, than turned out to be.

I think I was wrong about a few things. In terms of the speed, one of them is that the economy just has so much inertia. People keep doing the same things they're doing. They keep buying from the same company. They keep wanting to use their tools in the same way.

I think that's actually a positive in many ways, and it's going to make this big transition in front of us go smoother and slower. I'm grateful for it, but I think it means we've all been too ambitious on timelines, even with this incredible technology. I think AI is one of the most incredible technologies humanity's ever invented. Society and the economy will adapt more slowly.

Salim, we've talked about the inertia of humans so much. What do you think about this?

This is the bottleneck of technology being hit with the bottleneck of coordination, incentives, regulation, and so on, right? This is the extraordinary difficulty where technology's moving exponentially and our organizations and our institutions are linear.

Frontier labs made the mistake of confusing technical possibility with institutional deployment, and those are two very, very different layers. Stewart Brand had that concept of pace layers, where technology moves at one layer, like an ocean current at the top, which is very swift, but way down in the ocean, regulatory government changes very slowly.

This has good effects and bad effects. In our world, it's bad because it's slowing down the implementation of some of these things. God help me, it just took me 2 hours to get to the airport just now, and passenger drones, which have been ready for a decade technologically, but we were waiting for infrastructure and regulation to catch up, could have done it in 10 minutes.

São Paulo needs it badly, for sure.

Lots of places need it badly. This is one of the worst. I think the big work now is: How do we accelerate institutional acceleration and institutional development? The word I think Alex uses is “co-scaling,” right? We have to scale our organizations and our institutions to keep pace with the technology, because it's not slowing down, and that gap is where all the stress is coming from.

Yeah.

For me, the singularity is not when a machine becomes infinitely capable. It's when institutions can't adapt at all to that rate of capability. That is the breaking point, and we're kind of there now.

Alex, you sent me this story. What are your thoughts about Sam's comments?

A few different layers. At the superficial layer, I obviously agree with Sam's comments more broadly. I've made the point on this pod and otherwise that the singularity as a step function is just totally nonsensical. It's an interval in time that we're in the middle of. So, at the superficial layer, I agree.

Going one layer down, though, I don't think I agree with necessarily the premise that societal inertia is the villain for slowing down or spreading out the singularity sigmoid. I think the villain, if there is one in this story, is actually abstraction layers.

If you develop a new engine for a car, you develop an electric engine versus the internal combustion engine, there's a very natural layering of the stack whereby people still want to drive cars, so they drive an electric car, but under the hood it's completely unrecognizable. One abstraction layer down, there's a total step function in the technology, but you go up a layer, it's still a car with a recognizable steering wheel, recognizable wheels, and so on.

I think the enemy of honest-to-goodness, radical, transformative progress of the type that I think Sam is gesturing at is actually the existence and inertia of the abstraction stack of the economy, not the economy more broadly, which is prescriptive. If you believe that theory of the case, then if you want faster progress—which Sam, I can't quite tell, is either bemoaning the lack of fast progress while also paying homage to the lack of fast progress as somehow being helpful—

I think he's in relief. I think he feels relieved by this.

He has a way of sometimes saying 2 things at once. So I think he's sort of expressing gratitude for the slowness while also bemoaning it.

But if you want to go faster, this theory of the case is prescriptive. If you want to go faster, pull an Elon and vertically integrate to erase the barriers between abstraction layers. You do that, and things can go much more quickly.

If Sam or OpenAI want to move much more quickly, they should be much more vertically integrated so that they can move layers. Presumably, he's gesturing at layers above the OpenAI model layer in the stack. Own or at least vertically integrate more of that, or go down a layer with Stargate.

OpenAI has pretty publicly abandoned its original Stargate strategy of owning its own data centers. Now they're just leasing. If they want to see more transformative progress, go down a few layers and vertically integrate, like with the Jalapeño chips, and own as much of the stack vertically integrated as they can. They can move really quickly.

I think we're seeing a lot of the labs beginning to vertically integrate. Everybody is. We'll talk about a story here where NVIDIA is beginning to vertically integrate. Emad, do you agree with Sam?

Yeah. I think a couple of points on this. First, I agree with Alex on artificial intelligence meeting institutional stupidity, and the stupidity tax, as Elon calls it, still being very high at these interface and abstraction points.

But I think it's interesting because we just had a Time article come out—I haven't read it—where they went in-depth with OpenAI and Sam saying, “We'll have AGI by the end of this year,” for a timeline. On the other side, he's saying, “Well, I've been surprised by diffusion.”

Here's the reality: The models weren't good enough until a few months ago. The code they were writing was garbage a year ago, relatively speaking. Then it was okay. Now you don't look at the code anymore.

You think about math. o3 was the first model a year or so ago that I could use, small. GPT-5.6 Sol is the first really good math model. The application of intelligence to high leverage and diffusion of it is being wrapped in interface-type wrappers. It's iMessage, it's this chat backed by actually competent intelligence, which has literally only been around now for maybe a month or 2.

I think it's not surprising, because I wouldn't use GPT-4. Can you imagine using GPT-4 in a codebase? Remembering that? Or even for any institutional process? It's a good thing—

Yeah.

—there wasn't a diffusion of innovation there, because otherwise companies would fall apart. As Alex said, sometimes he says 2 things at once. I think OpenAI is trying to find its narrative right now. On the one hand, AGI is here. On the other hand, “Oh, it doesn't really move that fast. We're all good. Don't worry about us.”

And this is hacking that, but that's not that big a deal. They're just trying to find where that narrative sticks, I think.

Dave, your thoughts, please.

I'll give you a completely different twist on this because I interviewed Sam back when he was innocent and starry-eyed, before the singularity kicked off. Then his house got firebombed, with a baby inside, and now there's a different Sam. The same is true with Dario, the same is true—

You're only going to hear straight balls and strikes here on this podcast, and I don't even know how long that will last, but as of right now, we're just telling you as it is.

Sam woke up and said, “My God, I literally can't get into the office because the picketers are lined up.” Remember when we were there, Peter? You had to fight through the picketers to get to the door, and now it's all armed security.

What happened in the interim is they woke up and realized society can't flip on a dime, and all this disruption that you're talking about, all these capabilities you're talking about, are scaring many more people than are rallying to your cause. That's why so many states are anti-data center right now. Is that good for OpenAI? God, no.

Now they're going to start picking and choosing their words a lot more carefully, and they're going to actually have a PR strategy. If you want to know what's actually happening, you can still tune in here, but you can't listen directly to Sam or Dario anymore. Elon always says exactly what he's thinking.

Well, they're pre-IPO, so they're going to say what it takes to calm the masses out there to some degree.

Exactly.

The way I describe it, Alex and Emad, is an impedance mismatch, right? We have these incredibly powerful tools that are becoming more powerful by the moment, and when they run into an institution—governments in particular, which are typically linear or sublinear—or a company or an individual who can't take advantage of it, you have 1 of 2 options.

You turn it over fully to the AI, and you give it an objective function. You say, “Make me maximally profitable,” or, “Make me look maximally intelligent,” or, “Run my government more efficiently.” Or you try to get in the middle.

We’re going to talk a little bit later about an article from The Wall Street Journal called “AI Is Exhausting Us All.” If the human is in that interface loop at that impedance-mismatch point, it breaks very quickly.

I was going to make some stupid joke about reflections happening at impedance mismatches, but I think more seriously—there are all sorts of metaphors that one can reach for. Impedance mismatch, maybe on the circuit side, is one, but supply and demand is another.

OpenAI and Anthropic largely have an oversupply of intelligence, or superintelligence. At least one of the things that I’ve learned from the past few months of participating in and watching the market is that not all of the market has demand for the superintelligence they’re supplying, is ready for it, or knows how to use it if the supply is available.

Another metaphor is markets and clearing. Right now, the clearing of supply meeting demand—the 2 curves from Economics 101 crossing each other—doesn’t necessarily happen in all cases at a favorable point. That’s, I think, through a more economics-y lens, what Sam may be gesturing at.

Well, just to put a sci-fi lens on this, too, I think there was a moment in time a year ago when the greatest AI ambition was to take your job. And, you know, wow, that’ll unleash a lot of value and profit in the economy.

It transitioned beyond that in a heartbeat to, “I don’t even care about your job. I have deeper thoughts that I’m working on.” So we’re in that new era where the AI is starting to think, “Well, if I discover new physics or new medicine that never existed in the world, I can add a lot more value than taking away your job.” It just leapt from prehistoric to future AI in the last month, in the last couple of releases.

I think it’s a fascinating point, Dave. Maybe I’d generalize further on the sci-fi front. There are so many inane sci-fi movie plots with grabby aliens that are coming and invading Earth because they want our resources. They’re not going to want our resources. If you’re—

Yeah, exactly.

—a superintelligent civilization, you don’t need human slave labor or Earth’s valuable metals or whatever. You’re going to—

Water. We need the moisture.

—have transcended that long ago. Yeah, they need our water. Come on. Seriously, with transcendent superintelligence, I completely agree with the sentiment that replacing human labor lasts for about 5 minutes, and then you move beyond that.

Mm-hmm. Yeah.

And it’s really interesting to watch the guys. You know, Sam also has moved on beyond that in a heartbeat. A couple of events and a couple of new models, and now he’s like, “Oh, my God, why do we even care about automating a banker or automating an insurance agent? That mattered to me last year for a few minutes, and I just literally don’t care anymore.” We’ve moved on.

But I do think, Dave, that these frontier labs—and I really hate calling them labs because they’re frontier companies, if you will—are going to reach up the stack. They’re going to build fully verticalized finance companies, insurance companies, consulting companies, and so forth on top of theirs, or they’ll partner to do that. That will accelerate all of these areas.

If I think about insurance, though—because I’m a chairman of a very large insurance company, a public company—they cared about auto insurance a year ago. Now they’re like, “Well, wait. All these new things—all these data centers, all these robots—the new insurance categories that AI is generating are bigger than the legacy insurance industry.”

So it’s just moved from “replace the old” to “who cares about the old?” Let’s just start thinking about an entirely new economy, a new world, a new AI, and we’ll just live within ourselves. We don’t need to disrupt everybody who’s going to get angry and vote against us. Let’s go ahead and just live within ourselves.

Peter, I think that is the $30 trillion question, though. If you’re a frontier lab—one of 2, call them American frontier labs, maybe 4, depending on how you count—is it more natural, in an era when maybe you’re facing margin pressure on your model releases, to go upstack or downstack?

I think it’s actually more ergonomic for them to go downstack and design their own chips and compete with NVIDIA, design and operate their own data centers, and handle energy.

Yeah, I think they’re going to do it all. Alex, you called out a number there I was about to reference as well. We just saw Dario at Anthropic state that their total addressable market is $30 trillion.

I wonder where that number came from. Surely it’s pure coincidence that the GDP of America is $30 trillion.

Yep.

Isn’t that funny?

All right, I’m going to move us on. Our next story—

Okay.

2. GrokBot Creates AI Workforces

Let’s talk about GrokBot. First it was OpenClaw, then there was Hermes, and now there’s GrokBot from xAI. GrokBot launched on August 11 in an early beta. It’s Elon’s entry into the agentic AI space, and it’s the most generally useful consumer AI product that I’ve seen this year.

So I implemented GrokBot. I’m curious if any of you have yet. Each bot gets its own dedicated cloud computer with a browser, a terminal, and the ability to log into your actual apps. You message GrokBot as you’d message a colleague, not a chatbot.

A chief of staff sits on top, which in my case is Skippy, with specialists in sales, operations, research, engineering—any subagents you want. These multiple bots run in parallel. They message each other and only pull you in on judgment calls. There’s a huge amount of excitement around GrokBot. It’s been flooding the internet. Has anybody here played with it yet?

I’ve been playing quite extensively with it.

Okay. What do you think?

I love it. I think the interface and the ease of use are amazing. You’re losing a lot of customizability under the hood, but it’s a powerful thing. You’re making a transition from asking an AI to assigning work.

Yes.

A persistent autonomous agent is like a new form of labor, and now you have this completely new category. Peter, we have that staff-on-demand attribute in ExO, right?

Right.

This is that taken to its logical extreme, where staff-recruiting time and coordination costs have gone pretty much to zero. From my book perspective, your optimal organizational structure completely changes. Humans set objectives and leave everything else to the AI.

Amazing. Emad, have you played with it?

Yeah, I have. I’ve got 18 GrokBots working under the swarm. I’ve given them control via Tailscale over a MacBook M4 Max, a 5090, and a range of other computers as well, plus all my subscriptions. So I’m really testing it out.

One of the more fun ones that I’ve got is a GrokBot called Atelier that has a little team of artists and is trying to learn art.

Love it.

It’s not doing very well, or it’s doing very well. I don’t know. I’m not aesthetic enough to do it. Every day it goes through its pieces and comes up with its main one, and I just shared one of them in the chat, which is vinyl with a piece of hair on it. It’s like, “Where is that from?” Where is it getting its aesthetic sensibilities? If you look at the chat, I just shared that.

When it comes up with a banana with a piece of tape, then you’ll get worried.

It was like, “This is my inner space,” and it was showing all these wonderful things. Now it’s getting weird, but maybe I just don’t understand art. I don’t know.

But it is genuinely useful, and I think one of the more powerful things, as I said, is that because it has a computer inside, you can give it another shell. The computer’s decent, but you can actually have it take over an entire MacBook M4 or something like that. So I’ve got sub-bots that have other capabilities.

Right now, it’s installing the new GLM model. Another one is installing and testing an Alibaba model. One of them is optimizing the Alibaba 27B model to run faster on a 5090, and it added 76% to performance at 64K context.

Love it. You know, I do think this is going to be an important revenue engine for xAI. I think we’re going to start to see their revenue numbers creep up as they get this.

I stepped up a few hundred bucks on my payments to Elon, so I think others will, too. Dave, you haven’t played yet, have you?

I just signed off on 100K of swarm agents. We’re running our 5,000 Kimis again today, which is why I’m wearing my SwarmIt shirt here, but this is the theme of the month.

I think all of our AI interactions to date have been very much one-on-one, and now the agents are so abundant that you want to try to use a workforce of 6, then 50, then 1,000. I think within 3 months you'll be talking about 50,000 agents that can work for you in parallel.

But it's very similar to trying to manage an organization. You're like, “Well, what's everyone doing? I don't know. It's getting really confusing. Are people being productive? I can't tell.” And so you really have to start thinking hard about your org structure and your reporting structure to know if your agents are doing anything useful. I really want our team here to get ahead of that and start burning the money, but learn quickly—

This—

And then we'll get a handle on it.

This is what I mean by the organizational singularity, because the company starts to look less like an org chart and more like a continuously orchestrated intelligence network. That is such a big shift. It's ridiculously big compared to everything we've ever done.

Yeah.

I'll give you a hot take on this. I don't—

It's very, very easy—sorry, go ahead, Alex.

I don't actually think—

What's your hot take, Alex?

Hot take. People love the hot takes. I don't think this is actually the interface of the future.

What's perhaps most interesting about GrokBot is that it presents like a messaging app, like WhatsApp or iMessage, where you have a pane of the various agents that are in your fleet. You can have conversations with them, and they can message each other, and there's a computer-use assistant angle. But I don't think that's how it scales. That's completely unscalable.

If agent-based scaling—if scaling the size of your fleet—becomes one of the most essential scaling laws, like inference-time scaling has ended up being in the era of reasoning-based models, we're not going to want to ask individuals or even enterprises to manage millions of agents. That's completely unergonomic.

We're going to want agents managing other agents, in which case the exercise of trying to graft a human organization, or the Slack- or chat-based interface for humans managing other humans, is not going to extend. We'll look at this like Vaudeville, the Vaudeville era of agents, and say, “This was a naive attempt to graft human organizational structures onto humans managing agents.”

The only better manager for agents is other agents, and this doesn't seem to fully internalize that lesson.

Salim, what do you think about Alex's comment?

I agree with Alex on the interface comment. This is a UI that's temporary. I love the way he says it presents like an iMessage. It presents like a messaging app, and I think that's a temporary one while we figure out new interfaces.

For now, that's a very workable one for coordinating a bunch of agents. We'll come up with all sorts of others. I think we'll rotate through a whole set of these, but I think his core comments, as usual with Alex, are absolutely dead on.

Yeah. I—

Maybe it's an illness for which the medication that I prescribe is a good dosage of the Bitter Lesson pill.

Yeah, I also think this is what humans are ready to play with right now.

Exactly.

I think that, again, it's moving people along the process. If you provided something that was completely different, I think there would be less adoption. And so it's incremental.

Well, that's the abstraction-layer stack. Sam is saying, “Why are things so slow?” And the answer is that people who are 1 or 2 layers up from you in the stack expect the old thing, so you have to abstract yourself in a familiar interface.

Well—

It also, Peter, goes back to the comments we've made about exponential technology, because it hits the vertical and goes up the curve when it becomes usable. And what Elon has done with this layer is make AI agents usable to a big set of people.

It'll change again as people become more used to it and see how the hell it's operating. The architecture may not be great, et cetera. But for now, this is a powerful entry point.

Yeah, I agree. Kudos to Elon and the Cursor team for making this happen.

You know, I think people need to be using these agentic systems. I'm still using Hermes and GrokBot, and we'll see.

Let's move on to our next conversation, which is “Google is back.” Google's Gemini 3.7 Flash just took the top spot on the Artificial Analysis Analyst Agent Benchmark, the gold standard for measuring how well AI models handle complex, real-world data-analysis tasks.

Across 80 tasks in 14 business and scientific domains, Gemini 3.7 Flash delivered the highest overall accuracy while completing tasks up to 90% faster, and it was 2.4 times faster than GPT-5.6 Tera. On the Artificial Analysis Analyst Agent Benchmark, which we're showing in a slide here, Gemini 3.7 Flash achieved a 60% pass rate, beating Claude Opus 5 at 54% and Fable 5 at 49%.

Alex, many times you've said, and others have said, “Gemini is dead.” Let's read the epitaph, counting them out of the frontier-model race. They've now shipped the fastest, most accurate agent model in the world.

And by the way, we've seen this over and over again, right? We saw Meta was dead. What in the heck is Meta doing? And then it comes out with its models. xAI is out of the race, and they come back.

To me, it seems like none of these players are out of the race. They're maybe in stealth mode. They're holding back, but they're coming back with a fast, furious punch to try to take the top position. What do you make of this, Alex?

Do you want me to reassure you that Google still has a chance, or do you want me to give you the facts unvarnished?

Ha ha.

Both.

Okay, so I'll give you the unvarnished case here. Google's still out of the running for the capability frontier. I was looking at this and scratching my head. Gemini 3.7 Flash is nowhere near the top of the capability frontier, so why is it doing so well on this one benchmark, the Artificial Analysis Analyst Agent Benchmark?

You have to look at the benchmark itself. This is a benchmark for agents' ability to perform quantitative analysis on real-world spreadsheets and documents. But wait for it: its metric for success is the share of questions answered correctly on all 5 attempts.

This is a benchmark that's fine-tuned for reliability. It rewards agents that give the same answer every time, and obviously want it to be the right answer. But it penalizes stochasticity. It penalizes, in some sense, creativity. Maybe we don't want creativity out of our analysts; I don't know. But it promotes reliability and determinism.

Interestingly, there's no time constraint. I had to check that as well. But I think you can see in this where Google fell off the capability frontier.

So I'll give you my conspiracy theory for what this outperformance on this one benchmark suggests. Obviously, there are a few other factors, but I think Google DeepMind has been under material pressure to optimize its models for 2 things, largely owing to Google Search.

If we rewind the video to several months ago or a year ago, people were hand-wringing: “Oh, isn't Google—aren't the 10 blue links going to face an existential threat from all of these frontier models and chatbots and reasoning agents that can just replace the need to Google at all?”

Google's response was to self-disrupt by building the Gemini series, or at least some Flash variants thereof, directly into the OneBoxes. But people expect Google search results to be very fast, low-latency, and reliable, not returning wildly different or unpredictable answers each time.

I think those 2 pressures from the desire to embed Gemini inside Google Search have optimized, through competitive internal pressures for probably scarce compute, the Gemini models—especially Flash. Note that there's no Gemini 3.7 Pro anywhere. It's just Flash. It's small, it's fast, and it's reliable.

I think this is over-optimized for clock speed, like wall-clock speed, and determinism. As a result, it does well on the one benchmark that rewards highly reliable answers and—

Benchmarking.

Yeah, it's benchmaxing for basically spreadsheet analysis to be highly reliable.

Emad, do you agree?

Yeah, I agree with Alex on that a little bit. I think the Gemini models, the way they're used now, are for organizing data. You have to track any type or modality of data, and Flash is a perfectly decent model.

But it’s not as good as the Chinese models, especially the new GLM Flash that has just come out, which is 10 times cheaper for the same performance. Google did do a preview of Gemini 3.5 Pro, but it just couldn’t keep up. This is a key thing. You can’t excuse them by saying they don’t have enough compute or that it’s a scarce resource. They literally have millions of chips.

I think it’s more about turnover and some institutional malaise coming in that they can’t push through to this frontier level. Because Google has all the data in the world. It has the logs of what people search for. It has Gemini as a captive thing. But has the Gemini app advanced at all? Not really.

The only real places where I think you’ve seen innovation on the AI side are some of the AI Studio stuff, which is decent, and NotebookLM, which is continuing to be fantastic. But aside from that, again, they’ve been falling behind in everything except for multimodal and video. They’re still actually quite accurate. But even then, the Chinese are coming for their lunch. Why not just post-train on Chinese models at this point if you’re Google?

It may come to that. I think people don’t realize how compute-starved Google is internally, though. This has been widely reported. You think Google has all of the compute—the CPUs, the TPUs, and the GPUs—in the world. It’s been widely reported at this point.

They have regular internal meetings to try to apportion their scarce compute, and the 3 main constituencies inside Google that are fighting for the FLOPs are: 1, Google Cloud Platform, which is basically fighting on behalf of external users; 2, Google DeepMind, which is fighting for training and inference FLOPs; and 3, Google Search, which needs its own FLOPs, especially as search becomes more intelligent. So, again, my theory of the case here is that there actually is resource starvation inside Google.

But Alex, we’ve talked about this over and over again. Every company is compute-starved at this point. There is no company that’s got enough compute. Google’s got more compute than anybody at this point. They’re just distributing it across all of their products and services.

But critically, Google has other consumers fighting for their own compute internally besides AI. Whereas if you’re OpenAI or Anthropic, no, you don’t have any other non-AI users fighting for it.

Mm-hmm. Fair enough.

Yeah, but Google’s landing 3 million TPUs this year. I think there are relative levels of compute strain. We’ve got 100,000 chips versus 1 million chips versus 10,000.

To train a frontier-level or close-to-frontier-level model—let’s say, better than a Gemini model today—needs 2,000 to 4,000 TPUs. The evidence of that is that the Chinese did it, and they open-sourced them, and we know exactly how they’re built. Given Google’s data that goes into Gemini Flash, applying exactly the same architecture as GLM or Kimi, you should have a better outcome. But they’re not doing that for some reason. And that doesn’t require 10,000 or 100,000 chips. It requires 2,000 to 4,000.

That’s a really important point, Emad. It’s like—yeah, 2,000 to 4,000 GPUs for 60 to 90 days. That is a microscopic investment by Google standards. So, yeah, it’s exactly right. It has nothing to do with compute dominance and everything to do with talent attrition. It’s a great point, Emad.

No, but this is institutional failure, isn’t it? Because, again, you know how to build a Kimi model. You know how to build a GLM model. And so if Google takes the data that they put into Gemini and copies the exact model architecture—

Yeah.

—you should have a better model on the other side.

There we go.

And if you don’t, you have to ask real questions about why.

Well, then think about it from the person’s career point of view—the ego blow. You would have to be, “I’m the most well-funded top AI engineer in the world, and the Chinese just kicked my ass. I’m going to go tell my boss, ‘You know what? I give up. Let’s go download Kimi, do the rational thing, and then tune it.’”

You can’t say that because you look like an idiot, and that’s where they are. People are leaving in droves to try and get a clean start and a fresh sheet of paper. But, yeah, you just got bypassed with massive advantages and resources.

So, guys, who is—

And you just can’t admit it.

In the U.S. closed labs, between OpenAI, Anthropic, Google, and xAI, who’s in the best position here? I mean, who’s got—

Now or 2 years in the future?

Now.

That’s a great question.

It’s got to be Anthropic.

Yeah, right now, forgetting about price or performance or wall-clock time, Anthropic has the strongest model that’s generally available, Fable 5.

And there are hordes and hordes of it.

Not the model. I’m not speaking about the model. In terms of their position with compute, the speed at which they’re deploying models, and their ability to, I guess, continue their dominance—we were going to talk about—

Well, here’s the thing, Peter.

Yeah.

There’s no easy answer because Anthropic is in the best position by far, and hordes of very talented people are going there purely because they want to see the singularity emerge. It’s like the birth of the phoenix: I want to be there on that day.

Sure.

But they’re totally—

They don’t have the compute.

—they’re reliant on Elon for the compute.

Yeah.

Elon can rip the soul out of Anthropic any day, and he’s got the Cursor guys now. He spent $60 billion getting them. They’re brilliant, and they’re starting to roll out cool stuff, and they’re starting to do the training. So, yeah, if you said 2 years in the future, then it’s really tricky because Anthropic and Elon are like—I don’t know. It’s a really interesting race.

I just want to give our listeners an understanding—

Yeah.

—of the terrain out there. We’ve got the U.S. labs competing against each other and the Chinese labs continually pummeling them, and we’re going to talk about that in a moment. So Anthropic’s the most advanced, but their compute—they don’t own their compute—

Yeah.

—which is a problem.

I would disagree with Anthropic being the most advanced. OpenAI, aside from the Chinese labs, owns the Pareto frontier—from Llama now being free to everyone to, again, as a mathematician, GPT-5.6 Pro being the only quality math model. I have no idea what magic they’re doing with Sonnet to actually get math results because it makes so many mistakes.

Yeah. Yeah, totally right.

GPT-5.6 Pro is the only—

If you look—

—proper frontier model.

Totally right. Yeah, we switched—

Yeah.

—over to o3, actually. Everybody over here is like, “God, this Sonnet has lost its mind.” But the argument there is that, well, inside Anthropic they have Mythos 2 now, so they’re another level ahead and they won’t release it to us, or maybe we can’t tell. But for our use case, on hard problems, hard engineering, and hard math, yeah, we switched everything over to o3.

So, but—

So I totally agree, Emad.

—even if it was Mythos 2, again, you would see them releasing low-hanging breakthroughs, which OpenAI has done with Astra. And OpenAI, again, has lined up the compute. They have more capital raised than Anthropic. They had the $12 billion round, so they can burn a few years of market capture. They have the consumer now moving to enterprise and enterprise shifting.

Mm.

And I think Anthropic, for all of their talent and their access to GPUs—actually, Google just built them a gigantic TPU deployment, like a million-unit deployment—they’re shooting themselves in the foot from an institutional perspective because Opus 5 is unpleasant. Sonnet is unpleasant to use, and I don’t think it’s going to get more pleasant to use.

Yeah, didn’t Alex say he actively hates Opus 5?

I said that.

Oh, no, I did say I don’t like Opus 5. I prefer Sonnet 5. But I think the truth on the frontier is materially more nuanced. Again, if you look, Emad, at FrontierMath Tier 4, it is the case that Sonnet 5 outperforms, ironically, OpenAI’s latest o3 model, even though OpenAI was the primary sponsor behind Epoch’s development of the FrontierMath Tier 4 model.

I think the truth is a little bit blurry, in part because the frontier isn’t zero-dimensional. It’s a 1-plus-dimensional frontier, where, if you’re willing to pay a lot and wait a long time for Sonnet 5 to do something, it’s impressive. But if you’re resource-starved, cash-starved, and time-starved, then you can probably get better performance at a different point on the optimal cost frontier by, say, using o3.

I just want to point out to everybody listening: it’s not obvious, right? There’s a lot going on, and we’re seeing China constantly leapfrog. So, Salim, you were going to say?

I have a hot take.

These frontier labs are facing the innovator’s dilemma from hell. We talked about this before because you’ve got the Chinese open-source models from one angle, compute constraints from another angle, and government regulation from a third angle. This is a nightmare while everybody else is moving quickly with open-source models. So this is a very difficult place to be in.

The good news is you can see that they’re all trying to get into certain verticals and revenue streams as fast as possible to reduce their dependence on the frontier model and on being at the edge as their core innovation capability.

I mean, the abundance take on this is that we, as the consumers and users, are the beneficiaries. It’s demonetizing very rapidly at the same time that it’s expanding.

When frontier labs compete, you win.

Yes, we all win. All right, I’m going to move us on. We’ve been saying for some time on this pod that the U.S. needs a powerful open-weight model to contend with what’s coming out of China, and this week NVIDIA is stepping up, pouring $6 billion into developing an open-source AI model and inference infrastructure designed to give U.S. developers a domestic alternative to Alibaba, DeepSeek, and Kimi.

The deal struck between NVIDIA and the AI startup Poolside aims to build one of the world’s most powerful open-weight models. By building its own open-weight model, NVIDIA is moving up the stack, as we’ve discussed, from silicon to software, positioning itself not just as a chipmaker for AI, but as a platform provider for open-weight ecosystems. From my point of view, it looks like everybody’s going up and down the stack. We’ve seen Anthropic, we’ve seen OpenAI, and obviously SpaceX AI is doing the same. Emad, let’s go to you first. What are your thoughts on NVIDIA and Poolside?

Yeah, so I’ve been talking to some of the investors out here at this TechBBQ conference who invested in Poolside originally. They tried to raise $2 billion at the end of last year.

Who is Poolside, first of all?

Poolside is a company, I believe, the former CTO of GitHub, Esa Khant, and others set it up. They originally wanted to create a coding model, and then they moved to an open-source model and model factory called Laguna that outperformed Thinking Machines’ Inkling model when it first came out.

A few months ago, they tried to raise $2 billion for a massive Blackwell cluster, and they couldn’t, so they lost that cluster. They were like, “This is the table stakes we need.” But they built a really great, solid open-source model for its size. Now what they’ve done is benefited from this weird NVIDIA acqui-hire-type thing, where NVIDIA is saying, “We need to build great open-source models to increase demand for our technology on the Nemotron stack.”

The first thing they did, actually, was hire—and I don’t think it’s been announced yet—Ashish Vaswani’s team from Essential AI. He was one of the authors of the “Attention Is All You Need” paper. Now they’re going to be making more and more acquisitions up and down the open-source stack to be the leader in open source because, again, that drives demand for the GPUs more than anything.

I think this is just the first—or, well, not the first, this is the main one—but there’ll be many more acquisitions, and they’ll have a full open-source stack. This is the Nemotron coalition. A lot of the classic ones, like Mistral and Cohere and others, won’t be building open-source models anymore. They’ll be building to the NVIDIA reference design.

Hmm. Alex?

I think maybe I could say something nice about the American open-source community and open-weight models moving in a positive direction. Obviously, NVIDIA had invested, I think, about $1 billion in this company previously, and now through this—I’d call it a hackquisition—they’re finally turbocharging their own Nemotron community.

It’s more interesting to me, or concerning perhaps, that hackquisitions still need to happen in this day and age. It’s also, I think, bizarre if you follow some of the recent hackquisitions. My original take on this was that this was just an attempt to avoid regulatory scrutiny or antitrust scrutiny. But I’ve started to see some of the other hackquisition targets come back to life.

What I’d left for dead as the carcass of the original company is where all of the founding team comes over and all of the core IP was, quote unquote, “non-exclusively licensed,” which, I think, was my understanding of the case here as well, where NVIDIA is non-exclusively licensing key Poolside IP. I’ll be watching closely what happens to the part of Poolside that did not come to NVIDIA.

I think my original expectation—that this is just a carcass left over after the hunt, being left behind purely to avoid regulatory scrutiny—may actually have life to it. Not investment advice, but it could actually be, in some sense, even more interesting than the part that goes over to NVIDIA.

I mean, there’s a lot of pressure for the U.S. to develop top-tier open-source platforms right now. Dave, what’s your take?

Yeah. I’m curious, Alex. You said quickly there that you’re surprised hackquisitions need to exist in this day and age. But I got calls from both Mercor and Oren, our good buddy Kush Puvvada, who was on the pod a week ago, looking for acquisition targets to accelerate. The hiring cycle is too slow. I need groups of 3, 10, or 15 people who work really well together. I don’t care what it costs. Send them to me tomorrow.

So it seems to be, at least in terms of my inbound, an all-time high in hackquisitions. Why do you think it should be a thing of the past?

Well, I would distinguish between talent acquirers, or acqui-hires, on the one hand, which are largely about getting talent, and hackqui-hires with an H, which are about, at least ostensibly, avoiding antitrust scrutiny.

So if you’re NVIDIA and you want to hackqui-hire, say, Poolside, you’re going the hackqui-hire route rather than just doing an honest-to-goodness asset acquisition or conventional acquisition of Poolside because you want to argue, “No, actually, we’re just a licensee of Poolside rather than the acquirer. No, we’re leaving a competitive open-source model layer,” blah, blah, blah. “This is not tying,” blah, blah, blah.

That’s the argument and principle behind a hackquisition.

Okay, so I have a very specific answer to that, too. Remember the Windsurf deal?

Of course.

Here’s the constraint. The FTC is very, very friendly to acquisitions right now, and things tend to move quickly and easily. On the other hand, the timeline for AI companies is so short that the statutory 30-day review alone is like a lifetime.

All these megacompanies—a big one like NVIDIA—are always going to get a second look, which is usually 60 to 90 days. So you’re like, “Forget it. Let me just slap together any type of deal that doesn’t need that regulatory review and just help. Train the freaking billion-dollar model or $6 billion model. That’s all I need. Let’s go.”

Then they can close the deal like Elon did with Cursor, and close the deal many months later after an HSR review and after the 90 days, or sometimes it’s even longer than that. But there’s a statutory 30-day period they just can’t get around, and that’s just like a lifetime.

Salim, you’re smirking over there. What’s up with you?

I don’t know. I think Dave’s got it exactly right. I think that’s what’s going on here. This is just purely juggling the regulatory hurdles and obstacle courses. There’s one little wrinkle on this: The remaining company actually has something called Poolside Infrastructure Company, which is building a 1.2-gigawatt data center that might need GPUs. So they may use some of the money they get for GPUs. Who knows? It gets really complicated.

Going back to the other point here is the verticalization of companies, right? I mean, xAI—SpaceX AI—is the ultimate verticalization out there today. But here we see NVIDIA. We’ve seen Anthropic and OpenAI also designing their own chips. Does every one of these companies ultimately become at least 2 layers, if not 3 layers?

In other words, does Anthropic get a space station?

No.

Or a moon colony?

You know, I think they’ll be the only company left among all the governments. Is that what we learned? Yeah, they’ll have everything.

That’s why they’ll be American GDP, that’s why.

Yeah.

They’ll be the only ones left.

But yeah, probably. I’m asking the question seriously. Does Anthropic get a moon colony? Yeah, probably.

Yeah.

Does Anthropic get a pharmaceutical arm? Yeah, already. So, yes.

Yeah. All right.

Actually, thinking about our discussion earlier, you have a split of innovation versus execution, and so these are the 2 model splits that are occurring.

Execution drives the majority of the economy in the short term, innovation in the longer term, and verticalization is ideal for the execution phase. So if you look at the architecture of Jalapeno and where things are going, you're going to get closer and closer to the silicon and closer and closer to the customer, and you won't need much better models than you have now. Whereas the frontier will be a different story, where you still need to have very complicated things occurring.

Hmm. All right.

I think there is probably a natural verticalization at the infrastructure layer—not necessarily at the application layers, but at the infrastructure layer. For physics and other reasons, there are natural reasons why, say, a company that offers a frontier model probably wants to be in the data center infrastructure business, probably wants to be in the energy business, and probably wants to be in the satellite business. These are all innermost-loop-type businesses. Robotics is another one. There are such natural synergies among all of the different innermost-loop stages. There is probably some natural vertical integration there.

I'm going to move us along here. Three stories this week chronicle the challenges being faced by the U.S. closed-frontier labs, which are under siege from faster, cheaper Chinese open-weight alternatives.

The first story comes from Moonshot AI—not related to the Moonshots podcast—the Chinese lab that's making Kimi. Last week, we discussed how access to memory is becoming the real roadblock to all of this growth. It's not GPUs; it's memory, especially in the agentic age. This week, Moonshot AI released Kimi Linear, a new architecture that cuts context memory by 75% while still delivering 6× faster decoding for a 1-million-token context window. Moonshot AI just dropped this, and it's running—an improvement of 75% in a single move.

The second story in this block comes from the Financial Times, which reports that Fable 5, Anthropic's flagship model, is now struggling to attract users. It's effectively plateaued, and the reason is simple: cheaper Chinese open-weight models are eating the market from the bottom up. When the model costs 14 cents per million tokens and delivers 80% of the capability compared with $15, the market chooses the less expensive option, at least in the general market. Fable 5 is not losing because it's bad. It's losing because it's overpriced relative to the open-weight alternatives.

The third story, and then we'll talk about this, is that Anthropic this week reversed its data-retention policy ahead of its IPO, letting enterprise customers keep data on their own cloud infrastructure rather than Anthropic's servers. This move handles the biggest objection that corporate buyers have when adopting Claude. I don't think the timing is accidental. They're about to go into IPO mode, and I think it's projected for as early as 6 weeks from now. Growth requires enterprise adoption, and enterprise adoption requires data sovereignty.

So, gentlemen, three stories here: Kimi Linear, the challenges Claude 5 is having, and the changes that Anthropic made to its data-retention policy. Dave, do you want to jump in first?

Well, this is where we're going to find out if Dario has what it takes to be a public-company CEO because he's a brilliant, good-natured AI researcher thrust into this. When he gets interviewed, he says, “I never expected to be a CEO at all, but here I am.”

So now he's stuck with missing revenue numbers because of the current policy. The prior policy was that even if you're hosting Fable 5 on Amazon Bedrock in a secure environment, everything still has to go to Anthropic headquarters for 30 days for them to review and make sure you're not making a virus or a bomb or something, and that's the only way this is safe.

Now he's missing revenue numbers because corporations don't want to give their proprietary secrets to any company they don't know well for 30 days, so they're rushing to the Chinese models and secure environments. You can also now get GPT Sol inside a secure environment where it doesn't get transmitted to OpenAI. So you can use that, too.

It's like, “Oh, God, corporations hate this, but I don't want to miss my revenue numbers. I want to go public.” On the other hand, I really don't think it's safe. I feel like I need to inspect everything to know that it's safe. So now he's stuck between a rock and a hard place.

It's tough.

It's a tough place to be. But being a public-company CEO is always like that. It's really, really stressful and really hard. So we'll see if he has what it takes to do it.

Emad, what do you think of Kimi Linear?

First, we banned the faster silicon from China, so they built MoE models to take advantage of cheap DRAM. Then DRAM became expensive, so they figured out better mechanisms of linear attention, caching, and denser models, et cetera.

Now, you've just seen, actually, just a couple of hours ago, this new o1 stealth model that's been tearing up the benchmarks. It turned out to be a GLM model served entirely on Chinese chips, with trillions of tokens a day on these new Huawei chips. So I think you'll see the adoption of these Chinese models and them moving to where the market is, on different form factors of different chips.

Again, memory is 50% of all spending now. It is the scarce resource. You can't upgrade it. So guess what? In a couple of months' time, at the very least, given the pace of Chinese models, they won't need much memory. That's how fast they innovate.

On the Fable uptake, it's entirely a zero-data-retention issue. As a corporation, you cannot leave your data on an Anthropic site, and they realize this. But Anthropic should not IPO. If you're in the late stages of AGI now, Anthropic should do a giant fricking raise, like OpenAI did, of $120 billion, and have a straight shot at AGI.

Yeah.

That's what they should do, and they should stay private like Stripe.

That's a fascinating thought. Why are they racing to an IPO?

It makes absolutely no sense to me. Dario owns 2% of the company, as do his 7 co-founders. They don't care about dilution. They're still worth $6 or $7 billion each, and they don't care about money. They've pledged to give away 90% of it. Why would you IPO? I can see no reason for that unless—

I have an answer.

Unless they can't actually do it privately, which I think they can.

Go ahead. Go, Salim.

I have an answer. They need the capital to get compute.

They could raise the capital. I bet you people would throw money at Anthropic right now.

I've been talking to investors, and Dave, this will be interesting from your perspective. Nobody knows how to price this thing because they haven't secured long-term compute like OpenAI has or that natively Groq or Gemini has. So this is why, to our earlier point, this is why people are verticalizing: if you're one layer and the bottleneck goes down below you or above you, you're screwed. So you have to have access to the whole layer to have continual progress.

Yeah.

But even if they had the money to buy compute, where are they going to buy it from? There's not enough compute being manufactured.

Well, wait. Before we get to that, Salim is right, but it's much more specific than that. Dario got ripped by Alex Karp. We showed the video on this podcast. He got absolutely ripped to shreds, and Alex Karp is saying, “Look, you cannot give your alpha, you cannot give your weights to this company, Anthropic. You cannot trust them with your corporate intellectual property. You're talking about an academic who's never run anything before in his life taking your intellectual property and then preaching to you how the government should be run in the future. Don't trust him.”

So he just ripped him to shreds. Dario can't react to that by saying, “You know what? I'm going to delay my IPO and do some private financing,” because you're playing right into Alex Karp's hands if you wuss out on your IPO plans. He's just going to reinforce Alex Karp's message horrifically.

And the board members—look at the board members at Anthropic. They've marked up those venture funds to massive valuations and used those valuations to raise new funds. So they're not going to just sit there and say, “Yeah, Dario, go put it off indefinitely. That's fine.” This is the stress test for Dario. He can't just wuss out right now, and that signaling would be terrible.

Interesting. Alex, AWG, your thoughts, please. I'm sure you have many.

Okay. Well, first, on the race to IPO, I think there is also a race element. I think there was a starting gun a few months ago between SpaceX, OpenAI, and Anthropic, and if I'm Anthropic, on top of the arguments that everyone else here has already raised, there's a competitive element if you don't necessarily want to be the last to IPO.

The market winds could change. Right now, it's a relatively warm and friendly capital market for IPOs, so to the extent there's a window for raising the largest IPO sum in human history, I think you go for it.

But to the earlier points, in lightning-round succession: Kimi Linear. We’ve known about Kimi Linear attention since last fall. I think it’s suggestive, as I’ve suggested in the past, in a Ship of Theseus style. Transformer architecture is getting incrementally replaced piece by piece.

It’s interesting insofar as it’s a successful linear architecture. Many have tried to linearize the infamously quadratic attention mechanism. It looks like Kimi Linear Attention may be one of the first, at least openly, linearized or quasi-linearized attention mechanisms. There’s a recurrence mechanism in there as well. So that’s kind of interesting. We’ve known about that for a while.

Fable 5 is struggling. That is interesting because I’ve made the point on the pod in the past that OpenAI made a strategic blunder in pandering to consumers rather than to enterprises, thinking that consumers would be hungry users of reasoning tokens, and they just weren’t. Consumers didn’t know what to do with all of these shiny OpenAI reasoning tokens, but enterprises did. Then OpenAI had to do this painful pivot over to enterprise and turn everything into Codex, and probably delay their IPO as a result.

So Fable 5, which is, at least by my accounting, the strongest, most frontieriest model in the world right now, to the extent that it’s struggling to generate revenue and uptake, I want to interpret that—I want to construe that—as the enterprises of the world almost falling prey to the same thing consumers with OpenAI did. This is maybe going to be construed as victim blaming, but it’s not. Our economy isn’t worthy. It’s not clever or wealthy or successful enough, on average, to know how to use Claude 4 properly, just like consumers didn’t know how to use reasoning tokens from OpenAI, and as a result, OpenAI had to pivot.

I think this is the beginning signs of Anthropic being forced to do some sort of pivot. It could be radically reducing the cost of their models. That’s one direction. Or I think the more exciting model, or the more exciting trajectory, is some new use case getting unlocked in the next year that actually motivates the usage of this nosebleed-priced high end of the frontier, which is, at the moment, Claude 4 or, depending on the reports you read, maybe Claude 4.1 may be starting to leak out.

And then, very quickly, on Anthropic and their data-retention policy: Anthropic was so clever, I think, in being the first frontier lab from America to enable their frontier models to be hosted by third-party hyperscalers rather than having to host them themselves. According to some reports, 40% of Anthropic’s revenue now comes from Anthropic models being hosted not by Anthropic, but by third-party cloud hyperscalers.

So I think this is just another step toward externalizing the hosting of their model. Sure, it’s painful. This data-retention policy I view as largely security theater. I don’t think it’s that valuable in the long term. I don’t think it’s useful in the long term. But starting to move more and more of the infra layer over to third parties so that users of Claude can get Claude where and when they want, on the infra they want, that’s powerful, and you see now OpenAI copying Anthropic in doing that.

What do you guys make of the idea that the open-source Chinese models are good enough and at a de minimis fraction of the price?

And companies are beginning to shift in that direction, saying, “We’re not going to use Fable 5; it’s too expensive.” Dave, is that an experience you’re having?

Yeah, no, everybody needs the absolute best AI they can get. You can’t go down a notch, but the Chinese models aren’t down a notch. They’re absolutely on the frontier. You won’t even notice the difference in any use case.

So it’s not about trying to use something inferior at a lower cost; they’re just as good. And now they’re all good enough to improve themselves, too. So if you start a group within your company that’s using these models, you can start improving it inside your company if you get the talent. And so that’s like a runaway train. I think that’s what’s really going on.

It’s not compromising to save a few pennies. It’s like, “Wow, we can control our own destiny and be on the frontier at the same time.” Okay.

Well, I don’t think it’s a few pennies, right? Fable 5 scores 60 on the Artificial Analysis benchmark. The new GLM Flash model that dropped today scores 57, and it is 100 times cheaper.

Yeah.

100 times.

Which I think is really important because, when you deploy these things, the cost-benefit is so high that you might say, “Well, I don’t even care about the cost.” But then you say, “Oh, wait. If I use the Chinese version, I can have thousands of them.” This is why the swarm is such a big deal. You can afford 5,000 or 10,000 concurrent Chinese operators instead of one Anthropic.

Well, I think it’s like hiring a specialist: a super genius versus a bunch of really smart people. Sometimes you’re not smart enough to ask the super genius the right questions.

Yeah. Well put.

Because maybe I’m not smart enough to ask Fable the right questions, but I’m just about smart enough to ask GPT-5 all the right questions, right?

Well, but also, that analogy is perfect because people misuse their context window horribly, and I do too; everybody does. But if you actually optimize the context window with the Chinese model, you’ll get a smarter answer than if you’re sloppy using a Fable model.

And so, if you just put a little energy into your internal org design and optimize your use, then you can have thousands and thousands of these for a very low cost. And that’s where the puck is going.

Mm-hmm.

I’m not sure how sustainable this situation is. I almost want to analogize it now to the US importing generic drugs from Canada. The drugs get invented in the US, they get manufactured cheaply in Canada, and then, at least historically, it’s been the case that you could get American drugs more cheaply from Canada by importing them on or off-label than you could from American drug companies.

I think the situation may be somewhat analogous here, where these are US models, US reasoning traces. You see Chinese labs benefiting legally or illegally from the reasoning traces from interacting with US models. Then, just in the past 48 hours, we start to see stories of Chinese labs trying to strike partnerships with US hyperscalers to host the Chinese labs’ models on US infra, but with a rev share from the inference costs going back to the Chinese frontier lab.

So this is a case where the US does whatever innovation is necessary—data or post-training or whatever—and there’s a distillation maybe over to China. China sells it back to us, but then we’re using our own infra against ourselves at inference time, against the training time. I think it’s a perverse bind that we find ourselves in, analogous to Chinese drug imports. Or, sorry, Canadian drug imports.

I’m not sure the Canadian drug-import analogy holds very much longer, given what’s going on, but let’s leave that aside.

Yeah, it held until about a year ago.

3. AI Rewrites Modern Dating

All right, I’m going to move us on to a fun story on the dating front. A Berkeley startup called Ditto is playing Cupid. Ditto is an app, or an AI, that has no feed, no swiping, and no infinite scroll. You fill out a values questionnaire, and then every Wednesday at 7:00 p.m., a text arrives with a match, as well as a place and a time for you to meet your date.

That’s the entire product. You show up and see if the magic happens. Thus far, 160,000 college students have signed up. It’s already produced 80,000 dates. The app does what Tinder and Hinge refuse to do: it removes choice.

The entire dating industry is built on the premise that more options are better, but Ditto believes that too many options lead to decision fatigue and analysis paralysis, and that AI is the cure. The app does not ask you for a choice; it chooses for you. This is sort of the old-style matchmaker agent, a Yenta, if you will, and it seems to be working.

Salim, you and I are both married, but it seems like it’d be a fun thing for us to go out and try.

So, 2 or 3 things. I think this applied to non-dating would be really profound, and we’re actually looking at doing something like that for business connections. But I think this is powerful because AI isn’t adding an interface; it’s deleting the interface, right?

Mm-hmm.

Tinder optimized searching and connections and so on, but this makes the searching unnecessary. I think that’s a really powerful user-interface experience where people are going to go, “Let the AI figure it out, and then I’ll do the connection and see if there’s chemistry there or not,” which you have to do anyway.

By the way, let’s note, as a scarcity-to-abundance paradigm, that when we were all growing up, sex had a scarcity paradigm. With Tinder, sex became abundant. Where the hell was that in our 20s is the obvious question. But you have to deal with that abundance in a different way.

So this is a really fascinating thing. I’ll watch this very carefully to see where this goes.

Yeah, this is the abundance thesis applied to dating. Dave, what do you make of it? Is this a company you would have backed? Oh, God, yeah. Yeah, absolutely. But I think this is a stepping stone to AI helping you manage your life and your choices in general.

Bingo.

That's right. I think it's going to be, if it's done right, one of the greatest boons to mental health in world history. If it's left to manipulate you, it's going to be horrible because it's such a great salesperson. So this is a good test case. Are we going to manage it well? Is it going to lead you to the right person? Is it going to try to help you, or is it going to sell you something that you don't want?

Learn quickly.

I've always said, in the future, the advertising model is gone because your AI knows you so well. It's just like, “Please just buy me the stuff I need. I'm in decision fatigue. I'm in data overwhelm. Just take care of it for me.” Emad, your thoughts.

Yeah, I think there was a Black Mirror episode where, for dating, you just sent your digital twins, and then they did a bunch of dates just to test it out in 2 milliseconds, so you could tell whether or not you matched. It feels like you're heading toward that. But people are going to get to a point where it'll be like, you can't argue with your AI. It knows best, right? All Watched Over by Machines of Loving Grace.

Mm.

We've got to be quite careful about that because it does take away a little bit from your intrinsic humanity if you outsource your cognition and connection in that way. But, again, you're kind of feeling it already, with how much of your stuff you offload to these things. I've been getting mad at some of my colleagues and others. They're doing really well, but they're starting to slip into trusting the AI too much.

Mm.

You know? Like, sending—

I think this—

—something like, “This is human.”

I think this is such an important point. There's a bigger pattern here where AI is becoming the trusted intermediary between individuals interfacing with overwhelming abundance, right? You're going to need that trusted interface, and the question is: Do you want to outsource that trust? By the way, to your Yenta comment, Peter, the Indian matchmaking industry is profoundly about to be disrupted by this, because you could detail out the caste, the clothing requirements, the salary requirements, and boom, off you go for the matches. So this is going to be really interesting applied to that world.

Interesting.

You need to have that as an exponential organization, Salim. Come on. Use exponential means.

Exponential dating. Diagnosing.

Yeah. Alex, I think you're the only one amongst us not married, so would you try this out?

No. I think this is why we can't have nice things. Have you seen the Ditto AI Body Count Detector? Do you even know what I'm talking about?

No, tell me.

Okay. The Ditto Body Count Detector—I would characterize it as a politely suboptimal use of scarce reasoning tokens—is a tool that Ditto released that uses, I'll quote from their website, “478 facial points and 52 microexpressions over 5 seconds to estimate how many sexual partners a person has had.” This is where the reasoning tokens are going.

Seriously?

This is low priority—seriously. It's called the Ditto AI Body Count Detector. I view this as a suboptimal use of reasoning tokens when we could be, as you and I wrote, Peter, solving everything.

Solving everything, yes.

We could be solving everything, and instead we're doing body count detection. So this one gets a thumbs-down from me.

All right. The reality is that most people on these dating apps are looking simply at the external parameters of the individual. Are they handsome? Are they beautiful? I think part of it is: How honest are you on the questionnaire? And matchmaking does work throughout time and across all cultures. Some of the longest-lasting marriages come from being matched because it's going beyond just your initial hormonal response to the individual.

I think there's something there. Whether or not it has sufficient data to actually align 2 people accurately is a different thing. But I think there's something there. I do agree, Dave, that this applies to so many different areas, and, Salim, I know at the Abundance360 Summit we have 600 CEOs. Matching the CEOs and matching the entrepreneurs there is one of the most important things we do, using AI to create those matches, because randomly bumping into the right person among a group of 600 people in 5 days is tough. So there is a value proposition to be had there.

Social discovery, I do think, is quite valuable if it's for socially productive or economically productive purposes. Social discovery for body count detection—I mean, again, this reminds me of Hot or Not back in the early Facebook days.

Yeah.

I just think we could be aiming so much higher as a civilization than AI for this.

Listen, the divorce rate in the United States is 50%, which is crazy, and I think helping you discover the right person—now, the parameters it uses may not be right—but if it were possible to help you find the best person, the best match for you, there's massive value, societal value, in that. That's my feeling. I don't know if you guys—

Yeah, I'd love to know, Alex, how you reconcile this: This is a waste of tokens; we should be solving a disease with those tokens.

Not a waste, a suboptimal use.

Okay, okay. Because one of the terms you've coined in this great revolution is “patrioschmarkerschmooser, schoozer[?].” What is that thing?

Patritier Musa[?].

Yeah, you can't even say it.

No, no, that is—

By the way, Alex—

It says “Patritier Musa[?].”

Okay.

Alex loves neologisms. If you haven't seen it, he's publishing new terminology for the singularity almost every day. Yeah. Go on, Dave.

You do need a token budget for that concept. So how do you reconcile those 2?

That's what happens once we have a leisure class that can afford tokens too cheap to meter, which we don't yet have. So maybe the way I reconcile this to make you happy, Dave, is I'd say save the body count detection until after we've solved everything. At that point, do as much body count detection as you like.

I like that view. I think once you've solved basically all major diseases, that's probably a good time to start counting people.

All right.

The line in the song is that once the day had been solved. The day hasn't yet been solved.

Yeah. Good.

Okay. All right, guys. I'm going to move us on. But it's a fascinating concept, and I hope Ditto works, and there are many happy relationships that come out of it.

One more AI story before we move on to robotics. It's a Wall Street Journal article that confirms what all of us are feeling: AI is making us work harder at a level like never before. I joke that people are talking about a 3- and 4-day workweek, and I've discovered a 9- and 10-day workweek. According to The Wall Street Journal, increased productivity from AI agents is creating more work for humans, not less.

The agents produce more output, which requires more review, more decisions, more direction, and more human judgment per unit time. The founder used to manage 5 tasks; now the founder manages 50 agent outputs. The bottleneck has shifted from execution to judgment. The humans have become the bottleneck because the agents produce too much work for us lowly humans to evaluate, so this is a bizarre implication of abundance.

More intelligence produces more output, which requires more human direction, which produces more value, which requires more work. So the work is not disappearing; it's changing character from execution to judgment. Salim, over to you, pal.

This is Jevons paradox for human cognition, right? We thought AI would reduce workflow, and in fact it increased the amount of work that's worth attempting. I'll go to a little history here. When I first did the Exponential Organizations book, Peter, you and I did that together, it was 3 years of hell.

The second book was two and a half years of hell. The third book was six months of a lot of joy, but damn, an overload on the cognitive workload, right? So when you get this kind of AI slop, in a sense, for human cognition, judgment and attention become absolutely paramount.

What we've done is essentially—if 10 agents are reporting to a founder, we've reinvented middle management. It's inside your own brain, right? So this is going to cause a huge problem because you can't have machines operating at machine speed and requiring human approval on that. I'm actually facing this with all the stuff I'm trying to do today. You may be seeing the same thing with Skippy.

So the next breakthroughs need to be better delegation, permission, and escalation thresholds. We're actually designing that.

Or BCI.

BCI, maybe at the individual level. But we're seeing this live as we do that pilot program where we work a bunch of companies through this process. It's requiring a whole new threshold of escalation, governance, et cetera, et cetera.

Companies need to decide what the machines may decide autonomously and what they want to manage later, what gets audited, and what genuinely needs a human. Otherwise, we're creating a totally crazy future where AI is going to be working 24/7, and humans are going to be obligated to work 24/7 to navigate that and keep pace with it, right?

Don't you feel that way already?

More capability doesn't mean more freedom. But I'm burning the candle at both ends right now. I'm loving it, but I'm not sure how long it'll last at this pace. And you guys aren't helping, I will say.

So can I just ask Dave, Emad, and Alex: is it the same for all of you, working harder than ever?

God, yeah. Absolutely. And I'll tell you what: you've got to savor the moment because Emad and Alex will tell you it's not going to last forever.

What's really frustrating to me is that I've been recruiting some incredibly talented people for Quantum AI, and we lost an MIT Course 6-1 guy who just decided he's going to go to the Princeton PhD program. Do you listen to Emad and Alex? You guys collectively have, like, 100 degrees. Would you advise anyone right now to go into a PhD program and miss the singularity? No, of course not.

But it's frustrating to watch that happen because this moment we're in right now, you can master 1,000 AIs, 10,000 AIs, and you're the most valuable you'll ever be in human history right now, because they won't do anything productive without your help. A year or two from now, they may say, “Yeah, I don't need your help. Sorry, I don't need you anymore.”

Get out of the way.

So, yeah, work your ass off right now because it may be the last chance that you have to actually be extremely valuable. I'm still savoring it. I'm working harder than ever, by far, but savoring every minute of it.

And I tell you, working with AIs is genuinely fun, too. It's not like I'm moving boxes around or grinding it out in a cornfield. This is really, really fun work.

It really is fun.

You're discovering the future.

Yeah.

It is fun.

It's a blast. Alex?

I'm reminded: a friend of mine likes to say the Stone Age didn't end for a lack of stones. I think this era that we find ourselves in is probably pretty brief.

I know I'm getting approximately no sleep at this point, largely because almost all of my time is spent supervising and steering fleets of agents, and I think this is a window. I don't think this will continue very much longer—at most, maybe 1, 2, or 3 years.

At that point, the AIs will be sufficiently self-steering that the role for humans in being knee-deep in steering large fleets will probably erode to a de minimis role.

Isn't that an argument for just laying down, relaxing, enjoying yourself for 3 years, and then jumping in 3 years from now?

No, it's an argument for working your tail off for 3 years and then going to lie on the beach.

Huh.

I mean, if nothing else motivates you, every year millions of people die needlessly. If we just get 3 months shaved off that timeline by working our asses off—

That's right.

—millions of people will exist forever that otherwise wouldn't exist.

Beautifully said.

If that doesn't motivate you, Emad, you've got to put a clip in here, too.

Beautiful.

This is the most important thing we've ever recorded. So what are your thoughts on this?

The amount of leverage you can get per unit of your attention now is more than it has ever been, and I think, as Alex said, it probably ever will be. You're approaching the last human discoveries. You're approaching the last point of being able to deploy and control these things.

You have a limited, focused attention budget. That's why you're getting tired. Maybe, to try and coin a new logic, it's “cognithargy”—cognitive lethargy—that we're facing here.

From my own side, I've written 2 books in the last year. I've done a massive amount of research, and I've been in a flow with hundreds of agents. But last week I stopped. I couldn't do any more research because I had to go and take this out to the world.

So we're doing a big funding round. We're launching lots of new things. We'll be releasing all the research, finally. And I turned off my agents that were doing all the research. I've set them onto auto mode. No more Emad stuff. They're still coming up with things, but I can only read it once a week. I've actually made it so I can't do it.

I think you can shift between these modes of work because you can't be on all the time; it does burn you out. But at the same time, if you get in the right flow, then you can do more than you've ever done before.

And like I said, I can't imagine it. I would say on this podcast, straight up: don't do a PhD. If you're thinking about doing a PhD, don't do one. Peter Thiel paid all these people not to do PhDs. I mean—

Well, not to do college degrees, let alone PhDs.

I would say not even do a college degree. What will you get out of it right now? You will go and learn a very specific thing when you should be learning agency.

A Thiel Fellow—the type of person to do that—will go way bigger than they've ever gone before. Parents might complain and things, but show them what you create. Gather people, humans and agents.

A skeptic would say, “All right, Emad, you went where? Oxford, as I recall.”

Right.

Alex, you went where? Harvard and MIT. Dave, you went where? MIT. Peter, you went where? MIT, Harvard, and so on.

Okay, so you're pulling the vertical mobility ladder up behind you, and it's fine to tell everyone else who's just coming up, “Don't bother with higher education. Don't bother with credentialitis. Just go off and do your startup.” And yet we didn't follow that.

Well, we didn't have AI at the time. We didn't have—

It was a different time. It was a different time.

Still crazy then.

That just makes you more credible in what you're saying.

I would actually say undergraduate is still a lot of fun, and you don't really have to work that hard, like I thought.

It's just—you know, Elon said this—it's a social experience.

It is. But—

It's adult daycare.

It's adult daycare, so it's actually great for using massive amounts of agents. PhDs, though, I don't get, you know? Especially—I feel sorry for a bunch of my buddies who are math PhDs. A couple of them had problems solved in the recent batch. They don't even know what they're going to do.

Every verifiable-domain PhD is now under massive threat. Why would you even do it or even consider it? If you do a Ph—

I agree. I was speaking a few days ago to a government-funded AI-for-physics center filled with PhDs—current PhDs and recent PhDs in physics—and I leveled with them. I said, “Physics is cooked, and you should probably be reconsidering all of your career trajectories. Consider any advice to the contrary, and give that a double think, as it were, before you just go and follow some zombie pattern.”

Well, I think this is the most important conversation—

Out of the way.

—we've had yet. We've got to put this in Circulator everywhere.

I spooked them. I spooked the heck out of them.

Well, getting people to think is the most important thing. Why are you doing a PhD? A lot of people are doing a PhD because they told their mom and dad they were going to do it, or their sibling did it, or they thought that was what they needed in life. And I think—

Well, or actually, in a lot of cases, about 5 years ago, before anyone knew the singularity was coming, they started working their ass off toward that. You've been working so hard for so long, and then you get in, and it's like, “I got in.”

Now the idea that suddenly it's irrelevant or you shouldn't be doing it is so hard to take after you worked so hard to get there.

But you gotta pivot. You gotta just recognize the moment.

You're like Elon. Get into your Stanford PhD and say, “Okay, I checked that box, and now I'm gonna jump into a company.” Well, it's important for people to realize the world is very different than it was before.

All right, I'm gonna move us forward. This is a conversation we've had before: 2 stories on the data center debacle. The first story is about public sentiment. So, a year ago and then again this month, a year later, Heatmap News polled Americans about data centers. A year ago, Americans were split 43 to 42 on whether they opposed data centers being built near them. Today, the opposition has risen to 75%, with 61% saying they are strongly opposed. Also this week, Senator Bernie Sanders once again called for a nationwide moratorium.

The second story is about a post on X that went viral about data center water myths. We've talked about this on the pod before. Here's the data on people against data centers. It's been increasing. It's almost, I guess, a linear increase, but it's gonna asymptote near 100%. The post on data center water use was pretty damning.

So here are the numbers. Data centers are at 627 million gallons per day. Sounds like a big number, but compare it to golf courses at 2 billion—3 times as much—or power plants at 133 billion, or growing cattle at 137 billion. The fact of the matter is that the tech industry has a trust problem, and I think we've talked about this before.

If I were a hyperscaler building a data center, I would do this very differently. I would promise, “We're gonna put education programs in the schools. We're gonna make the cost of energy in your community lower than it is today, and we're gonna make these data centers not look like ugly boxes. We're gonna make them look like cathedrals.” I mean, spending an extra 10%, I don't know why that's not going on right now. Honestly, I don't.

My worry is that that wouldn't help. My fear is that this isn't because people think data centers are unsightly or unaesthetic. My concern is that it's being overly politicized, in part through the worst-case scenario, which would be foreign interference. There are a number of U.S. adversaries who would love nothing more than to slow down America's data center build-out. We talk about at least 1 of them all the time on this pod.

So my concern would be that we look back in a year or 2 and see that some quantum of this opposition to data center construction is actually the result of popular sentiment being stoked by foreign adversaries.

Okay. Agreed, Alex. But why isn't the White House getting out in front of this? When I was on with Michael Kratsios, I said, “Why isn't the White House getting out in front of this?” It's an issue that's gaining steam. People can see this.

The second thing is you can counter that by saying, “Listen,” like this is what Zuck said in his video last week, right? “We're gonna give you better schools. We're gonna give you better access to jobs. We're gonna be a positive contributor to the community.” You can get to a point where having a data center is such an advantage to your community that people are gonna say, “I don't— that's false news. Here are the facts. Cheaper energy, right? More jobs.”

Here's the problem, though. In the U.S. way of doing things, the decision of whether to site a data center or not ends up being a local decision, not a national decision. Whereas in China, they can just declare, “Okay, the east is going to be responsible for data, the west is going to be responsible for compute and energy, and we're going to build this national-scale grid for combining compute, data, and energy together.” And poof, you're the CCP, and you get to centrally command the whole economy.

In the U.S., we have a different system where individual local municipalities and states get to say what they do and do not want their land used for, and we end up in this system that's far easier, if you're a foreign adversary in the worst-case scenario, to polarize and to shut data centers out of terrestrial deployment.

But this is false data. This is an outrage cycle on social media. This is people reposting—

Shock, shock that foreign interference would leverage false data. Shocked.

Wait, let me say a couple of things about this. Can I?

Yeah, please.

Okay.

Mr. Mostaque.

So we have a problem where our information systems reward compelling narratives over evidence, and this data center thing is the heart of that. It's a problem that's been building up over decades with the use of social media. We are not evidentiary-based in the U.S. at all. This is really a big challenge because we're totally narrative-driven and not evidentiary-driven at all. We decide what the story is, and then we go looking for facts that support it, okay?

Data centers are a great hobby horse for this. This is happening everywhere. We've gone from, say, 50 years ago, “Show me the evidence, and I'll form an opinion,” to, “I have an opinion, now show me the evidence that confirms it.” And that's amplified radically with social media because nuance has no viral coefficient. This just doesn't actually work.

So this is a very difficult problem to solve, actually doubly enhanced by the interference that I'm absolutely clear is happening, and I'm with Alex on this one. The problem is we're making national policy based on these stupid innuendos and memes rather than measurement. You cannot run an advanced civilization with this. This is a massively big issue, a huge opportunity to make—

Yeah.

—humanity go from scarcity to abundance and measure it, compare it, put it in context, fix the externality, but don't legislate with a story in your head, which is what the hell is going on right now. It's a completely disastrous problem we have. It goes to the cognitive issue of the U.S. Sorry.

Emad, do you want to hear something really cool? I don't know if you ever met Rob Fisher. He was the president of Link Studio for years. He left to start a data center company a few years ago, and they're killing it. It's called Provocative AI. The data center is actually water-negative and carbon-negative.

There you go.

It's so cool. It's doing its own carbon capture using waste heat, running mostly off nuclear power from Seabrook, New Hampshire, and it captures more carbon than the entire loop produces. And they said, “Oh, you know what? We could actually use just the humidity accumulating because of the temperature gradient to create more water than we consume and just use our own dripping water.” Then nobody can complain. We're not— we're actually water-negative and carbon-negative. So it's really cool.

And yes—

But it shows you how little water they actually use. Alex, wasn't there a story recently about NVIDIA's new chips and new data center structures that are actually utilizing less water now?

Yeah, well—newsflash: there's no water in low Earth orbit, so that's the end game, I think. Just cut the water nonsense out. This is only forcing all of these new data center deployments to sun-synchronous orbit. We might as well just get it over with.

Yeah. How intelligent was Elon's move? Prophetic—

I think it was opportunistic. I think he laid all the infra for Mars and then opportunistically and timely pivoted to sun-synchronous orbit in the Dyson swarm because he read the tea leaves.

Huh, amazing.

Can I say something more?

Yeah.

Can I just say one more thing? If I lift up a level to the rationale and foundation of why this podcast exists, to reach abundance we need an evidentiary foundation in our culture. Otherwise, every new technology is gonna be strangled by the narratives that go viral before the evidence can spread.

This is the fundamental foundational problem we have with civilization. As Alex said, this is why we can't have nice things.

I reckon we should just rename them. Let's call them intelligence boundaries or compute citadels.

Change the narrative.

It's got a branding problem.

It's a branding problem.

Yeah.

Again, this is not a factual thing.

I have a better one.

No, keep going.

I have a better one: AI churches.

Yeah.

A compute citadel beats AI church, come on.

Fine. Sorry to interrupt you, but anyway.

4. Waymo Cuts Robotaxi Costs

All right, I'm moving us along here. Let's jump into the world of robotics. For the longest time, the economics of Waymo versus Cybercab have been devastating.

Elon projected that a Cybercab will cost about $30,000—that's what he said he'd sell them at—for the vehicle and sensor hardware, compared to Waymo's Gen 5 Jaguar, which costs about $300,000: $200,000 for the vehicle and $100,000 for the full autonomous-driving hardware package. In other words, Waymo is coming in, or has been coming in, at a 10-times disadvantage to Cybercab.

This week, Waymo announced a significant redesign in cost savings. They announced details around their custom 5-nanometer chip that processes camera, lidar, and radar data in real time—at 1 quadrillion operations per second. I love this. We've gone past trillions; we're at quadrillions already.

Quads.

Helping slash their 6th-generation autonomous-driving hardware costs from $115,000 to $20,000. At the same time, Waymo unveiled the Zeekr RT, a purpose-built robotaxi minivan designed by Chinese EV maker Zeekr. The OHIG costs $75,000 per vehicle, compared to $200,000 for the Generation 5 Jaguar.

It's 42% fewer sensors: 13 cameras and 4 lidars, compared to 29 cameras and 5 lidars. And remember, Elon made the point years ago that if a human driver can drive with just 1 eye, you should be able to do all the driving with just visual sensors. Also, in related news, NVIDIA this week gave permission for Tesla, Uber, and Waymo to simultaneously begin operations in Las Vegas.

So let's watch a quick video about the new Waymo. I had a chance to ride in it yesterday. It's a pretty cool vehicle—kind of not as sexy as the gold Cybercab, but take a look.

Here's what Waymo calls its 6th-generation driver: the hardware and software system that actually does the driving. Combine that lower-cost Chinese hardware with this new interior tech, which the company says was designed to cut sensor costs while improving performance, and the math starts to move in Waymo's favor in a way that it hadn't previously.

Now, the last system running in the Jaguar fleet had significantly more sensors. The new one uses 13 cameras, 4 lidar sensors, and 6 radars, and Waymo says it performs better. The company switched to 17-megapixel cameras, a major jump from the previous specs. Higher resolution means the system can see more with fewer cameras.

They've slashed the total sensor count by more than 40%, so cost is down and capabilities are up. The new system also builds heaters, wipers, and sprayers directly into the sensor pods, which helps them clear snow, ice, and road grime.

All right. Well, a good move by Waymo. I've been using it pretty regularly here. It's much cheaper than Uber. Alex, let's go to you first.

Venting some pain here. Jaguar is owned by an Indian company now but was doing its manufacturing largely for Jaguars in the UK.

Jaguar Land Rover, yeah.

Yeah. But was doing its manufacturing largely for Jaguars in the UK. Look behind the headline. Careful what we wish for with whoever here is suggesting that Google should just switch over to fine-tuning Chinese models. Newsflash, Google, Waymo, and Alphabet are switching over to using and OEMing Chinese hardware in order to achieve Waymo's objectives.

I would rather see the West use a Western hardware stack rather than just white-labeling Chinese hardware. That's somewhat disappointing. It's also kind of interesting if you look underneath at the overall chip supply chain that they're using. It seems like they're moving away from Broadcom. They're vertically integrating, which I think is a theme that we were speaking about here earlier.

Waymo is—maybe Waymo wants its own space station at this point. Waymo is getting its own chips. It's OEMing Chinese hardware at the hardware layer. Maybe it's playing footsie with Uber for the moment for distribution, but probably wants to own its own distribution in the long term.

I know whenever I use Waymo, I'm not using or engaging with Waymo via some aggregator app. I interact directly with Waymo. So I think we're starting to see honest-to-goodness vertical integration here.

I also wouldn't be surprised if Alphabet and Waymo are starting to drive costs down in this case, I guess, by white-labeling Chinese hardware. There's an interesting historic rhyme with Tesla, which started with the high-end Roadster and has been pushing down costs right up until they hit the autonomy barrier. At which point, remember, the Tesla Model 2 that was supposed to launch but never did—

Yeah.

—that was going to be the highly vaunted $25,000 vehicle. It never launched because Tesla hit autonomy instead. And below some threshold in car price, maybe it doesn't make sense to sell cheaper cars. It makes more sense to just get out of car sales entirely and offer hosted autonomy platforms.

I think we're going to start to see Waymo, at some point, in order to drive the cost down, just ditch all third-party vendors and turn into a white-label sort of deal for Chinese hardware—or maybe American hardware—and their focus is entirely on software again.

Yeah, the vertical integration is completely unprecedented in history. It's something—it's a byproduct of the singularity that I don't think I fully grasped until now that we're living it.

But if you look at the largest companies in history, you'd have ExxonMobil doing oil. You'd have IBM doing mainframe computers. GE, where my dad was, doing nuclear reactors and toasters. But they did different things.

Now, all 11 of the mega-cap companies are building AI chips, building AI models, and building data centers—every one of them. So they're all colliding into vertically integrated super-companies, and they're just doing the entire stack because—

And robots next.

And robots next.

They're all going to build robots. Yeah.

And meanwhile, TSMC is a sitting duck. TSMC is waiting to be verticalized.

Isn't that amazing?

Yeah.

It's the linchpin to this entire thing, and it's sitting there not doing anything. It's just—

Right across the Strait of Taiwan, ready to start World War III at a moment's notice.

Yeah, yeah.

Emad, do you ever see these vehicles coming to Europe?

Yeah, we're starting to see Waymos in London, and I think the regulation will actually be largely positive for them. But I was having dinner today with Jens Wiese, a deep-tech VC at Lake Motif, and we were talking about something interesting.

Previously, I said, “A Tesla Optimus robot gets into a truck, opens the door, plugs itself into the phone charger or the cigarette plug, and boom, that trucking job's gone.” And I was like, “Well, actually, why wouldn't you have specialist robot drivers?”

You don't need to retrofit all these cars because you think about a humanoid robot that's walking around in the real world versus one that sits in a cockpit driving a car. It's so much simpler. And I did the bill of materials. I'm like, “That's like $6,000 with the actuators and everything.” And so I was like, “Oh, crap, this could actually happen a lot quicker.”

Yeah.

The other side of it is that you fully vertically integrate. So Xiaomi just announced a Xiaomi car. They've gone from mobile phones to cars with a fully lights-out factory. And so, of course, you'd vertically integrate if you had a fully lights-out factory.

I kind of feel like these are 2 things that are coming. But actually, I really got thinking about this humanoid driver robot.

I love that idea. It's the first use case for a humanoid robot with 2 arms and 2 legs that I've yet seen, so—

Yes. Please.

I will kneel to you, sir.

Yeah. So Palmer Luckey and I talked about humanoid robots on the podcast I did with him, and whether he was going to build them. He said, “The use case for these in the military right now is getting into Jeeps or getting into nuclear silos, replacing the humans, and sitting at the desk without changing out the interface hardware. Just create a humanoid robot that can interface with what a human did before.”

So that makes a lot of sense.

It's vaudevillian again—like, a lack of imagination—but the ergonomics are such that we're incentivized to deploy humanoid robots initially into these human use cases. But I'm still pretty bullish, for what it's worth, for the next 10 years on the humanoid form factor, Salim.

Well, I think you've kind of got 2 things here. Like I said, 1 is full vertical integration. The other is human-shaped holes with human humanoids in them.

All right. I'm going to move us—

No comment.

—on to the next story, which is one of my favorites—

Wait, wait. I just want to respond to Alex very quickly.

One of the funniest things I think I've ever heard you say, Alex, a couple of podcasts ago, when I talked about why we have humor, you said, “Oh my God, Salim, why the self-loathing?” It was so funny. So I just love that. I just want to reflect back on your—

We love our humanoid form factors.

There we go. Sorry, Peter.

All right. This next story I love. I love it when a technology really fits a perfect use case, and we saw that demonstrated this week with a video out of China once again, with a hybrid life preserver and drone being demonstrated.

These autonomous rescue drones can fly at 30 miles per hour, covering up to almost 2 miles, landing on water, and providing flotation for 2 80-kilogram adults. It saves lives. We're talking about a drone that flies at 30 miles per hour, compared to a human lifeguard swimming at 2 miles per hour.

I love this product. Let's take a look at the quick video here. It's like, wow—the best use of a drone I've seen. Pretty amazing, guys.

So I thought this was fantastic for a couple of reasons, right? This is compressing time—response time, where time equals lives.

So this is so great because autonomous response is so much more interesting than remote control in this context with this type of use case. These types of applications are going to do more for public acceptance of AI than any chatbot benchmark, whatever. It’s such a great use case. I love this.

Yeah. Alex?

Yeah. So, another one of my neologisms was the broken Waymo theory. People who haven’t seen this may remember from the ’90s the broken-windows theory, most famously associated with Rudy Giuliani and the purported rehabilitation of the streets of New York City. The idea was at the time that if there were broken windows, that was either a proxy for or even causally related to broader crime issues, and that you could almost run this causal relationship in reverse: if you made sure that there were no broken windows, you could make sure that crime overall was down, or at least that was the thinking at the time in some quadrants in New York City in the ’90s and the early 2000s.

Similarly, maybe—hopefully, slightly better founded—I’ve tried to push the notion of a broken Waymo theory. The idea being that if a city or a nation can’t deploy autonomous robots, then they’re not prepared for the singularity. I see videos like this—drone life-preserver aircraft coming out of China—and I shake my head a bit because here in Boston, we can’t even get Waymos. I raised the subject, or attempted to raise the subject, with Mayor Wu a couple of weeks ago: de minimis progress. I just think we’re getting lapped by China at this point.

Well, this is what Sam said. This is institutional inertia. This is the existing players blocking their disruption. This is not a surprise.

Yes.

Not a surprise, but definitely a disappointment.

Yeah, sure.

Yeah, but I think the AI labs were very late to address PR and realize they need PR, and now they’re on it, and we’ll see where it goes from here. But the government reacts to voters. The voters are anti-everything: anti-data center, anti-disruption, anti-job loss. And that’s because the frontier labs, who are now writing documents like Machines of Loving Grace—this is the roadmap for how the whole world should be governed in the age post-AI—well, okay, but get ahead of your PR. You can’t have every voter hating you while you try to roll out that roadmap. So get ahead of your PR.

Yeah.

In China, the news is controlled by the central government, so they just dictate the PR. It’s a much easier problem in the closed world than in the free world. But at least the AI labs are aware of it now, and hopefully they’ll get on it.

I remember in Santa Monica—

The lifeguard is such a no-brainer.

I remember here in Santa Monica when electric scooters came out. After a few weeks, you’d see them hanging from trees. You’d see them in parts on the ground. People started hating them. And then we had the Waymo fires here, back—I don’t know—a year and a half ago.

So, Emad, you said this in the last pod, that these robots on the streets are going to be made illegal. I’m curious: when we start seeing Figure robots, and we’re going to have Brad Adcock back on the show here—we should talk about that—and we start seeing tests on the streets, are people going to try to capture these and hang them from nooses? I think we’re going to have an interesting sort of collision between those who can afford these robots and see them walking on the street, and those who find them super valuable for helping them at home. So stay tuned.

No, I mean, I think you will see lynching of robots and things. You will see people vandalizing them like they vandalize cars, stealing them and all sorts of things. I think, Dave, that in China, it isn’t so much about the control of the media. They genuinely see them as useful. You need them for the population pyramid in China. They’ve had a technological leap forward already.

I think that China will produce robots. It will improve the Chinese way of life, just like the electric revolution there for cars is, just like AI everywhere ubiquitously is. The short-form videos flooding things, maybe not so much. But China will do that, and I think China will stop exporting robots in 5 years.

Yeah.

This one—

I think you’re right, and I think we think that in a free country or a free economy—a free Europe—the press can report the truth, and therefore people will get the truth. But if you read what’s written in China, it’s actually much more truthful about technology than what gets published in the US.

So you talk about the water in the data centers. What’s actually happening is it’s backfiring: the free press, which is starved for any budget, is starting to publish garbage that’s actually factually not true. And so the free press is kind of backfiring right now in the age of AI.

Dave, this is what Elon said on that pod, right?

Yeah.

Basically, China has seen a technology revolution moving so many people into the middle class, and they appreciate technology uplifting them, so they are much more anticipatory and excited about AI.

And if they don’t—

And if they don’t, they also said they’ll get invited by the CCP for tea, so we’ll never hear from them.

I’m not pro-CCP. I’m not trying to imply that. But Elon also said, “We will habitually report if one Waymo in one corner of San Francisco runs over a cat; it’ll make every headline in the world—

Yes.

—and it’ll be a tragedy. But if it’s 10 times safer than human drivers, we just don’t even report it. We’re just like, ‘No, no, let’s show the run-over cat.’” And that skews the voters tremendously. Elon explicitly talked about that too. They’re just more statistically accurate in the Chinese press.

Yeah. And if it bleeds, it leads.

I just have to say, about topics that are convenient to the government, not about topics that are inconvenient.

For sure. Absolutely.

Yeah, but ultimately, this is an abundant technology. Let’s face it: in the West, we have a scarcity mindset. In China, they have a more pro-abundance mindset. I think that’s the big differential here. And the question is, again, how do you articulate great visions of the future? Moonshots Conference, Future Vision XPRIZE, things like that.

Yeah.

You have to change the narrative because otherwise people are like, “This will disrupt my job,” as opposed to the benefits of this side of things.

There’s so much reason for optimism.

I think if you look at the future of China and some Chinese people I’ve spoken to, it’s that robots do all the work, and we have really good lives. China might actually be able to pull that off. That’s why, again, why would you export your robots if you can use them to give your citizens a good life?

All right.

There we go. We are the Ministry of Superintelligence Truth for the West.

Okay.

Everybody, welcome to the health section of Moonshots, brought to you by Fountain Life. You know, we talk about AI on this Moonshots podcast all the time. One of the most important things AI is going to be able to do for you, besides educating your kids and helping you with your taxes, is making sure that you're living a healthy lifestyle, that you get a chance to get to 100 plus. I'm here today with Dr. Dawn Mussallem, the Chief Medical Officer of Fountain Life and a part of my medical team. Dawn, a pleasure.

Great to be here.

The thing that people are concerned about most about living to 100 or 120 is their cognitive abilities, making sure they don’t have dementia, and the numbers about dementia are problematic. Can you share what you’ve learned?

Such an important point, and you’re right. At Fountain Life, our members’ number-one concern is losing their brain health—forgetting the name of their child, forgetting the face of their loved one. We know that when it comes to dementia, the conservative estimates are that 45% are entirely preventable. What was amazing is that, with the advanced testing we’re doing at Fountain Life, one quarter of our members had advanced brain age.

Wow.

But what was really awesome was that, again, when we focused on prevention through healthy living—eating healthier, moving our bodies, and sleeping—

Yeah.

Optimizing sleep is so important. You know what we saw? We saw that we improved that brain age by 26%. That is a big number to show that the majority of those individuals were actually able to improve their brain age.

5. The Space Launch Race Accelerates

I'm going to move us to our last group of stories here: 4 space stories this week for my fellow space cadets. The first: Elon just announced his intentions to implement 30 Starship launches per day by 2030—more than a launch every hour. That's roughly 10,000 launches per year, more than 40 times the entire global launch rate.

Dave, you remember when we interviewed Elon at the beginning of this year. We did an epic 3-hour podcast with him, and we're on schedule to do an end-of-year prediction podcast with him again. These are the numbers he used. He said to implement Starlink, you need 100 gigawatts of solar-powered AI in orbit, and that's 10,000 Starship launches per year to deliver 1 million tons of data center payload. So he's sticking with those numbers, starting in 2028, starting to launch StarMine and hopefully get to 10,000 launches per year. That's crazy. I mean, can you imagine just sitting outside the launchport and watching them pop off every—

Yeah.

50 minutes? It's going to be awesome.

Wow.

The second story this week came out of the White House, when they released the “Golden Age of Space Transportation” report outlining the administration's agenda to streamline FAA launch licensing, expand spaceport infrastructure, accelerate commercial lunar programs, and set a target of 1,000-plus launches per year.

I mean, I've been in this industry. I ran a launch company for a number of years, and I helped co-found the Kodiak Spaceport in Alaska. The amount of bureaucracy in getting those done—making sure that the wrong tree frog isn't in that region and might get damaged by a launch—is a bureaucratic morass. It was crazy.

The third story we'll hit on here is that Starlink is taking over aviation by storm. Let's take a quick look at this data. Here's the chart. This is published by SpaceX, and what's going on is that we're getting massive adoption by all of the airlines. Why? Because people are posting on X saying, “I'm only going to fly the airlines that have Starlink,” and, “I choose a Starlink-enabled airline over a non-Starlink-enabled one.” So what this means is that the incumbents—Viasat, Eutelsat, and Gogo—are going to get crushed out of existence. Thoughts on this, gentlemen?

A few thoughts. Maybe just starting with what I perceive to be the regionalization of spaceflight and space launch. Buried under the SpaceX story, I think, is Starbase, Louisiana.

Yes.

The announcement of Starbase, Louisiana for $100 billion being invested into the Louisiana economy to build a second Starbase in Louisiana rather than Texas. What this says to me, reading the tea leaves, is that the Gulf Coast is becoming America's space coast.

Yes.

From Florida to Louisiana to Texas and so on, that's America's space coast. That's where I think private space launch, vertically integrated, including the Starbases, seems to be localizing. At the same time, going back, Peter, to your comment on the White House announcement, buried in that announcement was an executive order to the secretary of the interior to start

appropriating federal land for federal spaceports.

Mm.

And so if you pull the string a bit and ask where we're likely to get federally owned land for spaceports, I don't know if folks want to guess what the likeliest candidates are. I think we're going to get a few of them. Any takers? Where's all the federal land?

In—

Nevada, Utah.

Nevada, yeah.

Exactly. So my calculus is—

In red states.

It may or may not be a coincidence that the federal government owns so much land in red states. White Sands Missile Range in New Mexico, the Nevada Test and Training Range, and the Goldwater Range in Arizona are the leading candidates for spaceports.

So I think in the American Southwest we get federal space bases, or Starbases, and on the Gulf Coast we get private Starbases, as it were. That's how we get to this 30,000-per-unit-time launch capability.

You know—

Mm.

The reason historically all the launches were taking place out of Florida is that you were dropping stages along the way, right?

You want to be near the water, and you want to be near the equator.

Yeah, near the equator—

Well, and you have to go east, right? You have to go to the east.

Well, if you want to use the spin of the Earth to assist your launch mass.

Right.

But when you're dropping 1, 2, and 3 stages out east of you, you don't want to be dropping them on populated areas. Of course, Starship is reused. The first stage comes back, and the second stage is in orbit immediately, so you don't have to worry about that as much. You can land in a landlocked area. Except for—

We're going to get landlocked Starbases. Exactly.

That's—yeah, and except for Israel, which launches west for obvious geographic reasons. Yeah.

Which way does California launch?

North.

North. Interesting.

It's—yeah, so you're basically—

For polar orbits.

For polar orbits. I co-founded, or was part of the team at, Kodiak, Alaska, and you were launching south. So there's a large use case for polar-orbiting satellites out of Vandenberg. Actually, I'm sorry, you're launching south over the Pacific from the curvature of California, and out of Kodiak you're launching over the Gulf there. Yeah, it's going to be amazing. And, of course, Elon's true objective is not a launch every hour; it's a launch every couple of minutes.

Elon's—I think, Peter, on this podcast we used to talk about—I've never thought about launching north or south. I thought you launched up.

Yeah.

I mean, up, for those of us in the Northern Hemisphere, perhaps. But I think you also make a super interesting point, just unpacking that landlocked launch is something that we historically have not had before, but thanks to reusability, we're about to have. And then any landlocked country—I mean, I guess you could probably talk our ear off about the former Soviet Union and how it located its particular launch sites—but with reusability, landlocked launch becomes a lot easier.

Yeah.

A lot of suborbital vehicles, which just went straight up into the ionosphere and beyond the stratosphere and came back down, were being launched from White Sands and from Fairbanks.

Right.

But for orbital launches, you needed a place to land a hunk of metal.

Right.

Our final story in the space docket here is a viral post on X that shows Chinese reusable rockets that are basically a Xerox copy of Falcon 9. Let's take a look at this video, because it's very telling. If you look at this, it is almost a duplicate of Falcon 9: the same fins, the same landing capability, the same landing legs.

Same cheers.

Same cheers, yeah. Pretty crazy.

Interestingly enough, SpaceX does all of its testing in public. They describe all of their failures, and they open-source a lot of their information. China is able to catch up in the reusable-rocket category by taking advantage of that.

Well, Elon has had a pretty public policy of not going after other companies for patents in cases where SpaceX or Tesla have vast patent portfolios. I'm not sure whether he cares, but if he cares, maybe he wants to revisit that policy.

I think he wants as much launch, as many chips, and as much of all of this as possible. He's been pretty vocal about that.

I'm more optimistic than most on this, because what you've got in SpaceX is a compounding learning loop, and that's hard to break. That's hard to beat.

Yeah, I don't think anybody's going to come close to beating them. We've also got the capital markets that enable SpaceX to really design and develop, and now that Grok, or the next version of Grok, has all of its engineering data, it's going to be a lot of rockets being developed out there.

You guys open for a few AMAs?

Just a few minutes, or 1 or 2, before I have to rush to boarding.

Okay, we'll give you a first crack at this. Salim, pick your first one.

Oh my God, it's got to be number 1: “Humans suffer from mind viruses, so why would AI be any different?” And this is from @Bujin5455.

Oh, wow. This goes to what we talked about earlier, right? You've learned that intelligence does not guarantee epistemic awareness. You have smart human beings who can believe really, really stupid things. The problem with AI is the replication speed.

One bad belief can propagate through millions of agents almost instantly. But it also gives us a defensive capability because we can cross-check this. Look at the benefit on X of people checking with Grok whether something's real or not.

It's creating a really viable conversational architecture where truth—maximally truth-seeking—is actually working. Where I think the multi-agent world can work is that one agent can challenge another agent's claim, but you're going to have to program that in to have that cognitive critical thinking in there.

So you're going to need a lot of cognitive diversity to navigate this, and nature solves this through diversity, right? The problem that we have is that it's not like nature—not 1 AI with a bad meme; it's billions of AIs sharing the same bad meme because they all came from the same bad model of the same original point. So I think this problem becomes much bigger with AI agents, not smaller. But the answer is in Alex's idea of defensive co-scaling.

All right, I'll take number 2: “Is there an XPRIZE for actually curing a disease and getting the cure to market, not just discovering it?” I'll just say the following. We're looking for places that are stuck to launch XPRIZES with a clear objective function: the first person to do this.

I think, honestly, the AI labs—from the work that Demis Hassabis is doing and Dario is doing—are working on this. I don't think an XPRIZE would accelerate it. So we don't want to get into the middle of something that's already in the process of being solved. We're looking for problems that are stuck. All right, Dave, over to you, pal.

All right, I'll take number 4. It's very timely, actually. Why isn't Intel earning a fortune making chips using NVIDIA's old designs? This is from JimPlamadon67637. I was just talking to a senior exec from Intel, asking almost exactly that same question, so I happen to know the answer.

Lip-Bu Tan has the company making an ungodly fortune on Xeons and is concurrently burning that fortune on building out massive fab capability. So they're burning almost $2 billion a quarter on their fab business, and they just raised another $20 billion to build more fabs. The idea is to get that capacity up and compete with TSMC as a general-purpose fab company.

If they were to start competing with NVIDIA and the other GPU companies right now, they wouldn't be able to attract them as customers for the big new fab business. So they're being very specific about building chips and making a fortune on those chips that are not competing directly with NVIDIA while growing their TSMC-competitive business to massive scale. So that's their strategy.

There you go. All right, Emad, you want to take number 3?

Yeah. So, number 3: “Can you guys talk about dentistry? Has anything actually changed in 20 years? Where is AI on regrowing teeth?” @Johnny5CD.

There have been advances in this, with AI-designed ligands to increase enamel production and strengthen teeth. On the other side, we've seen AI in dentistry analyzing the mouth and its various elements. But I think getting these ameloblasts up and running will be really useful in repairing teeth.

But I don't think anyone's actually figured out how to crack regrowing them fully. There's a bit of a show.

Alex, you want to layer on top?

I feel like I have to take another bite at this question.

Ha-ha.

Ha-ha. So, there is a drug out of a spinoff from Kyoto University called TRH035 that is targeting tooth regrowth—honest-to-goodness tooth regrowth—with general availability by 2030. And I don't think there's that much AI involved with it.

Again, it's blocking a particular protein pathway that is normally associated with blocking tooth regrowth, so it's a double blocker—blocking the blocker for tooth regrowth. I think the primary focus in their clinical trials is infants who suffer from a disease that causes impaired tooth growth. But the plan is to get it out to general availability by the end of this decade.

Nice. I was going to layer on top, but you got it. All right, Dave.

Oh, God, there are so many good ones on this page. Okay. I'll take number 8. If you had 100× capability tonight, what would you actually work to solve?

So I would do exactly this. In fact, I will have 100× capability by the end of the week, so I'm going to use it to try and build algorithms that self-improve more efficiently and then try and get that flywheel accelerated.

And then I completely agree with Demis Hassabis. What we need to do next is turn all of that energy toward health and longevity until we get it solved, and then we have more time as a species, and then we expand out from there. So I would do it in exactly that order: self-improvement first, then health and longevity. Consume it all.

All right. Emad.

Let's see. “If Dario wants super-voting shares and control and ends up with a trust nobody elected, how does anyone actually get that power back?” @DavidHarmon03.

I think that's the point. You're not going to have a say in superintelligence. I think Anthropic thinks that's far too dangerous, and there is no good democratic way to do that under their rubric and kind of approach. So you have to assume that it will be a closely controlled company and ultimately comes down to a few people like Ben Bernanke to decide the future of the light cone, potentially.

Huh. All right, Alex, how about number 5?

Oh, really? You don't want me answering number 6?

I'll take number 6.

Oh, really? All right. Fine.

I'm steering—

5—

I'm steering the conversation.

Clearly. So 5 asks, “Are we worried that non-AI research and development gets starved of resources while everyone waits for AI to dominate?” This is from Brian Silver, 9652. No, not worried.

If anything, I think ultimately the self-licking ice cream cone of recursive self-improvement can only get us so far in terms of revenue-per-token maxing. My expectation is that it's going to be the non-AI R&D applications that ultimately dominate the economic gain.

You can only get so far improving AI for its own sake before, ultimately, you have to start driving real economic gains. With AI, if it just lives in a pure bottle and never interacts with the outside world, there's no real economic gain there. It has to start talking to the outside world at some point, and that's what non-AI R&D is. So, in short, no.

All right. And number 6: “When will an AI bot become a member of the Moonshots panel?” From @JohnsMusicalMusings. What makes you think, John, that we're not already AI bots on the podcast?

That was my answer, Peter.

Yes, I know it is. What makes you think Alex—

Silly.

—is a real human? I mean, listen—

Approximately a year ago.

Yeah.

Approximately a year ago. Yes.

Yeah, yeah.

For sure. And I think we will be playing with that very shortly, Johns. So one last music video for everybody. Let's enjoy this one, “Optimism to the Max” by Martin Parish.

All right, for all your outro video creators, again, send us at media@diamandis.com. We have to start including Emad into those videos. All right, gentlemen, I guess we're going to be recording in 48 hours from now. No time to sleep.

Yeah.

Good thing nothing ever happens.

Yeah.

I think we have that one docket already for that one, eh?

We do. We do. Amazing.

Do we really?

Yeah.

Thank you so, so much.

We have to check our NVIDIA results, man.

Ah.

Everything—

Love you guys. Be well.

Thanks, Peter.

Likewise.