Google向 Anthropic 投资400亿美元，GPT-5.5发布，Google Cloud占据主导 | EP #252

Google commits to a $40 billion investment in Anthropic. Dario needs compute, so he signs up with Amazon. Google is already a shareholder in Anthropic. They're trying to maximize the economic value per token.

It's all bottlenecked at TSMC. That's the actual bottleneck to all of AI, and only Elon will talk about it. Google Cloud is dominating. They unveiled their 8th generation of TPUs, in particular TPU 8T for training and TPU 8I for inference.

I still believe Google is the winner in the long run here. OpenAI unveiled GPT-5.5. It very much feels like a release that's intended to strengthen OpenAI's coffers. MATH is cooked. A bunch of other things are cooked as well.

Things are moving so quickly now that, on a month-by-month basis, we're able to see the hardest of these benchmarks creep up 1% per month. So, not long now.

Hey everybody, welcome to another episode of Moonshot. Your favorite AI exponential tech pod out there in the universe. I'm here with my incredible Moonshot mates, AWG, back with his orchid-filled room DB2 in his headquarters of all exponential investments. And, of course, Seem is on the road. I mean, you remember the book Where's Waldo? I think we're going to replace that with Where's Salim. So, Salim, where are you today?

I'm in a car in Guadalajara, Mexico, transiting to the airport, and this was the only way I could do this: in the car. So, hopefully, the friend's hotspot we're piggybacking off lasts. We'll see how it goes.

I can't believe you brought up Where's Waldo. Do you know, Peter, that we're still the exclusive licensee of Where's Waldo for data mining?

Okay. We used to go to trade shows and have an actor dressed up in that Where's Waldo suit. We'd say, “Hey, our neural nets can find anything in your data. It's like a Where's Waldo.” We gave out all the books and everything. It's amazing that you still remember that.

So, you're in Mexico. The Blitzy team is in Mexico, and they're raving about the podcast, by the way, so I guess we have a big fan base down there.

We do, it turns out. Big time.

I was at a conference of about 1,100 people, and quite a few of them are avid watchers.

That's awesome. What about the rest? Did you convert them?

Yeah. We've got to think internationally whenever we're commenting on these topics because everybody—it's a big world, and everybody out there is—

My Spanish is not quite up to snuff to say, “Everybody should watch Moonshots en español.”

You know, there are translators now.

I know.

I did my Meaning of Life session last night in Spanish with a translator, and you should have seen the translator at the end of the night. She was so fried.

And, of course, when you're touring through India, what do you speak? Hindi, or what?

No, I speak English. It's my native tongue because I come from a diplomatic family. I have pretty bad Hindi. I can get by, but my grammar is bad, my vocab is bad. I can get through about 50% of our conversation.

Well, we're at almost 500,000 subscribers. So, next time you're in front of our large audience, tell them to push us over to 500,000.

I do.

over to 500,000.

Okay, I'll tell them.

Let's jump in. Another incredible, crazy week.

Let's kick it off with a conversation around the AI race and the agentic boom. Check out this slide: 15 major releases in only 8 weeks. We're getting a pace of 2 major models per week. I think you've got to be retired and focusing only on this to keep up. There's no way otherwise.

In this segment, what I'd love to do, guys, is really hit on the last 3: Kimi K2.6, GPT-5.5, and DeepSeek V4. They're extraordinary releases, each of them hitting new capabilities. One thing we saw, Dave, was the acquisition—or the announced acquisition—of Cursor by xAI.

I think what's interesting is that the winners in this crazy model race are going to be those that are providing the best abstraction layer. It doesn't matter what model is underneath. Do you agree with that?

Dave Blakely

Yeah, totally. I just had a meeting with a data center company here in Cambridge, and the amount of effort going into the TPUs and the NVIDIA B100s and B300s is incredible. But at the abstraction layer, factors of 5 and 10 are just being thrown away by mismanagement of the context window.

There's so much opportunity in this stack, which makes sense because it's all brand-new. There's also a lot of vertical integration going on. The warfare is really stepping up.

I can't believe how Kimi K2.6 is keeping up. I mean, it is just shocking that the open-source world is actually on the radar and keeping up. We'll get to that in a minute.

What's interesting is the speed of these releases. I'm guessing that these new models are sort of competitive marketing, where the models are probably already cooked and they're just waiting for someone else to release, then releasing right on top of it.

Dave Blakely

Anthropic is holding back on Mythos. So, there's at least 1 case where you're exactly proven to be right, which means there may be others as well. It's funny—the dot releases are coming faster and faster and faster.

What's shocking about this list is that it's US versus China, right? There are no European models. There are no UK models, no Japanese or Indian models. It's just all US and China. Everyone else is a spectator, it looks like, at this point. I don't know if you agree with that, but—

Dave Blakely

Well, the models are definitely self-improving now.

Well, no, you're 100% right, but the models are self-improving now. So, the rate is accelerating—exactly what singularity theory would have predicted. The rate is accelerating, but because the models are improving themselves, it's hard to start from a cold start and catch up.

I'm surprised that other countries aren't using the Kimi K2.6 model to bootstrap their own internal research. Maybe they are, and it hasn't popped under the radar yet. But I'm not finding it too hard to design new neural nets using existing neural nets. It's a very doable thing.

And I'm curious, Alex, about that chart down below on this slide that's showing all the leapfrogging. It's leapfrogging all the time, but is it that they're all just cherry-picking? They're all just studying for the test on the particular benchmark and then releasing whatever the latest benchmark they're best at? Or is this truly—

Alex

Yeah. I think we're down in the West to a 3-way race at the frontier between OpenAI, Anthropic, and Google. I think those 3 labs have been pretty good about not benchmark-maxing or overfocusing on just 1 benchmark. They're pretty good generalist models.

I think we're seeing an honest-to-goodness arms race, horse race, or rat race, depending on which metaphor you prefer. My friends at the frontier labs often call it a rat race.

As to the Chinese models, it's interesting—the aphorism, “Why do you rob banks? Because that's where the money is.” To the earlier point about why there are no European models, where's Mistral in all of this, for example? It's because the US and China are where all the compute is.

Ultimately, I think OpenAI's Noam Brown, who of course is quite famous for having led their reasoning approach, has recently started almost pondering, with a bit of awe, whether the weights actually matter as much as they used to, or whether it's really turning into a race for compute, in some sense, as inference-time reasoning becomes more and more important.

His argument—not mine, but I think it's a credible one—is that the weights themselves start to become less important, in the same sense that, say, individual units within a transformer-style architecture become less important as the transformer itself starts to scale. The overall weights for an entire model may become less important as more and more reasoning gets used, and you see, in effect, a space-time transformer that's rolled out over time in reasoning-token space.

If that argument holds—and I think it's a pretty interesting one that I hadn't heard elsewhere before—that would almost suggest that, while at the same time we're seeing a race to the bottom on, say, per-token intelligence densities between American models and Chinese models, in open source, the American models are still about 6 months ahead. This has been pretty consistent for the past couple of years.

In closed source, it may not matter in the end. What may matter in the end, at least according to the scaling laws we have at the moment, is who has more compute at the end of the day to do more reasoning.

We're going to see that in a minute. But 15 models over the course of 2 months is insane. Some of these are just improvements on existing models, and some of these are completely new pre-trained models. I think that difference needs to be pointed out.

Salim Ismail, any thoughts on this insanity? The fact that we have that many releases in 8 weeks kind of blows my mind.

We're watching the cost of cognition, coordination, and execution all collapse at the same time. I think it's not so much the breakthroughs; it's the compression density that's crazy.

Alex

Well, the capabilities are mind-blowing. These are not just fake little dot releases that are benchmark-maxing. If you use them firsthand, what’s really helpful is to look at our podcast or at any postings on the internet from 3 months ago, 6 months ago, 9 months ago, and 12 months ago and look at the predictions of capabilities. We’re so far ahead of what even the upper bound of predictions would be in terms of the capabilities as the parameter count grows.

If you extrapolate from there, we’re just on this steep knee curve of the acceleration in the singularity.

And raw parameter count and more chain-of-thought reasoning are just going to push us to limits that are way beyond human.

Alex

So, what does the average person care about?

What does the average person care about this? Right? One of my boys says, “Okay, great. A new release with new numbers over and over and over again.”

At the end of the day, stuff is getting better, cheaper, and faster. What does the average user care about? Do you recommend someone sticking with a particular model? Should they say, “I’m just going to use OpenAI. I’m just going to use Anthropic. I’m just going to use Google”? Any thoughts there?

Alex Isenstein

I think the question itself is a red herring. Why? OpenAI bet the company on consumers using all these reasoning tokens. That consumer-oriented strategy for all these billions of dollars of capex that they’re building out would work, and they’ve had to pivot rather prominently in the past few months back to enterprise.

So I think the question of what the average user cares about—which I construe as what the average consumer cares about—I almost think the market is telling us that the average consumer, in the short term, isn’t even part of the equation anymore. This really should be: What does the average enterprise care about? Because they’re the ones—

—asking for—I’m asking for our listeners, right? A lot of them are entrepreneurs or general consumers. At the end of the day, is it okay for someone to say, “I’m just using ChatGPT. I’m just using Gemini 3.1 Pro. I’m just using the latest version of Anthropic’s models”?

Is it important for people to be driving to the latest model, or is it okay? Ultimately, everybody’s basically leapfrogging everybody else. If you’re a mom, a dad, a student, or maybe an entrepreneur just getting going, this insanity of 15 models in 8 weeks, bouncing back and forth—Dave, you’re using 2, 3, or 4 models all the time, right?

Dave Blakely

Many more now. That’s the biggest change. The coordinator model can now manage dozens or hundreds of other models successfully, and 6 months ago—or 3 months ago—that wasn’t true.

For the average consumer, the ability for this stuff to install itself is a massive change. You can go into the model now, and you have to use the latest ones. It doesn’t matter a lot whether you’re using Claude 4.7 or GPT-5.5. Just use one of the latest ones, but ask it to install itself. Ask it to build something on your laptop for you, and it just works.

You don’t have to understand the Linux command line. You don’t have to understand any of the underlying infrastructure. It’s smart enough now to explain itself to you as it goes. So I think for the average listener, that’s a massive unlock. Someone who’s never built software before can just think of something and create it in an hour. That just wasn’t true 6 months ago.

Yeah, let’s jump into our first model story here, which is Moonshot AI launching Kimi K2.6. I just downloaded it onto my Mac Studios this weekend, on top of Skippy, which is orchestrated by Opus 4.6.

Kimi K2.6 is a 1-trillion-parameter, open-weight, open-source model that activates 32 billion of the parameters at a time. It runs 300 parallel agents. Very importantly, it can natively process text, images, and video all at the same time, and it costs 30 times less than the most capable closed models.

Interestingly enough, Moonshot AI didn’t get its name from us. The 3 founders are based in Beijing, and their favorite album is Dark Side of the Moon, so that’s where it came from. The company is backed by about $4.7 billion in capital from Alibaba, Tencent, and IDG.

If you look at the numbers at the bottom, on the benchmarks compared to GPT-5.4, Opus 4.6, and Gemini 3.1 Pro, it does amazingly well against all those models. This one was trained, they report, for a total of $4.66 million, compared to hundreds of millions or billions on the other closed-source models. Dave, I find that amazing. I mean, almost—

Dave Blakely

Almost incredible. There’s so much to say about this, starting from the fact that Alex said a minute ago that the Chinese models are running about 6 months behind the US models. But if you look at the benchmarks, this is up there with or beating Claude Opus 4.6, which came out only 3 months ago, in February. So that’s not a 6-month lead; that’s a 3-month lead.

On the price-performance side, most people, when they first start, don’t care too much. It’s cheap. All AIs are pretty cheap. But then when you realize that you can run 10 or 100 of them concurrently, you’re like, “Well, this is going to start to add up.”

If you run this on Fireworks AI, it’s about 1/8 the cost of running the Claude API or the OpenAI API. So 1/8 is a pretty damn big price cut. If you download it and run it like you did with Skippy, then you’re running it at about 1/30th the cost. That’s a big deal.

The caveat, of course, is that, as Alex has pointed out many times, you’re not 100% sure it isn’t spying or doing code injection. It’s probably not, but you can’t guarantee that. Somebody tells me this is 1/30th the price, and I’m like, “Try it.” Then you’re thinking, “I’m a little suspicious. Why is it 1/30th the price?”

I doubt it’s code-injecting on you, but you can’t be sure. Whereas, if you use Anthropic or OpenAI, it’s definitely not code-injecting on you. In fact, it’s safeguarded all over the place. So there’s your landscape: chaotic as always. It’s only going to get more chaotic.

Alex, how big a deal is Kimi K2.6?

Alex

I think it’s helpful for certain enterprise use cases where you want to be able to self-host the model and you don’t want to use AWS Bedrock, which now hosts GPT-5.5 in addition to the Opus models. I think it’s helpful in that respect. It’s helpful if you want to be able to self-host fine-tuned models for yourself. The same goes for DeepSeek V4.

I think in general, again, it’s a few months behind for other use cases. For consumers that want to be able to self-host for whatever reason—privacy or otherwise—it’s probably very helpful for folks who want to self-host their own Claude. Very helpful.

I think there are many use cases where these typically Chinese open-weight models, like Kimi K2.6 and DeepSeek V4, are very helpful. I do think, however, that they’re not at the frontier. To me, the big headline is that the disparity between the American frontier closed-weight models and the Chinese frontier open-weight models seems, at least for the moment, to be in place.

Dave Blakely

Yeah, Peter, I think your setup is perfect. It’s exactly what I do, too. I use Opus 4.7 as my orchestrator because you want that extra notch of intelligence. Then, if you have simple tasks or subtasks, you can farm them out and save money using Kimi K2.6.

If the results coming back don’t make perfect sense, your orchestrator will tell you, “Hey, this is garbage.” So you can actually rely on Opus 4.7 to give you the straight truth on what the underlying models did for you. It’s exactly the way you set it up, Peter.

Yeah, Dave, a week ago you said you moved from 4.7 back to 4.6. Did you move back to 4.7?

Dave Blakely

I have both running now. 4.7 is kind of wordy and sounds kind of PhD-ish, which annoys me sometimes. 4.6 is friendlier, but it’s clear that 4.7 is a little smarter.

Sometimes you just need the right answer, no matter what. So I actually have both running in parallel in agent windows now.

I’m curious—in Guadalajara, Mexico City, parts of South America, and parts of Asia—what are you hearing about the use of US models versus open-weight, open-source Chinese models?

Guest

I get a mixture of both. A bunch of people use the hosted models, the big ones, just because it’s easy. There’s a subset of people that use the open-source models and the Chinese models, and they don’t really care.

I think they should care. At some point, that’s going to come up. One question I have for Alex and Dave is: How do you protect against code or prompt injection in these open-source models? Is there a way of defending against that?

If there is, then there’s a huge case for this, because everybody here is looking for the low-cost approach. But for the most part, I’ll be blunt: The conversation isn’t around which model, open-source or closed-source. It’s “What do we do with AI?” That’s literally at a level of lack of sophistication that you would expect.

But the opportunity is also there for startups to leapfrog lots of people and build aggressively for the coming madness that's upon us. It's almost an impossible question, it seems, because if you sit on the sideline and don't use this stuff aggressively, you fall way, way behind.

Alex Isenstein

Yeah.

But if you start using it aggressively, you're generating thousands or millions of lines of code before you even know it.

Alex Isenstein

And so the odds go up, right? I think what you're trusting right now is that the guardrails that Anthropic and OpenAI put on their models are very, very cautious when they're pulling in code, open source or otherwise. I mean, they're almost annoyingly cautious, so you kind of assume they've done a very, very good job of filtering out nasty code injection. But the numbers work against you at scale, so there's no simple answer.

When I got into it, I was like, "Hey, I'm just going to look at the code. I'm not going to just run it. I'm going to see what it does." That's a joke, right? That's just laughable. It's generating so quickly now that there's no chance you could even scroll through it. So you have to use AI to protect against AI. There's no other way to get the scale. So it's tricky. I know that wasn't much of an answer, but it's tricky.

One thing I'd love to point out here—we talk about this on occasion, but I don't think we've ever really spoken about it in detail. Kimi K2.6 uses something called a mixture of experts, or MoE. If, in fact, you have a trillion-parameter model and ask a question, it's basically accessing all trillion parameters every time to analyze every token.

What they did here is create a set of 30-plus experts, so some percentage of all the parameters are dedicated to one expert system. If you ask a coding question, the orchestrator looks at it and says, "Okay, this is a coding question. We're going to send it to experts number 3, 7, and 12," and it only uses a portion of the parameters. Instead of all the experts, it uses some subfraction thereof, which saves money and time. How many different models are using that right now, Alex?

Alex Isenstein

Sparsity, which is the term of art I think we're talking about here, is endemic to all frontier models at this point. It's also the basis of the human brain. If we look at the brain, most neurons don't, at any given point in time, have action potentials going in and out.

Sparsity is a great way to reduce the memory footprint of models. To my knowledge, all of the frontier models use sparsity one way or another. It's also a good way to regularize the models, to make sure that particular weights or parameters in the models aren't overfitting to the training data. One of the age-old techniques is just blasting away individual weights or parameters in the neurons, making them disappear entirely as a so-called regularization technique.

Sparsity is everywhere at this point, and I think it's only going to become more important with time. One of my holy grails, as I've mentioned on the pod previously, is to see a million-parameter-or-smaller diamond or black hole of a model at the end of the scaling race. I think sparsity, and cranking the knob on increasing sparsification in these models, is one possible path to getting us there.

David Friedberg

And just to add something to what Peter said, the mixture-of-experts innovation that came from DeepSeek is actually layer by layer. Most of these neural nets are about 140 layers deep now, and it'll route the expert layer by layer. It'll say, "Look, within this layer, I'm just doing basic image classification; in this layer, I'm doing deeper thinking; and in this layer, I'm doing higher-level math."

As it moves through the neural net, it'll actually route to, I think, up to 128 different experts layer by layer. It'll find the optimal pathway through the entire neural net. On top of that, you can also have dedicated experts—here's a surgeon, here's an artist, here's a coder—above and beyond that, but this is actually within the neural net, layer by layer.

All right, next story: OpenAI unveiled GPT-5.5, literally just 7 weeks after GPT-5.4. Greg Brockman calls it a new class of intelligence. It's natively omni-modal, able to process text, audio, video, and images all in a single, unified, end-to-end architecture.

It has a 37-point increase for 5.5 over 5.4 in long-context reasoning. Both 5.4 and 5.5 have million-token windows, but 5.5 can actually remember the beginning of the million tokens and provide complete context across the entire thing. Token efficiency is up, with 40% fewer tokens at the same latency. And I love this: hallucination is down 60% over 5.4. Let's go to our resident genius, Alex. What do you make of 5.5? How important is it?

Alex

I think it's very important, both intrinsically and relative to 5.4. I want to highlight 2 key stats here. The first is the leap from GPT-5.4 Thinking to GPT-5.5 Thinking. That's probably the biggest leap overall on Terminal-Bench 2.0 specifically.

One way to interpret this is that Terminal-Bench is a benchmark focused on the ability to agentically operate from a command-line terminal. That's useful for Codex and Claude Code-type environments. One way to construe this huge leap, which is larger than most or all of the other leaps that we see in terms of other benchmarks, is that 5.5 is being very seriously benchmarked, if you like.

Although, having used it, I really don't think it's narrowly overfitting just to making Codex a better Claude Code competitor. It very much feels like a release that's intended to strengthen OpenAI's Codex. That's thought 1.

Thought 2 is that my favorite benchmark among all of these is FrontierMath Tier 4. FrontierMath Tier 4, which I think we even had a New Year's bet about that we're going to have to revisit sometime later this year, is one of the best proxies for the ability of AIs to solve professional-level research problems in math.

What do we see? We see approximately a 2% leap from GPT-5.4 Pro to 5.5 Pro in approximately the last 2 months. What does that tell me? It tells me that we're seeing approximately 1% gains per month in research-level math coming from frontier AIs, and we're getting closer to approximately half of all the FrontierMath Tier 4 problems being solved.

You can extrapolate this and realize that if the present rate just stays the same—which I guarantee it won't; it's going to accelerate—even at the present pace, we're talking about essentially all FrontierMath Tier 4, all professional research-grade math problems, being solved in the next 4 or 5 years. So math is cooked. I'll say it a second time: math is cooked. A bunch of other things are cooked as well.

Things are moving so quickly now that on a month-by-month basis, we're able to see the hardest of these benchmarks creep up 1% per month. So, not long now.

It's worth pointing out that the API pricing on 5.5 is twice that of 5.4. It's $5 per million input tokens versus $2.50 on 5.4, and $30 per million output tokens versus $15. I like the simplicity of that pricing. Dave, have you been playing with this at all?

Dave

Yeah, absolutely. I think what Alex said earlier in the pod is really, really important and insightful. Like Noam Brown is saying, "Wow, maybe the weights don't matter so much as this chain-of-thought process is just way ahead of any expectations on how intelligent it can get."

From a user's point of view, that first benchmark—if you ask it to do something complicated, like configure an entire system for you, download some software, integrate it, make it all work, connect it to my Outlook, connect it to my whatever—it just works. That exactly ties to the first benchmark. It just flat-out works. It feels like this incredibly capable, brilliant assistant, no matter what you're trying to do, because of that first benchmark.

Then that last benchmark—the frontier, or second-to-last, FrontierMath—Demis Hassabis came out and said, "Yeah, I think it's kind of a coin flip." This was on Alex's Inner-Most Loop. It's kind of a coin flip now on whether just the existing architecture scaled up solves everything.

Alex

Yeah.

I think calling it a coin flip means he's moved a long way.

Alex

He has. We need new breakthroughs.

We're out of breakthroughs, apparently. I remember 10-plus years ago, when I was chatting with Demis, he used to say there were 5 breakthroughs remaining between where we were then and AGI as he construed it. Now we're out of them. It's half a breakthrough or 0 breakthroughs at this point.

I'm curious—we should ask him: What does he think is required to get to true AGI or ASI? Are we just going to extrapolate what we're doing, or do we need breakthroughs? I think that requirement's been falling, more or less. It's starting to feel a lot like—actually, Alex, you know what would be great for that? To put together a chart of Demis's number of breakthroughs, because at Davos it was down to 2.

Now it's down to 50/50 that it's zero, but you mentioned 5. That was maybe a year and a half ago. That would be a really cool chart.

Alex

The number 5 from him was when I was chatting with him—this was 10-plus years ago.

Yeah. Okay. Well, there's some sort of exponential decay of breakthroughs, clearly. Alex, you said it a little bit earlier: this is ultimately a compute race. So let's talk about that.

Google Cloud is dominating. What do we see? We see Google announcing at Google Cloud Next 2026, their major conference, that they unveiled their 8th generation of TPUs—in particular, TPU 8T for training and TPU 8i for inference. Right now, we have training and inference chips separately, just like Amazon has its Trainium chips for training and its Inferentia chips for inference.

These new TPUs are 3 times faster in training performance and have 80% better performance per dollar. They're designed to run millions of agents in real time, so Google is really all in on the agentic era. Sundar Pichai, the CEO, whom I had a chance to spend some time with last weekend, made it crystal clear. He says over 16 billion tokens per minute are being processed, and 75% of Google's code is now written by AI. So, fascinating. Dave, what do you make of this?

David Friedberg

Yeah, you know, what's surprising to me is that the price performance of the TPUs is landing right on top of NVIDIA—not much different at all—which is surprising because it's a completely different architecture. It uses a systolic array design. I mean, it could not be more different from a GPU under the covers, but for whatever reason, it's all kind of canceling out and landing identical, which is fine from Google's point of view because now they have their own total chip-fab-through-data-center-through-model solution.

They do everything.

David Friedberg

Yeah.

I'm still—I still believe Google's the winner in the long run here across the board. I don't know if you agree with that.

Alex

Terafab is a big thing.

That's true. I'm sorry. Yes. Okay. I'm thinking in the OpenAI and Anthropic ecosystem. Yeah, Terafab—

David Friedberg

On the other hand, Google owns a material percentage of SpaceX.

They do. They do. I don't know if you saw—there was a tweet out recently about Google's investments that were made. They just had massive returns on their investments in SpaceX, in Anthropic, across the board. Huge returns.

David Friedberg

Yeah. I was at a board meeting yesterday for a company that I'm the chairman of that has massive cash flow and a huge cash balance. They were like, "Well, I don't know if a public company can really do seed-stage investments." I was like, "Have you looked at Google?"

Yeah.

David Friedberg

They have multiple hundred-million-dollar gains on their investments. And they don't even do it for the money. They do it for the knowledge and for—

Well, and strategic relationships, right? Larry and Sergey were just bonding with Elon, and they said, "Okay, Google is going to invest $1 billion," and now it's worth, you know, God knows how many hundreds of times more.

David Friedberg

Yeah.

It's also just investing in the future. I remember conversations with Larry and Sergey about the nature of the frontier, and I think, to their credit, they're investing in the frontier. SpaceX is part of it, and also compute.

Epic put out this really eye-opening stat in the past week that Google now accounts for approximately a quarter of all the AI compute on the planet, and I'm sure 8th-generation TPUs will be part of it. I think it's also worth keeping in mind that the TPUs at this point are being designed by AI. I have a number of friends at Google who are responsible for designing next-generation TPUs, and they're all just using Google AI to do it. The recursive self-improvement goes all the way down to the silicon at this point.

All right. Our next story in the Google ecosystem, also announced at their large Cloud Next conference, is that Google commits to 960,000 NVIDIA Vera Rubin GPUs for its A5X. Pretty extraordinary. A5X is Google's new bare-metal virtual machine instance, delivering 10x lower inference costs and 10x higher token throughput.

Just an interesting FYI: Vera Rubin, for whom these chips are named, was an American astronomer who discovered the first conclusive evidence of dark matter. I love the fact that Jensen is naming chips and systems after famous individuals.

Now, what I find fascinating—this goes back to the conversation a minute ago—is that this cloud is 2 times bigger than Colossus 2 and 2.4 times bigger than Stargate Abene. So Google is winning, at least based on what they're building and plan to build. Again, Dave, thoughts here?

Dave

Well, you know, part of the acceleration we're seeing in society as a whole is that all the really, really smart people are working on real tech now—

Hard hardware.

Hardware and space and medicine—real tech.

And, you know, if you go back to the Web2 era, the Facebook era, the rewards were all in either cheesy consumer experiences or banking. Doing deep tech was kind of a way to die poor. So it's creating a whole new era for society.

The post-AI era—we all knew it was going to be very, very different. But now the rewards are in actual deep tech that benefits humanity in really big, fundamental ways.

If you just counted the number of people that you know who have been pulled into this vortex, it would have been just a few percent working on world-changing, deep-tech, real stuff 15, 20, or 30 years ago. Now it's almost everybody that you know getting pulled into doing something big and world-changing, and it's actually working. So that's a big change for society. That's helping accelerate things as well.

All right. Our next story is that Anthropic is cutting deals for cash and compute. I mean, a huge amount of capital is flying back and forth between the frontier labs and the hyperscalers here.

Google commits to a $40 billion investment in Anthropic. Last week, Google committed to a ton of money: $10 billion in cash right now at a $350 billion valuation. And, as we talked about last time, Anthropic on the secondary markets is now at a $1 trillion valuation. So this $350 billion is coming in at roughly one-third the cost of what others are paying for it.

They committed to another $30 billion if Anthropic hits certain performance targets as well. They're going to be providing 5 gigawatts of TPU compute committed over 5 years. That's the equivalent of providing power to 3 to 4 million people.

I'm finding this pretty extraordinary. We're going to see in a moment a conversation where Anthropic has cut deals with Amazon in a similar fashion. Actually, let me go ahead and hit that, and we'll talk about this cash-for-compute conversation that's going on.

Amazon and Anthropic are trading cash for compute. Here's a second deal: Amazon is investing a total of $33 billion. They've committed to $25 billion on top of the $8 billion they've already invested.

In return for Amazon's cash, Anthropic is committing to spend $100 billion or more on AWS over the next decade. Anthropic will run Claude on Amazon's custom Trainium chips, and Amazon will provide 5 gigawatts of AI compute capacity for Anthropic.

We're seeing Anthropic becoming beholden to both AWS and Google in a significant fashion. Gentlemen, thoughts on this one?

Well, it's so funny to me. Obviously, Anthropic needs much, much more compute and is growing. Actually, a very good friend of ours—Peter, I won't mention him on the podcast, but he was telling me at the board meeting yesterday—he can figure it out from that comment—that Anthropic under the covers is thinking they might hit between $40 billion and $50 billion, up to $70 billion in revenue by the end of the year.

We talked about $100 billion by the end of the year a few pods ago, but still, I mean, a billion last month—doubling, tripling.

Dave

It's extraordinary. And the only reason they wouldn't hit those numbers is because they can't get enough compute to keep up with the demand.

And one of the things was that they didn't release Mythos because they don't have enough compute to deal with it, right? So it's a limited release of the capabilities.

Dave

Yeah. And OpenAI cut Sora. I think one of the reasons is probably compute.

Which is energy. And I think that it's so funny to me to see all these deals. Okay, so Dario needs compute. He signs up with Amazon. Google's already a shareholder in Anthropic. They're all—and now OpenAI is going to be running on GCP and also on Amazon Bedrock.

So you can get it through Bedrock. Everybody's partnering with everybody else, but—

Dave

It's all bottlenecked at TSMC. This is all great. You can all partner with each other up the yin-yang, but whose chips are actually going to get made?

You don't see TSMC in any of these podcasts, in any of these deals, in any of these meetings. And you saw Jensen actually recently say he doesn't have any long-term agreement with TSMC; they just kind of make it up as they go.

So all of this is bottlenecked, and only Elon is talking about it. Look, the fundamental constraint to all of this is the Terafab, and I already locked up $16 billion—could be $45 billion—of Samsung's capacity. The only 3 companies in the world capable of making any of this are Samsung, Intel, and TSMC. That's the actual bottleneck to all of AI, and only Elon will talk about it.

Alex, is it compute or is it energy at the end of the day, right now?

Alex

I think they're indistinguishable at this point. I think permitting for on-site energy is a major limiting factor. I think it's probably, on balance, more of a limiting factor at this point—maybe not a year from now—than TSMC, but it is a limiting factor.

Having powered land and data centers that you can take all of these—infamously, Microsoft even in the past few months spoke about having lots of GPUs that they'd love to rack-mount in a data center, but lacking the powered land and lacking the data centers to plug them in. I think at this moment, energy, at least in the U.S., is the bigger constraint. But I agree with Dave that, in the medium to long term, semiconductor fabrication supply chains—doubly so if there's any geopolitical conflict—are likelier to be a stranglehold once we solve our energy story.

So let's talk about the not-investment-advice segment here. Where do you invest your capital? If compute and energy—I mean, I'm seeing energy stocks beginning to fly, right? A friend of mine just had this IPO of X-energy, and it popped like 30% on the first day. We're seeing Bloom Energy and other energy stocks beginning to skyrocket, creeping up over time.

So, I don't know, if you're going to invest in chips, do you invest? We saw Intel pop up and AMD—I mean, all of these guys, that entire ecosystem of chips and energy. Ultimately, if they're really the constraining part of the innermost loop here, I think the most demand is there. Any thoughts, Dave?

Oh, so many thoughts. I could go for an hour on just this topic. Invest like crazy in anybody who has access to chips and can find a power supply. That's pretty straightforward.

There are power supplies everywhere. All these legacy manufacturing operations—aluminum melting and all that—use a huge amount of electricity, and swapping them over to data centers is a massive increase in the value of that energy supply. But you have to have a line on the chips.

Then, at the kernel level, because the chips are so constrained and the demand is through the roof, anyone who's writing software at the kernel level that empowers AMD chips or legacy GPUs to participate, or just makes inference more efficient on NVIDIA chips, those companies are worth a fortune. Anyone who's building kernel-level software is a brilliant investment.

And then, in the vertical use cases, Anthropic rolled out something called Skills, which you should absolutely play with. It's just a way to use the context window more efficiently by designing skills that the AI can then pull in. Rather than having to reinvent everything every time, just build a skill.

And then you can call on the skill very efficiently. Companies are now discovering they can refactor their entire business—or their entire whatever they do—around 100 or 1,000 different defined skills, but those skills then become the defensible intellectual property—

—within that vertical domain.

So any vertical domain where you're racing to build out the entire skill database for that use case is also an unstoppable investment theme right now.

I could go forever. The other thing that's interesting is that both Google and Amazon are getting their shares in Anthropic at one-third the going rate. I find that extraordinary: a $350 billion valuation versus the $1 trillion valuation.

Well, it shows you how important the compute is. Again, you're going to be sold out forever if you can get the compute. These hyperscalers are kind of hedging their bets, right? They're not picking a winner; they're buying every horse in the race.

This AGI-ASI race is just way too important to lose, so they're investing left, right, and center.

Alex

I would also just parse these as the market doing what the market does. Some of the participants—some of the frontier labs, like Anthropic—have an insatiable hunger for compute, and they have the revenue generation to generate the demand and sustain the demand. If you're Anthropic, you're going to go to every possible source of compute at scale that you can find, whether it's Amazon, Google, or other sources. You're just going to seek out whatever the market will provide, as a hungry customer for compute.

I don't think necessarily the story needs to be any more complicated than that. It turns out the world demands a lot of compute to solve some of these really interesting problems in code generation and otherwise. What we're going to see over time is all of this demand translating into supply.

It's going to translate in the short term into what looks superficially like a bit of a circular economy between, call it, the top 10 or 12 companies after we see the IPOs of SpaceX, OpenAI, and Anthropic. But that's going to diffuse throughout the economy over the next few years, would be my prediction.

People are hungry for compute. Salim was hungry for bandwidth. Salim, welcome back. I see you're in a stationary spot now, rather than at the airport.

Seth Godin

Dude, I'm just going to call you Waldo from now on. All right.

Let's move on. A couple of fun stories. I'm going to add this segment every time for the podcast: What did Claude just kill?

This is the stock chart for eBay, and this comes out of Anthropic's new research project, Project Deal. They created a marketplace for employees in their San Francisco office with one big twist: They tasked Claude with buying, selling, and negotiating on their colleagues' behalf—basically doing what eBay does. Then we see a drop in the stock price.

I don't think eBay has really dropped anywhere beyond this, but I think this is going to be more and more common. Any thoughts, Dave?

David Friedberg

I think a lot of this is just an immediate, knee-jerk fear reaction, but then things kind of settle out and you realize, wait, Anthropic is going to build all kinds of marketplaces because they can, but it's not going to hurt eBay.

I think what you're going to see more and more is AI growing so quickly that it's going to largely grow around the legacy economy—around the banks, around the insurance guys. It's going to be its own world, and it's going to be feeding on itself and building colossally large constructs that some people are not even aware of. It'll all happen very, very quickly.

So I think eBay will be fine.

Any thoughts here?

Alex Salkever

I have a slightly different take. There are so many places where lots of problems in companies exist because coordination is hard, and AI makes coordination easy. That's going to threaten big chunks of places: marketplaces, customer support, listing optimization, dispute handling.

There are huge categories of these that will become agentic workflows. I think the bigger question about “What did the AI just kill?” is: What workflow category did it just encompass and automate?

You want to hear something cool related to this? The data center CEO that I met with this morning—we were talking about data centers going into space because power is basically free. Solar is basically free in space.

He said, “Data centers—there's power all over the planet that's not tapped, that doesn't disrupt society at all. That's not why data centers are going to space. Data centers are going to space because there's no regulatory authority preventing it.”

You try to do anything—I mean, that's not true. You still have to, if you're going to be flying all these data centers and communicating, get licensing domestically and through the ITU for bandwidth. There are going to be regulatory hurdles that Elon and Google need to clear, especially if you're launching 500,000 satellites.

When you're putting up a debris field like that, there's going to be pushback. There is going to be pushback.

Alex

Yeah, it's interesting. I'll square the circle.

Doing anything on land. Sorry—[laughter]—go ahead, Alex.

Alex Salkever

I'll square the circle here and say I think, in the short term, for sun-synchronous orbit, that requires FCC and other approvals. In the long term, if we start to launch AI data centers from the Moon—and we're building them on the Moon—that will probably require fewer approvals, at least under the current regulatory regime.

That's 20 years away to actually get manufacturing on the Moon.

Alex Salkever

I'm talking about 20 years away, Peter.

You know, to get—listen, if you look at it deeply, I know 20 years away is infinity, I get that, but we're talking about—I mean, just to be clear, the stuff I'm concerned about is the next 5 years, right?

If you're launching—we talked to Elon about this—500,000 V3 satellites in a constellation, there are going to be debris issues. Elon pushed it off by saying, “Oh, we'll have superintelligence to figure that out.”

I just think we have this tendency where everything looks amazing from far away, but the reality is, by the time it comes closer, there are real issues. It's not going to be just the promised land of going to space. We're going to have challenges going there still.

Alex Salkever

Yeah, it's interesting how the timelines line up, too, because between here and there, there's all kinds of constraints. But between here and there, we'll have solved all math, and we'll have discovered all kinds of new physics.

Listen, I’m the space cadet. I’m the super space enthusiast here, and I can hope for nothing more than that vision to happen. But it’s always easier in the promised land. Peter, I’m gobsmacked to hear that you think it’s going to be 20 years before we have fabs on the Moon. My goodness.

Fabs on the Moon—manufacturing and pumping into Earth orbit with mass drivers.

Alex Salkever

You think that’s 20 years away?

Well, okay. Maybe 15, but it’s not the next 5 years.

Alex Salkever

Do I hear 10?

It’s hard for me. I guess Optimus robots will improve that. Demand will improve that, but the concern is if you have a collision of spacecraft in orbit generating debris, we still don’t have any mechanism for removing debris from orbit. It’s going to be a challenge.

Alex

Briefly, your concern is that Kessler syndrome is going to sabotage—

Is going to sabotage moon-based fabs? No, it’s going to sabotage the next 5 years of 500,000 satellites in Earth orbit. I mean, right now we have 10,000 satellites from Starlink, right, which is the most ever pumped into orbit. And we’re talking about 50 times that.

We’re talking about not just the US. Amazon’s going to do their best, right? Jeff is not going to stand still while Elon is doing this. And then you’ve got Chinese constellations. So, do you double or triple that number of satellites in orbit? I mean, listen, I can’t wait, and it’s going to have challenges.

Alex Salkever

Yeah, I’ll give a couple of thoughts here, with just a finger in the air. I think humanoid robots are 5 to 7 years away, minimum, at mass scale—in widespread adoption. Okay, minimum.

I don’t agree with that, but that’s okay.

Alex Salkever

I agree. I understand.

Oh my goodness. This is lunacy. Utter lunacy here. Well, we are in the Moonshots podcast.

Alex

Yes. Can we get back, just maybe, to Project Vend and Anthropic? I think we’re missing an important point. Everyone who wrings their hands over the latest Anthropic project purportedly sabotaging or killing some SaaS company—Anthropic doesn’t want to be triggering SaaS apocalypses left and right. There’s relatively little economic motivation there.

I think if you look at the throughline through all of these Anthropic projects or research projects, other than the alignment ones, all of their projects—and their corporate strategies and unhobblings—can be explained by a very simple principle: they’re trying to maximize the economic value per token.

That’s all that they’re trying to do. Claude Code, it turns out, through code generation, is actually quite economically valuable per token. Per token, it’s more valuable to generate useful, working code than, say, to generate video or cat images or whatever other consumer plays OpenAI and some other frontier model providers were chasing.

They’ve dropped that now. Everyone’s focusing on code generation because, on a per-token basis, it’s so economically valuable. So I would look at projects like Project Vend—running a marketplace, running a business—as Anthropic looking for new ways to increase the per-token economic value of their output. It’s as simple as that.

Brilliant, Alex. That’s absolutely brilliant.

To move us along here, we’re coming into the battle season. It’s Elon versus Sam and OpenAI. This just got posted today. Today is the start of a very important day in the AI world: the trial between Elon and Sam and OpenAI begins in the Oakland Federal Court. Jury selection is happening right now.

So I just put this up to keep us posted. We’ll be learning a lot. Of course, discovery is unveiling a lot of texts, a lot of emails that I bet both Elon and Sam and a lot of other people would rather not have aired in public. Any thoughts here, Jaan?

Jent

I think it’s sort of sad that it’s come to this. I just remember the Bill Gates versus Steve Jobs docudramas that were made from the critical Apple-versus-Microsoft era. This has, I think, a similar feel to it.

It’s sort of sad that this ended up in court versus settling earlier on, but I do think history will probably view this as an iconic struggle that will get the full Aaron Sorkin, if not similar, movie treatment. This will be the full Hollywood-type, totally titanic battle.

Yeah. Dave, any thoughts here? How’s this playing in Guadalajara?

David Friedberg

No recognition or awareness at all, and that’s probably a good thing. This is kind of a soap opera. I’m with Alex on this one. It’s just heavy drama. We wish it hadn’t come to this.

It would have been great to get these guys to settle the thing, but their positions are hard-baked in, and so it’s going to come.

How do you unravel the movement of OpenAI to a for-profit company? I mean, do you back it up to a nonprofit? And what about all the capital invested in OpenAI? Does that disappear if they lose the case here?

Alex

I’d like to do it more. I mean, I’ve said, I think, on the pod in the past that if the model of changing large nonprofits to public benefit corporations can be scaled, I’d love to do this to a number of major American research universities.

My question is, what happens to all the capital invested—literally hundreds of billions, $122 billion in the last couple of months? There’s so much pressure for this court case not to be won by Elon.

Alex

Well, if you’re following the detailed tick-tock of the way that this trial is being structured, it’s being structured in 2 phases. The first phase is more about deciding whether the claims that Elon et al. have made are in fact the case, and the second is the equivalent of an award-type section, deciding what awards, if any, to make conditioned on the first phase.

But I think there are a number of details in this court case that are notable. One is jury selection. There has been public reporting that already-selected members of the jury are aware of entanglements that Elon has had with the present administration and may view him negatively as a result.

I think the fact that jury members are being selected, reportedly, with some political influence seeping in—I think that’s very interesting. I also think it’s interesting that the district judge in this case has, again reportedly, decided that she’s going to take the jury outcome as an advisory opinion, but that if there is an award, she’s going to decide ultimately from the bench on the final award. So, there are a lot of nuances here.

Wow. Dave, any thoughts or opinions here?

Dave

Yeah. Do we ever figure out if we get to see it live? It’s not being broadcast, but I’m sure there are going to be court reporters giving us a lot of details here.

You can wait for the full Hollywood treatment in a couple of years.

Alex

By the way, there will be a Hollywood treatment of this, as with every other major—of course, it may be an AI-generated feature film, but nonetheless—

Dave

It will be for sure.

Well, I’m surprised how many personal texts have already come out.

Dave

Yeah. You know, the emails get discovered right away, and all your email gets thrown out there for the world to read, which is crazy, but it happens. Texts traditionally have not been thrown out, but yet we’re seeing them all. So, I don’t know exactly how that’s happening.

But for Elon to win, he doesn’t have to win the case. He just has to slow down OpenAI. I mean, in the middle of the singularity, if you lose 3 months, you’re basically—you lost.

You’re correct.

Dave

Correct. Yeah.

All right. Another fun topic. A few stories here. It’s about AI surveillance and privacy. Let’s check this out.

OpenAI’s Chronicle uses agents to build memories from screenshots. Sam Altman described this one as telepathy-like. Chronicle runs on OpenAI’s Codex, where background agents are taking periodic snapshots of everything on your screen. The screenshots are sent to OpenAI’s servers for processing. Agents use optical character recognition and visual analysis to extract the context of what you’re doing every minute on your screen. Structured memory files are created and stored locally.

We talked about this before: AI monitoring everything. Ultimately, it’s sort of the camel’s nose under the tent of being able to replace any worker. We have significant privacy concerns that come up on this, and no one’s raising that.

I don’t know if you guys remember when I was researching this. Microsoft had launched something recently called Recall. It was a product that they put out there, and then they retracted it because all the cybersecurity people said this is a privacy nightmare. It’s litigation bait, and they pulled it back. But when OpenAI announced this product, no one pushed back.

Alex

Can I first of all point out what a beautiful double entendre from Microsoft’s crack product marketing department: naming a feature Recall and then recalling it?

Nice. I think what we’re seeing here is one big architectural kludge.

And I think it's going to be kludgy, both from Microsoft's perhaps ill-architected Recall as well as OpenAI's Chronicle. This wants to be built into the operating system and the hardware. It doesn't want to be an add-on. I don't want an agent taking constant screenshots of my desktop, sending them to a server, and then parsing them and sending back results.

This should all be built Apple-style. I would hope that Apple will get its act together in the next few months and build this into the window manager, the compositor, and the operating system. The operating system is rendering the screen. Why can't the operating system understand what it's rendering? I mean, this is ambient AI, the term of art here, where AI is monitoring everything all the time.

And enabling you, right? This is, in one sense, what I did this past weekend with my OpenClaw, with Skippy, where I gave it access to everything. Every single Granola gets put into memory, every WhatsApp message, every email, every calendar, everything, and it just makes it so much more useful. I think something like Chronicle as well would just enable it to be, like Sam said, telepathy.

Alex

Well, that's the quandary. A lot of people who get in trouble with AI, or they get stuck, are doing something on screen. The AI doesn't have visibility into it.

Yeah.

Alex

But if you unlock that, the AI can be incredibly helpful, but it's also seeing literally every mouse move. When we talk about how our moms are still not using AI, why not? This is a big unlock, a big part of it. The voice interface and this are the 2 big unlocks, because it can then say, “Oh, I see what you're doing wrong. In fact, let me just do it for you and save you the trouble.”

All these configuration screens on any Apple device—and the menus are ridiculous now. The number of layers of configuration you can do—I think there's some crazy stat, like 70% or 80% of all iPhone users never change any defaults.

It's just too confusing to do anything. This is a huge unlock for all of that, but you said it's hugely intrusive. Right now, I take screenshots and I send them to Claude or whomever and say, “Hey, can you please help me figure this out?” But this is going to have sort of an expert over your shoulder, always there to support you if you need it.

Dave

Well, most people, when they first start playing with AI—like Alex's standard first query to test a new model is, “Build me a first-person shooter.” It's a better prompt than that. Sorry. Sorry to [laughter], but people want to do something visual and graphical to learn how it all works. Then, when it doesn't work, they want to show the AI, “Hey, this doesn't look right to me. Fix it.” So they screenshot it, just like Alex or just like Peter, you just said.

Yeah.

Dave

They screenshot it. But here, this is just a much more convenient way to get video, not just a screenshot, back into the AI's brain and say, “Look, this doesn't look right. Fix it for me.” And so you have a much more fun dialogue with the AI.

But you have to accept that privacy is being compromised there.

Alex

I think any loss of privacy here is just due to this being an architectural atrocity. This wants to be built into an operating system like macOS. It wants to take advantage of the Secure Enclave. It wants to have secure hardware that's cryptographically guaranteeing that, as it captures pixels that come out of the compositor, the window manager, and the renderer, all of those are securely handled and kept local. The reason that this is one big privacy dumpster is because it's not being baked into the hardware and the local operating system. But that can be fixed.

It will, and it will be fixed, and I want that. I've often said I'm going to give up everything, every piece of detail, because I want my AI systems to be that much more powerful.

See, you're back with us. Talk to me about what you think about this.

Dave

I think this is—I agree with Alex. Two other things, though. One is that this is going to cause massive privacy issues for workers worried about Big Brother watching over them. Already today, there's a crazy statistic that 44% of Gen Z workers are sabotaging AI's efforts to automate their own work. They're putting in the wrong data, throwing off the AI training. It's really crazy what's happening right now in workplaces. So I think this will just exacerbate it and bring this whole conversation to the front.

Oh, talk about a losing battle. You're far better getting on the wagon than you are trying to do that. That's such poisonous behavior. Protect your job.

All right. Here's our next story. Basically, World ID verification integration into Zoom. And here it is. So the backstory I think that's important here: in 2024, an engineering firm called Arup lost $25 million after an employee in Hong Kong authorized a series of wire transfers during what appeared to be a routine video call with the company's CFO and several colleagues. The problem is that everyone on the call except the victim turned out to be an AI-generated deepfake.

We've seen similar attacks in multinational firms in Singapore. The impact of this is huge, right? What we saw in 2019 to 2023 was $130 million in losses due to deepfakes. 2024 was $400 million. 2025, last year, it was $1 billion. It's projected to reach $40 billion by 2027.

And so, in steps our friend Sam with his device called the Orb that takes a photo of the back of your retina, and you verify on Zoom that you're an actual human. It uses World ID and real-time face authentication from a selfie, as well as video, and it says, “Yes. Yay, verily, this person is a human.” So you get a verified human badge on your Zoom link.

Alex

Did you just say “yay, barely”? That's fantastic. We're right back in Shakespeare here. That's awesome.

“Yay, verily.” You're a human.

Alex Hormozi

You still have to go and actually scan your eyeball in one of these Orbs.

Has anybody done it? Have you guys done this yet?

David Friedberg

No. No. Apparently it's bouncing all over Africa. People are scanning away. But I haven't done it yet. I love it because—I don't know if I told you, Peter—but I was on stage here at a company-wide meeting, and we took a little 5-minute break in the middle. Our controller came up to me and said, “Dave, I'm so sorry. I only got half of those wire transfers to China out. I'll get the other out right away.”

“Seriously, what are you talking about?” And then I got back on stage. I'm like, “I wonder what she was talking about.” So the whole second half of the company meeting, in the back of my mind, I'm like, “Wait a minute.”

When I got off, she said, “Okay, I got $300,000 out.” And I was like, “What are you doing?” She's like, “Well, you told me it was an emergency and we got to get the money to China right away.” Why would we be wiring money to China? I don't understand. So anyway, only about $75,000 got across the border. We never got that back. Then the FBI got into it right away.

But I'm like, man, digital transfers like this—everything should be logged anyway. I really feel like the digital fraud world is going to get solved, and this is a big part of it. Everything should be logged all the time. It shouldn't be that hard to deal with digital stuff. I'm much more worried about chemical, biological, radiological—

Yes.

David Friedberg

—stuff than I am about digital stuff, because I think we're going to get it fixed. And this is part of it.

Yeah. Alex, any thoughts here?

Alex Hormozi

This is Minority Report. This is the sci-fi future that we're catching up with. Apple, with its Face ID, was focused on the face, not on the retina. But if you remember the Tom Cruise–Steven Spielberg Minority Report vision, this is it.

I think it's been interesting to watch as World evolved from Worldcoin, and it's been interesting to watch as the company bounced back and forth between more crypto-focused and the economics of it versus the identification of humans, the human side of it. But it seems, from a distance, like the human identity verification side is ultimately the bigger seller than the crypto side. To the extent that's the case, as resident crypto bear, I'm very supportive.

We will have that debate soon. Don't worry about it. There's a wild irony here: the more AI scales, the more valuable verified human identity becomes. This is kind of interesting.

Alex Hormozi

Yeah.

Yeah. No. So here's this next story that's related. Grok creates a realistic AI French woman with a reflective ID. I'm going to play this little video here, and take a look at it very carefully as she holds her driver's license up to the camera. Look how beautiful and real this looks.

This was posted and went viral by this gentleman, Dr. David Lutke. He says, “This AI French woman was created by Grok, complete with perfectly reflective ID. A few more months and video ID verification may no longer be reliable.”

So, I mean, how many times have you taken a picture of your license or your passport and uploaded it? It is going to become more and more difficult. We're going to have white-hat, black-hat competitions up the wazoo here, Alex.

Alex

Well, I would maybe just comment: the IDs themselves should be verifiable with a centralized database. That's how you can maintain a single source of truth, and whether people are flashing IDs or not may be less of a—

David Friedberg

—centralized database, not a blockchain, but [laughter] good one, Peter.

Good one. I'm just poking you, buddy. I also think there are so many other technologies that we have to bring to bear. We can do hardware-level cryptography—for example, chain of custody for video. It's not that, as a civilization, we lack the technologies to ensure that any video or images actually originated from the real world without tampering. It's just that we lack the demand for it right now.

I would predict that if ever the situation of deepfaking gets so bad that it's causing real problems at a societal level, that'll just unlock all of these technological solutions, including hardware-level crypto for cameras—cryptography, not cryptocurrencies. The market will speak for itself, and we'll get all that technology. I'm still waiting for the laws to come out that require Grok and every other video-generation tool to identify it as AI-generated. It's a bill I covered in my newsletter, Innermost Loop. There is a bipartisan bill right now working its way through the House that will cover elements of deepfake fingerprinting like that.

Yeah. All right. Let's move ourselves along here. We're going to talk about the economic impact of AI. There's a lot going on. Token-maxxing: word of the year. This is from a report from 404 Media. Startup CEOs who are token-maxxing are bragging that they are spending more money on AI compute than it would cost to hire human workers. Astronomical AI bills are now, in a certain corner of the tech world, supposed to be the marker of growth and success.

Look how much I'm spending on my tokens. Everybody, you should invest in me so I can spend more on tokens. Dave, comment.

David Friedberg

No, this is a warped story. This is a great thing. The way you get left behind is by not trying. That's the worst thing you can do right now: not get in the race, not play with AI, not try.

Token-maxxing is fine. A CEO that's proud of the fact that they're consuming a ton of tokens can optimize it later in the year, but get every one of your people on their AI platform now—yesterday—and go ahead and start burning the tokens. You'll have no trouble making it more efficient later if you get in the game now.

I think it's great when a startup CEO says, “I burned $3 million on compute.” Fine. You're learning a ton along the way. Nobody incinerates money for very long; they're not that irrational. So this is just the backlash story.

What was the Jensen factor? It was like half your salary in tokens.

David Friedberg

I'm saying, yeah. I'm telling everybody, by the end of the year, you have 9 months. 50/50 is a good target: half payroll, half token use. Again, you're not going to have any trouble optimizing it. The token use is effectively about a 10x force multiplier.

So if you're at 1:1, it's like I've got 1 human and 10 AI equivalents in my bucket of endeavor. I'm actually underinvested in tokens at that point relative to human salary. So 1:1 is a better target, I think.

So, are you doing that?

Alex Hormozi

Well, I think the bigger, more healthy question is: what's the ratio of tokens to reducing iterations and maximizing efficiency, rather than just a raw spend? For now, raw spend is fine, but that's kind of a vanity metric, right? You're better off looking at it in terms of the extent to which you can compress iteration cycles. That'll be where it ends up.

David Friedberg

It's what Alex said earlier. It's dollars per token—economic...

Alex Hormozi

Impact per token. I think that's a great—

David Friedberg

If you ask a great salesperson, “How many miles did you fly this year?” that's a terrible metric of sales productivity. But if the answer is zero, it tells you it's a bad salesperson.

I think it's great when a salesperson says, “Oh, I had a million-mile year last year,” and they're proud of it. That's great. It's not the right metric, but it tells you they're proud of what they do. Token-maxxing is a lot like that, I think.

Alex, close us out on this one.

Alex

Yeah, I've seen a variety of asset allocations in recent months between humans and AIs. I think tokens-to-humans is one interesting way of framing that. A pessimist will look at this and say, “This is replacement theory. This is humans being replaced by AIs. How awful.” An optimist will look at this and say, “How incredible. We're empowering fewer people to do more and achieving higher per-capita productivity within an organization.”

What I don't hear very many people asking is: where does this end? Right now, I see asset allocations—humans to AIs, or at least human labor versus AI compute budgets—ranging from 1:1 to 1:2 at some of the frontier labs. It's an even more asymmetric ratio. The question in my mind is: is there any stationary endpoint? Is there a fixed point as this evolves?

I tend to think it's going to trend toward 1:infinity, effectively. As we start to phase humans out of the service labor force, we're going to see all tokens and no humans.

It has to. There's no other way around it. Capitalism will demand it.

Alex Hormozi

Totally agree.

David Friedberg

Until the humans merge with the tokens, at least.

Token per neuron. All right, Salim, this was your story: UAE launches agentic AI government models. This is from Sheikh Mohammed bin Rashid Al Maktoum, the prime minister of the UAE and the ruler of Dubai. He says the UAE is launching a new government model. Within 2 years, 50% of all government sectors, services, and operations will be run on agentic AI. The UAE will be the first government globally to operate at the scale of autonomy. Salim, brief us on this one.

Yeah. I did a talk for His Highness 3 or 4 years ago and talked through where this is going. Minister Al Olama, the minister of AI, is a good friend of both yours and mine, and they are going full speed on this. I have to give them massive credit. This is the benefit of the authority you can wield when you have a benevolent dictatorship.

You can absolutely get things done. When you have that, you have to make sure that whoever is in charge is doing the right things for the country, and the ethos here is 100% alignment. They are going at massive speed on this. To give you an example, I was given a golden visa, and I was asked to be the test case. The question was, could you get a golden visa authorized and issued within 5 hours?

They were freaking out, saying, “Singapore takes 5 days,” and His Highness said, “Okay, do it in 5 hours.” They got it done. There's an ability to cut through legacy thinking in a very powerful way, and this is such a massive competitive advantage. We're actually working with a few of their folks in the prime minister's office on this, and so we're very, very excited about where this goes.

There's another quote from Sheikh Mohammed. He says, “AI is no longer a tool. It analyzes, decides, executes, and improves in real time. It will become our executive partner in enhancing services, accelerating decisions, and raising efficiency.”

You can do this in an absolute monarchy. You can move this fast. What's shocking about this story is the speed at which it's moving, right? There's no parliamentary approval, no public debate or consultation. The question is, can Western democracies even keep up? You're going to see this probably in Saudi Arabia, maybe in Singapore, and in other Middle Eastern nations. Can we see anything like this in the US?

Sele

Actually, yes, you can, and I think we will. I tell the story of how it used to take 6 months to get approval for a wind turbine in, I think it was Colorado, one of the western states. Then they finally got together and mapped all the power lines, water mains, and flight paths on a GIS, plotted it on Google Maps, and made it available. Now it takes about 30 seconds to get approval because it knows where everything is. It doesn't need to take 6 months.

I think there's an economic impetus for this. This is the basis for where I think AI can make the biggest and most incredible difference, because in prescriptive workflows, you can absolutely completely automate. Almost all of government—certainly the implementation of policy enforcement—is prescriptive workflows. We know exactly the steps to remedy drivers like this. We know exactly what needs to take place, so there's no reason why that can't be handled automatically with AI in the very short future.

Step 1: give a person a super-frustrating experience. Step 2: make them wait in line longer than they need to.

Yes. Anyway, Dave, do you want to jump in on this story?

Dave

Yeah. I don't think the US has ever copied a good idea back from another country since the American Revolution. We stole the British legal system, but since then, I don't think there's been anything.

Look, you're exactly right. A monarchy can move very, very quickly. The rate at which things need to be regulated and new services need to be rolled out is way, way, way faster than any government in history has ever run before. Only AI is going to be able to do it.

If we get a great system together in the UAE, we're inevitably going to want to copy it back to the US. I think Peter asked the right question, though. Is the US ever going to—given the way Congress works—take a good idea and bring it back in? Yeah, I'd bet against that.

Dave

But it’s the right thing to do.

Weirdly, on this one, I’m more optimistic than you guys, which is weird. All right, let’s move on. We’re going to have some fun here in the biomedical space. There’s a new wave of biomedical innovation that’s coming, and I want this segment to give people hope.

We talk about longevity escape velocity on this podcast. We talk about the healthspan revolution. Well, it’s happening. I was with Demis Hassabis last Saturday at the Breakthrough Awards, talking to him, and he’s absolutely convinced that we’re going to cure cancer and solve all disease within the next 5 to 10 years—hopefully on the 5-year side.

The first story comes out of OpenAI. OpenAI releases ChatGPT for Clinicians. It just gave all U.S. clinicians—physicians, nurses, and physician assistants—a free AI copilot. This copilot outperforms all human doctors.

There’s a HealthBench benchmark they use. It scored 59 versus 43.7 for human clinicians. Pretty extraordinary. They validated this on 700,000 model responses and got 99.6% accuracy when their physicians evaluated the AI against human responses. Pretty extraordinary—something that will, I think, uplevel medicine nationwide.

From my standpoint, I’ve been saying this for a while: I think it’s going to become malpractice to diagnose a patient without AI in the loop. There’s so much going on that no human doctor can possibly understand it. At Fountain Life, we upload 200 gigabytes of data about you, including your genome, full imaging, your full microbiome, metabolome, and 140 blood biomarkers. Humans can’t analyze all that, but AIs can. Jent, any thoughts on this? Alex, do you want to weigh in?

Jent

Yeah, I’ll chime in and say the professions are cooked.

Alex

This was a widely expected release. It wasn’t a surprise. Those of you watching early releases and leaks out of OpenAI saw this coming months in advance. You can even know from those leaks what the next one to drop is: the next profession is law. There’s also one coming for management consulting and financial work.

OpenAI, thanks to GPT evals, has in some sense mapped out all of the knowledge-work verticals and is in a good position, thanks to its own internal and now external benchmarking, to know the relative strengths of its model as appropriately fine-tuned or post-trained for different verticals. I would expect to see many, many more of these ChatGPT 4.x offerings for different verticals.

In the case of clinicians, thanks to OpenEvidence and work by Epic in the form of UpToDate and other clinical AIs, this is already a somewhat crowded market that OpenAI is coming into. If I were OpenAI, I would release this sort of product more as a reference design and a way to ensure that capabilities built into the underlying models and then post-trained via a variety of evals are broadly available, and that OpenAI maintains its status as a favored foundation model for clinical and biological work.

Maybe they’ll try to monetize this as best they can. Right now, it’s available for free, but I tend to think it’s worth more to OpenAI as a distribution channel for medical knowledge and one that they can build on. OpenAI has released a variety of statistics over the past year about how many people are self-diagnosing or otherwise trying to treat themselves using ChatGPT.

I think offering a standard, regulatory-compliant channel for that is a very clever way to then make a sales upsell to biomedical enterprise and life sciences in general, which is probably where the real money is.

It’s also a data-aggregation strategy, right? I mean, OpenAI is going to be getting a huge amount of data, far more verified than “I feel this way” or “I think I might have this.”

Jent

You’re bringing a million-plus clinicians into the loop. The other thing that’s worth saying here is that current estimates are that we’re going to have a shortage of 86,000 physicians in the next 10 years.

But it’s going to be interesting, right? I have 2 nieces who have gone through medical school, my sister, myself, and lots of friends. You’re spending, literally, between college, medical school, and postgraduate training in whatever field you’re going into, well over a decade and half a million, close to a million dollars, to get this degree. Will you even need it?

Is a medical doctor going to need to be in the loop, or is it going to be a nurse plus an AI giving us all our medical advice, diagnostics, and therapeutics, with an Optimus robot giving you surgery? There’s a lot of change coming here.

Are you going to spend that much money to go through medical school and get into this profession when AI is doing the diagnosing?

Alex

Of course, this is about replacing doctors. I mean, let’s call a spade a spade. Of course, when fully developed, this and comparable solutions are about automating away medical practice. How could they not be?

And, by the way, it’s also about nursing, HMOs, and drug design. OpenAI and other frontier labs are all pursuing drug design and drug delivery. Of course, it’s about the full picture: if you’re going to solve medicine, are you just going to leave millions of human doctors practicing as sort of meat puppets for the AI? No. This is going to be the end-to-end solution. We’re just seeing the beginning of it.

Jent

Agreed. A couple of comments here. One is, in an ideal world, a doctor is getting a cognitive exoskeleton with all of this, right? You get this amazing capability to expand your own intuitive thinking. But Alex is completely right.

Alex

Yeah, but you’re going to get a huge backlash here. This is a very regulated industry. Remember, a few years ago, Texas passed a law banning telemedicine—outright banning it—because for every spot on my hand, I must go to a physical doctor. I can never do that over video.

The immune-system response is going to be very fierce. I expect to see this battle play out heavily over the next few years because there are vested interests up the yin-yang, and healthcare has the third-worst immune system ever, behind religion, education, and academia.

Jent

I’m not so sure about the immune response. If you look at what happened with the broad transition to electronic medical records, like Epic-based systems, every clinician you speak with will complain about Epic and will complain about EMRs—how much EMRs distract from direct interaction with the patient, all of that.

And yet, every major medical system, at least in this country, has either completed or is in the late stages of its EMR transition. If they can’t resist EMRs, how are they going to resist strong AI that outperforms humans?

Wait, hold on. EMRs are kind of an add-on, helpful aid because they save you time documenting the process, et cetera.

Jent

Clinicians hate EMRs. They hate the interface. They hate the process.

Alex

Of course, but they’re going to hate this 10 times more because it’s a direct replacement for the cognitive ability they’ve trained for 10 years to do. My prediction is huge regulatory and immune-system backlash on this one.

AI labs have been using healthcare as the reason why they can’t slow down, as well as the fight with China. “If we slow this down, we’re going to lose lives” has been the rallying cry.

Jent

Totally agreed. Everything Alex said earlier about this needing to be a wholesale replacement of the medical system is absolutely correct. But the path there is littered with stones and speed bumps.

Go ahead, Alex. Yeah, go ahead.

Alex

This is an interesting micro-debate. For the record, my intuition—and I interact with a lot of clinicians—is the exact opposite. Clinicians hate EMRs, but they love the AI that helps them do a better job of what they want to do.

There may be an extent to which AI interfaces like this end up being framed as the solution to all of their EMR woes—until it takes their job. Of course, that’s the way this works. [Laughter]

Let’s move this along here. Our second story is AI to reduce wasted donor hearts. I just want to show a number of stories here about how AI is going to be interfacing with and changing medical practice. I don’t know if you guys are organ donors. I am. Anybody else?

Jent

Yeah.

Currently, there are 4,000 patients who need a cardiac transplant today. There are 103,000 who need some type of transplant—kidney, liver, or lung. When an organ donor is on the table, at the end of life, and the physician has to analyze the organs and decide whether they’re viable for transplant, you’ve got about 15 minutes, typically at 2:00 in the morning, to make that decision.

In the heart world, only a third of hearts are ever actually chosen for transplantation. So here comes something called TopHeart. Just a third make it out the door.

So here comes something called TopHeart from NYU and Stanford. TopHeart is able to look at 20 different variables, right? Typically, the physician is looking at how old this person is, whether they have a drug history, if they know it, and whether there is coronary artery disease to say, “Should we ship this off?” Their goal, by looking at 20 different variables, is to give that surgeon at 2:00 a.m. a second opinion, and they believe that they can get an additional 500 hearts into the organ-replacement ecosystem.

This is on top of the fact that there is an entire synthetic-biology world going on right now to provide an abundance of organs through bioprinting and xenotransplantation—pig organs, with the antigens being replaced by human antigens. This is the work of George Church at eGenesis and Martin Rothblatt at United Therapeutics. This is an abundance story of going from a limited number of organs to an abundant number of organs. Alex, are you tracking this as well?

Alex

I’m tracking the space broadly. There are other advances as well, like trying to create a national market for organ donation versus a bunch of state markets, which would be greatly enhanced with improvements in vitrification and cryopreservation.

I think it’s good that there is a vibrant and growing distribution channel for donor hearts. I think that’s great. But I also think it’s very painful that the need for one human to die—or at least that one human dies and donates a heart to another human—is such a zero-sum situation. It’s painful to think about, and while it’s great at the margin to have more efficient ways of distributing donated organs, I really would like us to get as soon as possible to a situation where donor organs are completely unnecessary.

Yeah, and we will. I think eGenesis, Dean Kamen’s company, is advancing organ generation. They go from your skin cell to a pluripotent stem cell to regrowing you a heart, liver, lung, or kidney. A lot of this is going to be up and operating by the end of this decade, hopefully sooner.

Alex

Can’t come soon enough.

Yeah. And, of course, as we have autonomous cars having fewer car accidents, the ability to have organ donors is going to become reduced, though motorcycle accidents are probably the number one reason we get organs donated.

Let’s move on to our next story, and this goes in line with the fact that we are at the beginning of the slaying of cancer. This is a great story: “Pancreatic cancer mRNA vaccines show lasting results in trials.” I don’t know if people have been tracking this, but we now have these cancer vaccines. This is using mRNA. We used it as a COVID vaccine; this is the ability to create an mRNA that activates your immune system against the cancer that you have.

There are more than 120 of these trials going on against lung, breast, prostate, melanoma, pancreatic, and brain cancer. In this particular case, the 5-year survival rate for pancreatic cancer has just gone through the roof. Historically, it’s 13%. If you have pancreatic cancer, it’s a death sentence. Only 13% of people are able to survive it.

So, in this report, 8 out of 16 patients who generated a strong immune response to the vaccine—87.5%—are still alive after 6 years. How does this work? You have surgery to remove as much of the tumor as you can. You sample the tumor, and it’s sequenced. That sequence identifies 20 unique mutations in your cancer. That is then built into a personalized mRNA that activates your immune system like killer missiles. It activates your killer T cells to go after and attack your cancer.

This is a breakthrough in how we deal with cancer. The fact that the response is durable after 6 years is pretty extraordinary.

Alex

I remember this incredible quote from Raymond McCauley, our biotech guy at Singularity. He said, “mRNA vaccines are the first battle in the last war against disease.”

Yeah.

Alex

Amazing for me. And I think this is showing up.

My daughter works on this over at Moderna, actually, on mRNA vaccines. Moderna got a bad rap on its mRNA for COVID, but this is the Holy Grail, right? I mean, being able to go from your cancer to, “Here’s the injection that’s going to save your life,” is extraordinary.

Alex

Well, and if it works, a pretty good rap. That’s the amazing thing. It’s a universal solution. Like, you know, when Alex talks about all of math being cooked, this is the difference: in the old days, I solved one math problem; now I have an AI that solves all math. This is the equivalent in biology, where if it works, it should work everywhere.

Yeah, Alex, you’re right. I mean, mRNA was an Operation Warp Speed project. I’m just saying, afterward, a lot of people are coming down on mRNA vaccines, but—

Alex

There’s a lot of politicized griping over mRNA vaccines in general, but there’s going to be political griping over almost anything at any scale. I do think back a quarter of a century to Eric Drexler and Engines of Creation and the National Nanotechnology Initiative, when the U.S. Congress was sold a story that, with billions of dollars of congressional and national investment, we would get medical nanobots that would swim through our bloodstreams and kill cancer cells.

Well, we’re getting it, though. But we’re not getting it with diamondoid nanobots. We’re getting it with these lipid nanoparticles and Moderna- and Pfizer-style mRNA vaccines. I think it’s interesting, almost as a retrospective, to say we actually got the nanobots. They’re just fat.

Alex

They’re not silicon. They’re not diamondoid. They’re fat.

Yeah. We’re using our own machinery to do the battle for us.

Alex

That’s the other angle. Do you have a prediction, Peter? Given that immunotherapies—in some sense, very, very coarsely—we’ve known about some form of immunotherapy for 100-plus years, and people who were infected with a virus or bacterial infection 100 years ago, in some cases, showed tumors shrinking, we’ve known at some level that some form of immunotherapy would work. We’re only now figuring out how to fully weaponize it and operationalize it.

Where do you think this goes? Do you think, in 10 years, we’re all wearing Apple smartwatches that are looking for evidence of tumor DNA or RNA in our bloodstream and then send our daily mRNA update to a programmable implant or something?

I think that is basically it. Either they’re implantables or you’ll be sampled on a regular basis. The goal, of course, is to find it at the very beginning, especially if there are solutions.

There’s one more point about this that I think is really powerful: personalized medicine is actually becoming operational.

Alex

And that’s a huge inflection point we’ve been waiting for for a long time.

Here’s another example, again, just to give people hope and to see the data. The longevity mindset is about seeing this over and over and over again, saying, “Yeah, the world is changing.” The things that used to kill us are being either solved or delayed.

The single-shot CAR-T infusion shows a strong response from melanoma. It’s not just a strong response: 100% cancer-free after a single shot. This was an unexpected result. Within 2 months of treatment, all 20 patients in this trial had minimal residual disease, or MRD, negative. There was no disease identified after they were assayed again, meaning that all patients had a median follow-up of 15.3 months without any recurrence of their melanoma.

It’s game-changing in timing. So how does this work? You draw blood, you identify that you have melanoma, and the doctor finds it. We should all be scanning ourselves all the time. We do this at Found using visual screening. At a minimum, if you have a family history of skin cancer, please have yourself checked on a regular basis.

The doctor draws blood, extracts your T cells from the patient, and genetically engineers them. A gene is inserted, giving those T cells a new receptor called a CAR, a chimeric antigen receptor, that is specifically programmed to recognize the protein from your melanoma. Your T cells are then reinjected back into your body—hundreds of millions of them—and they identify the melanoma and slay it.

For the first time ever, with this type of therapy, we’re using the term “cure.” This particular type of therapy is extraordinary.

Alex

Just another example of what’s coming. This is both amazing, but can you see the clumsiness of it, requiring blood extraction and then CAR-T cell creation in vitro? Why can’t we do this in vivo? Why can’t we do this in individual cells? We’re seeing the beginnings. This is almost like the horse-and-buggy era of immunotherapies. Surely, we should be able to do this in a fully autonomous, intracellular environment.

Take the win, Alex. Take the win.

Alex

Oh my God.

Yeah. I want my FSD.

Alex

Yes.

Here’s one more story, and this is a fun one. MRSA—people have probably heard about this—is methicillin-resistant Staphylococcus aureus. It’s a killer infection, and this has typically been in hospitals. It’s now getting out into the community. So, 2.8 million people have antibiotic-resistant infections every year. It kills 35,000 people in the U.S. alone.

The problem is that all the first-line antibiotics for MRSA have failed: methicillin, penicillin, amoxicillin, and now even vancomycin, which has been the antibiotic of last resort, is no longer working.

Alex

This particular drug, candesartan, is now being used. It's an FDA-approved blood-pressure medication, and it works to basically stop and inhibit an MRSA infection. This is an example of taking an existing drug and making it fully usable by the scientific and medical community because it's been approved. We know its safety profile.

I love this. Do you remember Salim on stage? At Abundance, we had David Fajgenbaum from Every Cure.

Alex

Yeah.

This is similar to his story. I just want to tell his story and congratulate him. I'm a donor to his foundation.

In 2010, he's a 25-year-old medical student. He comes down with this rare disease called Castleman disease. They throw everything they can at him, and he's literally read his last rites. He has 4 near-death experiences.

As a medical student, he starts experimenting on himself, and he discovers that his disease is caused by a hyperactivation of the mTOR pathway. He says, “Well, if it's the mTOR pathway, I can probably downregulate it using rapamycin.” He does that, and he finds out that it works. He's been in remission from it for 12 years.

He comes up with the idea: Are there other diseases out there for which an existing approved drug can be used to cure the disease? Here are the numbers: There are 18,000 recognized diseases out there, but only 4,000 FDA-approved drugs. He's now using AI to match existing drugs and repurpose them against new diseases. It's working.

Alex

I think that's such a great example of citizen science. Take a personal problem and start hacking your way through it. I think we're going to see hundreds of thousands of examples like this. This is where people should think and understand why we're so excited about technology: this is now possible.

Yes.

Alex

This was not possible 10 years ago, or even 5 years ago. Now it's just going to become more rampant, and any problem can be solved by focusing on it, attacking it with AI, and going after it. It was incredible.

Solve everything, right?

Alex

Yeah.

Solve everything. I would also say that, historically, off-label indications were a dirty word, or dirty drugs that had lots of off-target side effects—highly undesirable. But now, if we have amazing AI models of individual cells and the body, suddenly off-target side effects become a secret weapon. We can repurpose drugs, and we can combine repurposed drugs.

I'm very bullish on the space. I advise a portfolio company, Senjam Therapeutics, that's increasingly focused on using AI to repurpose medications—anti-inflammatories for other purposes. I think this space has enormous potential thanks to AI.

For folks who are interested, you can go to Every Cure and see what David's doing. It's a nonprofit, and you can support his work. He's brilliant.

Let's get into some fun conversations here. The robots are indeed coming. A few stories to report here today. The first is that the pingpong champion of the world is now an AI-driven robot. Let's take a look at this little bit of a match and discuss it.

Alex

The background music is killing me.

Sorry about that.

Alex

Anyway, the robot's using 9 cameras and 3 vision systems. It won 3 out of 5 games.

Oh, let me pause this here. It won 3 out of 5 games. I'm surprised it didn't win all 5 games.

Alex

It doesn't have a lot of topspin, actually. Very nimble.

Note that this is the worst it's ever going to be. That's kind of incredible. The speed of response is amazing.

Alex

Yeah. This robot's called ACE. I'm not sure if I would see this in the same lineage as Deep Blue or AlphaGo, but it's the beginning.

It's totally not. This is a much lower-dimensional game than any of those board games. It frankly is astounding to me that it took this long to reach human performance in table tennis because it's such a simple game.

You only have a handful of degrees of freedom in the ball. You have its position, its linear momentum, and its angular momentum, and I think that's about it. The rest is just modeling the trajectory and maybe doing a little bit of Monte Carlo research for tactics that your opponent might take.

This should have been solved years ago. I don't know why this took so long.

Alex

Let's answer that question, actually, because that's really well said. This is very similar to many robotic operations in your home, in a factory, and everywhere else. Whatever the barrier was, I think it's probably related to the vision system.

You know, it's not a high-margin problem. Really, investing $1 billion to solve it doesn't make sense. But now that the vision systems and the feedback systems are dirt-cheap and easy, I bet it was solved by 1 or 2 people in a few weeks.

Alex

Yeah.

And that means all these other home robots can now be built by 1 or 2 people in a few weeks.

Alex

Similarly, there's a tennis-playing robot, which I'm excited to play with. That would be really cool.

But it's all the same category.

Alex

Yeah. Same. Total no-brainer. If you ever use a ball machine, then you go pick up all the balls for 20 minutes. A robot that does that is literally MIT class 270. He could have done it.

What's the barrier? I'm sure the barrier is related just to the feedback control and the vision, which you can now use with a transformer.

Alex

Well, also, people who are in robot labs don't play tennis, so they don't have an incentive to go do that work.

They don't have social lives—not to generalize.

Alex

I don't want to go too far down this rabbit hole, but there's a massive correlation between successful founding entrepreneurs and the MIT tennis team. It's basically 100%. It's crazy, including Warren and me.

All right. Here's our next story. The Tesla Cybercab is now in production. Take a quick look at this video here.

Dave, you and I saw this, and we saw the production line. We were in Austin in December. Of course, there are no controls, no steering wheel, and no pedals. It has an operating cost of 20 cents per mile, and Elon has announced that he's going to sell it for $30,000.

I think an incredible investment, if you can afford it, is to buy 10 of these, put them out in your community, and have them earn money for you while you sleep.

Alex

If you're just listening to this podcast and you're not watching the video, go find this video clip.

Yeah. In the podcast, you've got to see the interior of this to believe it. It's like you're walking into a car, but it's just a loveseat.

Alex

And, interestingly, it's only a 2-seater.

Which is the average load for an Uber. It's like 1.2 people per Uber.

Alex

Right. So 2 seats makes total sense.

Well, look, if you have 4 people, just push the button twice and 2 of them come.

When is this expected, by the way? I need this to get my kid to school so I don't have to do that. I mean, production—

Alex

Production is going off the line.

Production officially started this past week, April 24. The challenge is whether they can really build at the rate they want. Their goal is 2 million of these per year.

Alex

Yeah. Regulatory hurdles?

That's been passed now, same as Waymo.

Alex

Okay.

Yeah. It's town by town, state by state, town by town. But if you're covered, you just get it.

The difference is Waymo, because of the LiDAR and all the camera systems, and just the base vehicle, probably tops out over $100,000—probably $150,000. I'm not sure, if they get into higher production, whether it's going to come down. But at $30,000—

Alex

Yeah.

This is insane.

Alex

Yeah. No, there are so many parts. If you look at the parts just laid out, because there was that great exploded car in the showroom—

For an ICE, compare it to a consumer gas-powered car just in raw part count, and it's got to be an 80% or 90% reduction in components versus a gas-powered car.

Alex

I'll give you the statistic I always have in my head. The number of parts in the drivetrain of a combustion-engine car is about 2,000. A Tesla has 17 moving parts in the drivetrain.

The future of transportation is so good. It's just better.

Alex

Yeah. And it doesn't need a huge battery range, either. It can just go and hang out and recharge itself whenever it wants. Another one will come.

And guess what? On the transportation technology line, here's the next story. Joby Aviation—this is JoeBen Bevirt, who started Velocity11 with Rob Nail, if you remember Rob Nail. So Joby just did its first air-taxi flight from Manhattan to JFK. Let's take a listen to this news report out of New York.

Guest

And hello, I live in New York. Hello.

I know. Well, this is going to help you out, buddy. Check this out.

Guest

Massive.

Guest 2

Getting to New York airports is a nightmare. An electric air-taxi demonstration took off from Kennedy Airport and made the short trip to the West 30th Street Heliport at 11:00 a.m. this morning. If you were to drive those 16 miles, it would take more than an hour. In this cutting-edge plane, it takes roughly 7 minutes.

Guest

Joby Aviation's goal is to make this type of travel the gold standard, pointing out several pluses, including zero emissions and how quiet it is—100 times quieter than a traditional helicopter. This so-called air taxi would shuttle people from JFK to the West 30th Street Helport, as well as the one at West 34th Street and the downtown Skyport. The aircraft seats up to 4 passengers. There's 1 pilot, room for luggage, and it will fly between 1,000 and 3,000 feet.

Right now, Joby does have the green light from the FAA for this phase. And if things pan out, the company hopes to have its fleet up in the air and running within the next year. But for now, for the next week, you will see this aircraft that kind of looks like a large drone buzzing over our area.

It's a time machine, gentlemen. Can you imagine?

Guest 2

I'm standing at the heliport with my bags ready. [laughter] I love it. You know, eVTOLs, though.

eVTOLs is a lousy name, though. I just call these things flying cars, for lack of a better term. We need a better term than flying cars—a better term than eVTOLs.

Alex

It took too long. We were supposed to have these by 2015 in Back to the Future Part II. Here we are in 2026. Why did it take so long?

Guest 2

We need Mr. Fusion—

You think Mr. Fusion is the reason we didn't get our flying cars on?

Guest 2

Absolutely. That's what the movie's about.

Oh, man. [laughter] Well, in our robotics segments here, we had 2 back-to-back “Alex, why did this take so long?” questions. So let's stay very much in a “why did everything take so long?” mode. Do you think it's regulatory? Do you think regulations are why we didn't get there?

Alex

The technology has been there for quite a while. It takes a long time to do permitting. Ask Peter how long it took him for the zero.

This was 11 years to get approval to do something that NASA had been doing for 20 years. Anyway, yes, the FAA is not happy till you're not happy. That's the rule.

I think we've got to answer that question because a lot of the AMA questions are around what the jobs of the future are going to be if white-collar work gets obliterated. But I think a lot of the answer lies in these last couple of segments. Robotic stuff is going to be abundant imminently, but it doesn't just naturally happen. And so, if we can answer Alex's 2 questions on what the bottlenecks are, those are jobs.

Alex

Yeah.

Whatever those bottlenecks are, those are your jobs.

Guest 2

Those are AI models. If there's a bottleneck there, the AI will solve it.

This episode is brought to you by Blitzy, autonomous software development with infinite code context. Blitzy uses thousands of specialized AI agents that think for hours to understand enterprise scale code bases with millions of lines of code. Engineers start every development sprint with the Blitzy platform, bringing in their development requirements. The Blitzy platform provides a plan, then generates and pre-compiles code for each task. Blitzy delivers 80% or more of the development work autonomously while providing a guide for the final 20% of human development work required to complete the sprint. Enterprises are achieving a 5x engineering velocity increase when incorporating Blitzy as their pre-IDE development tool, pairing it with their coding co-pilot of choice to bring an AI-native SDLC into their org. Ready to 5x your engineering velocity? Visit blitzy.com to schedule a demo and start building with Blitzy today.

All right, let's jump into AMA with the mates. So, guys, thank you again for all the comments that you give us on the YouTube. We read them all. I have Skippy read them all as well and summarize. We pick out eight questions that we can answer every week. So, please keep them coming. Let's go to those questions.

Gentlemen, pick your favorite question off list number 1. Seem, do you want to go first?

Seem

I'll go with number 4, as I know that world a little bit, which is, “What's the future of large consulting firms like Accenture or Capgemini?” This is from Steve Bottle501. This goes full-on into the transformational effort that's going into enterprises here.

Traditional consulting is in very big trouble if it remains a pyramid of junior labor producing analysis and decks. AI totally destroys that model. But consulting firms—you know, in the land of the blind, the one-eyed man is king. In a volatile world, your clients are slower than you are and they need help.

The model will have to change. The future of consulting won't be like a people pyramid. It's an intelligence platform plus domain expertise plus change management, and we've been holding the change management side for a while. The winners are going to be bringing agentic workflows and benchmarks and governance and implementation capacity to their clients. The loser is just going to keep selling headcount. From an ExO perspective, consulting moves from experts for rent to a transformation operating system, and the companies that help their clients do that will win.

Yeah. We've talked about this before: in the old scarcity model, you put a wall around all of your experts inside and meter them out by the hour, right? That's going to get collapsed. Alex, why don't you go next?

Alex

I'll take question number 2, which asks, “Everyone can be an entrepreneur with AI as a tool. However, what action do you take when you genuinely don't have a creative idea for a direction? For most, the answer is none.” This is from 3 billionth random user.

I don't agree with the premise. I think one of the reasons why one of my funds, O21T capital, backed a firm started by friend of the pod Alex Finn called Henry Intelligent Machines, or HIM, is to solve the problem of creative ideation for starting new ventures. I think just as AI can take over as an operator of a business or a fleet of businesses, AI can also automate the process of creative ideation for those businesses.

And I think in that world—in the HIM world, if you will—the role of the human, sort of a one-person owner or magnate overseeing a conglomerate of maybe hundreds or thousands of AI-run microbusinesses, becomes one of a tastemaker. You have opinions. Everyone has opinions as a consumer of goods and services. But those opinions can shape the taste of fleets of AIs that are providing the creative ideation for businesses that they bring to you. They say, “Hey, I want to start this microbusiness for you. Do you like it?” Yes. No. And then the human can have an opinion.

The AI is performing the ideation. The AI handles the generation part; the human provides the discipline and the taste and the discrimination for which ideas pass the filter and which ones don't. That's the solution. That's how we square this circle of humans not actually needing to generate all the creative ideas themselves.

Yeah, agreed. Idea generation has never been the limiting factor. You just have to get around different people or notice the problems around you. Historically, execution has been the issue.

Okay, fantastic. Dave, number 1 or 3?

Dave

I'll take 3 and leave you with the hard one. [laughter]

Happy to help.

Dave

If you eliminate entry-level jobs but keep experienced jobs, what happens when the experienced people retire? Isn't that like eliminating babies from humanity? This is from Todd Marshall 416.

I don't think it's quite that dire, Todd. [laughter] Eliminating babies from humanity would be about the worst thing that could possibly happen. If you eliminate entry-level jobs, well, look, this was going to happen anyway.

We have a weekend place up in Vermont, and the Simon Pearce glassblowing factory is up there. If you want to blow glass, you have to apprentice with a senior person for, like, a decade, and then they let you make glass. It's a page out of 200-year-old history. That mode of operation is going to go away in all forms of white-collar work, no matter what.

The rate of change of the world in the singularity is so fast that the entry-level career path was kind of a dead end anyway. So now Meta announced a 10% layoff, which is really going to be more like 30%, according to the insiders I know, and they're definitely not hiring new entry-level people in the middle of doing the layoff because AI can do all the coding. That was not the career path you wanted in the first place.

So we're going to have to find a new way forward. But I think AI is going to be the ultimate teacher. We're going to save a ton of time on, like Peter was saying earlier in the podcast, the 4 years of medical school followed by 4 years of fellowship and internship—8 years of your life after you're already done with undergrad. It's just way too much time.

It's all going to move to AI-based, nimble training. And this massively expanding economy creates huge amounts of new opportunity every day. But it's opportunity that didn't exist the prior day, so the entry-level job wasn't really likely to lead you on that path anyway.

So it’s all got to get refactored. It’s not like people are going to stop having babies.

I think it’s so well put, Dave. Really well put.

Question number 1 I’m left with is from Gianluca Pacani Pani 808, who asks, “You guys say AI will create jobs, but for whom? It looks like AI is creating jobs for AI, not for people.”

Gianluca, the fact of the matter is, in the long run, yes, AI will be able to do any job. I think that is the case. But people still like working with people. People still like hanging out with people.

I think, ultimately, 2 things are occurring. Number 1, as every technology destroys a layer of jobs, new jobs are created on top of that. The internet killed travel agents, but it spawned millions of social media managers, app developers, YouTubers, and everything else. So there are going to be new layers of jobs coming out, and yes, those may well be displaced by AI.

Again, at the end of the day, the question is: What are you passionate about, and how do you use AI to help deliver that? There’s going to be a human interface layer for a lot of things because people like hanging out and interfacing with people—people, us puppets. So it’s going to be navigated. It’s going to be important.

I’ll just remind you of one other thing: The idea of a job is a recent creation, and most people don’t love the jobs that they have. They have the jobs they have right now because they frankly need to put food on the table and get insurance for their families. So if you could do anything, what would it be? Would it be to work? I mean, in a future of universal high income, where everything is demonetized to such a point where you don’t have to work, then you start doing the things that you love. So that’s my take on it.

Let’s move on to our second set of questions. Alex, why don’t you go first?

Alex

Let’s go with question number 5. Wasn’t all of this originally predicted by Ray Kurzweil to be happening sometime around 2040? Are we genuinely that far ahead of schedule? And this is from Brett Avalon.

I’m not sure, Brett, what “all of this” you’re referring to may mean, but I do think, broadly, we’re well ahead of where friend of the pod Ray thought we’d be. I think we achieved, as I’ve mentioned on numerous occasions, AGI—which isn’t Ray’s concept, but was popularized by Nick Bostrom and co-conceived by Ben Goertzel and some others. I think we achieved that by no later than the summer of 2020.

Ray’s approximate timeline—Ray may say I’m misconstruing his timelines—was predicting his version of AGI by 2029. So, call that a 9-year gap. Ray, and I’ve discussed this with him on the pod, is predicting the singularity, his version of the singularity, by 2045. My version of the singularity isn’t a point in time, and it’s certainly not in 2045. It’s now, it’s an interval, and we’re right in the middle of it.

Are we genuinely far ahead of Ray’s schedule? I think we are. I think Ray would probably, at this point—and has arguably said that we are—in some ways ahead of his schedule. I think the benchmarks reflect that. I think the 2045 timeline that he provided, where the superintelligence would be collectively smarter than all of humanity, is something we’re going to hit so far ahead of 2045.

Ask him next week. We’ll be with him in 6 days. From the horse’s mouth.

Dave

Yes. All right, Dave, why don’t you go next?

I’ll take the hardest one on this one: number 8, P(doom)—probability of universal destruction of all humanity estimates. Musk and Hinton say 10% to 20%; Amodei says 25%; Altman says nonzero. He actually said more like 10% when I interviewed him.

How can any of these CEOs think it’s acceptable to have a 1-in-5 chance of human extinction? They all agree with you that it’s completely unacceptable, and they all say stopping research and letting China run forward isn’t going to solve the problem. And so they trust themselves. You can debate whether that’s good or bad, but they each individually trust themselves, and that’s why they don’t want to lose the race. That’s why they’re pushing forward at full speed.

I think Musk and, along the way, Amodei have both suggested a 6-month pause, but it wouldn’t work. At the same time, they say it’ll never work. It won’t happen in the real world. So I’m just going to keep moving as fast as I can. But they 100% agree with you: This is completely unacceptable, ridiculous.

The lack of government involvement across the world is utterly insane. So that doesn’t solve it in any way; it’s just what is actually happening, and that’s what’s going to continue to happen. I’m continually shocked, as is Alex, with our inability to get any kind of government reaction to what’s now obvious. We were telling them a year ago, when maybe it wasn’t 100% obvious, but now it’s 100% obvious, yet they’re still so slow. So anyway, there’s your answer.

Alex

I’d be curious, Dave. Do you think that they believe their own estimates here? Or is this a case of revealed preference, where they think maybe it’s more socially acceptable to estimate a higher number, but actually, through their actions, they’re revealing a preference that suggests their internal estimate is much lower?

Dave

I think it’s lower. I don’t know if it’s much lower. I think they all have the same view: chemical, biological, radiological, and nuclear terrorism is the number 1 risk. So I think it’s probably lower, but I don’t think it’s like 0.001% low.

Interesting. You have 2 to choose from.

Alex

I will take number 7. When white-collar jobs are erased, where does the consumer demand come from to buy from all these new entrepreneurial ventures? This is from Tilly.

This is a tough one. This is the central political economy question of AI. If productivity explodes but income does not flow to the people, demand collapses and the system becomes unstable. Capitalism needs customers, right? So we need new distribution mechanisms, lower costs, new ownership models, AI dividends, equity participation, and sovereign funds.

All of this points—and this is similar to the previous question—to an optimistic side where AI makes goods and services cheap while giving individuals more leverage to create income. The pessimistic side is that you have extreme concentration, and then you have a massive collapse of the economy. The path we take is a governance and institutional design choice, not a law of nature.

Governance and our institutions need to freaking wake up and smell the roses here. We have to rethink this whole thing. The social contract, which is what you’re basically talking about, is essentially being wiped out. We can be optimistic about it, but the pessimistic case has a very big downside here.

The final question in our AMA today comes from James Williams Cu2 QQQ. How can a new CS engineer get experience to become a lead AI engineer if you can’t get a job in the first place?

James, first of all, as we’ve said many times, getting a job is the old model. The old model of doing well in high school, getting into a good college, getting a diploma, getting hired as a junior person, and working your way up the chain is vaporized, or at least being fully vaporized right now. The option right now is to build yourself outside the job. Build in public.

Basically, go and find something that you’re passionate about. It’s based on your massive transformative purpose, something you care about. We’re going to be launching an XPRIZE in this area very shortly. Use the tools available today to build and ship. Your GitHub is now your résumé.

Companies are increasingly hiring, if you want to get a job versus start a company yourself, based upon what you’ve done. I remember Elon said, “I don’t care if you have a college degree. I care about what you’ve done.” That is your degree now. That is your résumé. Show me that you’re brilliant in what you build, not what you happen to learn in some college or graduate degree or entry-level job.

Build in public. The barrier to entry has never been lower for you to build something extraordinary that shows your capabilities. And once you do that, you’re probably unlikely to be going after a job. You’re probably going to want to partner with a couple of friends and build a product, a company, or a service yourself.

Alex

So, boom.

That’s my answer. I’m sticking to it.

Dave

Great advice. We’re advising a couple of universities around this, Peter, and one of them is an engineering university. They’re asking, “What is an engineering degree?” It’s pretty clear that the engineering degree of the future will be: Go build some stuff, and at the end, what did you build?

And you get a degree granted not on what you learned, but on what you built?

Yep. I love that. And if you haven't done anything to start yet other than listening to the podcast, add Alex's Innermost Loop to your daily regimen first thing in the morning, and that alone will inspire you to shift gears and get into this.

Dave

Oh, thank you, Peter. That's very sweet. For those who want to read the Innermost Loop, just go to alexw.org and I provide links to Substack and X and Spotify, et cetera.

All right, our outro music today, which is beautiful, is from Hitham Said. It's AItopia.

All right. Thank you to my brilliant moonshot mates AWG. I wish you a beautiful week. Dave and C, I can't wait to see you guys next Monday. We're all together again. We're going to be physically at MIT at the book launch of We Are as Gods. We're going to be recording a podcast episode there.

Guest

Can't wait to do it face-to-face.

Dave

May the fourth be with us.

Yes, for sure. As a Star Trek fan, I'm not allowed to say that.

Dave

By the way, check out what's right above me. It's a world vision camera identity camera. Right, right, I've got surveillance literally right over my head. No, it's just an Omni.

It's just a surveillance camera, but I couldn't resist.

Dave

By the way, everybody standing in Guadalajara Airport is holding a laptop at eye level because there's nowhere to put it down.

You did pretty damn well as a mobile.

Dave

I've got my exercises for the day.

I saw you moving around. Were you trying to avoid policemen or something?

Dave

I just have to shift positions down there, shift hands, and once in a while lean on something. There's nowhere to sit here that's easy, and I don't want to risk losing a connection that I fought so hard to get.

Oh my God. Okay, if you've got an outro song or an intro song, please send it to us. media@diamandis.com. We'd love to hear it, see it, and potentially play it. And thank you for subscribing to this, and thank you to all of the fans out there. I know all four of us run into you on the street, at the airports, at events, and it's so great. If you see us, do come up and say hi.

All right. Take care, folks. Bye.

If you made it to the end of this episode, which you obviously did, I consider you a moonshot mate. Every week, my moonshot mates and I spend a lot of energy and time to really deliver you the news that matters. If you're a subscriber, thank you. If you're not a subscriber yet, please consider subscribing so you get the news as it comes out. I also want to invite you to join me on my weekly newsletter called Metatrends. I have a research team. We spend the entire week looking at the meta trends that are impacting your family, your company, your industry, your nation. And I put this into a two-minute read every week. If you'd like to get access to the Metatrends newsletter every week, go to diamandis.com/tatrens. That's diamandis.com/metatrends. Thank you again for joining us today.