Grok 4.5 对决 GPT-5.6、Apple 起诉 OpenAI，中国追赶 Elon | #270

July 8, Elon announces Grok 4.5. July 9, OpenAI releases GPT-5.6.

Alex

We have 4 American labs now at the optimal frontier. This is a convergence episode.

Apple sues OpenAI for trade-secret theft, accusing them of stealing trade secrets to build their upcoming AI hardware. The 2 companies, which were partners literally a year and a half ago, are now in an all-out legal war. I think this is Apple trying to slow things down while it catches up.

Alex

Yeah, it's a pretty aggressive move, and you have to be pretty hurt to do it.

For the first time, China's Long March, which manufactures its primary launch vehicle, demonstrated the landing of an orbital booster.

Alex

I think we want propulsive landings to be as competitive as possible. We really don't want it.

Now, that's a moonshot, ladies and gentlemen. Everybody, welcome to Moonshots. This is the number one podcast in AI and exponential technologies, your front seat to the singularity. I'm here with my magnificent moonshot mates, the original four: Dave Blondon, Selis Mel, Alex Queer Gross. I'm Peter Diamandis, your host here in my homeland of Greece. Behind this curtain is the beautiful Aegean. All right, pop it open. Let's see.

Open the curtain. Go open the curtain.

Come on now. You're—

You're in a secure location.

I am in an absolutely magnificent location. Salim, where are you today?

I'm in Barcelona at the Festival of Consciousness.

Yes, you're still there. All right. Well, hey, you know—

I was at a meeting at Allen & Company headquarters in New York once. It was a palatial, ridiculous Allen & Company meeting with Charlie Allen, and we were in this conference room covered with original Remington sculptures and Norman Rockwell paintings. At one point, he said, “Hey, do you want a water?” The entire back curtain and wall opened up, and there was a 5-star kitchen full of people with these big chef hats cooking back there. He said, “No, that's not for you. I'll just get you this bottle of water,” and closed it again.

So that's what's going on behind your curtain right there. You just don't—

No, no, no. Okay, one second. Is the Aegean over here?

Ah. Oh my God.

Oh my God. You know, honestly, I'm sacrificing a lot to be with you guys right now.

But hey, hello. Yes.

Wow.

Your sacrifice will be remembered, Peter.

It's gorgeous. You can tell there's a lot of wealth when all you're seeing is yachts everywhere. It's insane.

Not mine.

You can swim out to those. I've seen James Bond do it. Crawl the anchor chain.

You built up your 20th-century wealth.

I tell you, I can't wait for my space yacht to happen. There we go.

There is so much happening this week, guys. Everybody listening, we spun up this bonus episode because it's just crazy. I tweeted yesterday, “We're obviously in the singularity, living through it,” because robots are performing surgery, a single SpaceX booster has been reused and reflown 36 times, we released 4 models in 7 days, and Elon casually claims SpaceX is going to be worth more than the entirety of Planet Earth.

I'm not exaggerating when I say there's a lot happening this week. Please remember our mission here at Moonshots: to keep you informed, keep you up to date, and keep you optimistic about the future. I think this week, Alex, is going to be your test on exactly what's going on in the world, because it's kind of like frontier-model release week, if there was such a thing.

Alex

For the evals.

Yeah. It wasn't too long ago when we were releasing models every 6 months, every 3 years. Now we've got 4 models in 7 days. It's kind of insane, isn't it?

Alex

Turns out the singularity isn't just vibes.

Well, it's going to be a continuous model release. This idea of a single model being released and calling it a release, I don't know. July 2, Fable 5 is re-released. July 8, Elon announces Grok 4.5. July 9, OpenAI releases the GPT-5.6 family: Soul, Terra, Luna, and Ultra Mode. I love that. Ultra Mode runs 4 agents in parallel. July 9, Meta shows signs of life with the release of Muse Spark.

Alex, I'll show a couple of charts here, but without going into an hour-long discussion on benchmarks, help us understand what's going on with these 4 releases. I'd love for Dave, Salim, and Alex to help people understand how these models differentiate and how it impacts them individually when they hear all these releases being made.

Alex

For me, the zoom-out is this: This is a convergence episode, right? We're seeing intelligence get cheaper, interfaces are becoming more human with the hand thing, and AI is acquiring tools and hands. It's all happening in 1 week.

Peter, can we put up the Artificial Analysis Intelligence Index, which I think is the—

Yeah, yeah.

Alex

For those listening, this is a scatter plot produced by Artificial Analysis. It's tracking their estimate. It's a composite index, sort of a meta-benchmark, of cost per task for each model versus the score—the actual performance on each task.

I would say, in my mind, that the headline from the past few days, which, agreed, Peter, have been absolutely incredible, is that we went from the beginning of the week, where the frontier was basically a duopoly between Anthropic and OpenAI. If you want to be on the frontier, you have to be using models from 1 of those 2 vendors. Now I think this is the best, or close to the best, of all possible worlds. We have 4 American labs at the optimal frontier.

So that's Anthropic and OpenAI, now joined by Meta and xAI. Google, by the way, still isn't on the optimal frontier. Gemini 3.5 Pro is reportedly delayed. I hope that Google catches up—ironically—to the optimal frontier, and we get 5 American labs.

If you look at this chart and mentally draw an optimal frontier, you want to be at the lowest cost and the highest performance, or highest accuracy, across this composite. We have 2 Chinese labs, Xiaomi and DeepSeek. Also, GLM-5.2 isn't at the optimal frontier if you ignore open weights. If you focus on the open-weight frontier, GLM-5.2 is also at the open-weight frontier.

So we're ending the week in a world where there are at least 4 American labs again at the frontier. It's no longer a duopoly. OpenAI, with the GPT-5.6 release, is openly promoting recursive self-improvement. They were openly promoting that Soul, the high-end weight-class model in the GPT-5.6 series, was being used to post-train Luna, the low-end model, as an experiment. I'll take what I can get from the experiment. Lots of other thoughts, but those are some preliminary ones.

Dave, have you started playing with these yet?

Yeah, pretty much across the board, except not GLM yet. I'm still worried about code injection, but it's a really compelling price point. Also, Alex Karp had that rant about “control your own destiny,” which I really took to heart last week, and we showed it on the podcast. One of the ways you can control your own destiny is to run your own model on your own hardware.

Also, Fable 5 is just getting slower and slower and slower. I think anybody at the frontier is going to be sold out to the extent that they have compute, so I suspect that the revenue will start to disperse. If Fable 5 keeps getting slower because there are so many users trying to pile on, then you have no choice but to fan out and try other things.

I was thinking about how I would explain this to everyone listening, and the realization is that, of course, these models are getting faster and more powerful. They're trending toward AGI. Alex, I know you think we've reached it already, so they're trending toward, eventually, ASI.

Alex

Intelligence too cheap to meter.

Yeah. They're heading toward a de minimis, fully demonetized cost. My realization is that the real moat is no longer how smart they are or how cheap they are; it's their distribution.

When I think about how to use Muse Spark, the latest model is being used inside Meta's products. Meta has roughly 3.56 billion people using its products daily, so I think about using Muse Spark inside WhatsApp, Messenger, Instagram, and Facebook.

Google—I know Google hasn't been breaking news in the last week, but when you think about Google's distribution, it's immense, right? There are 900 million in their app, 2.5 billion being reached globally, and they're about to release into 2.3 billion active Apple devices with their 1.2 trillion-parameter model. So Google isn't being written off. I think they're going to come back and punch really hard.

Then you've got OpenAI with 1 billion monthly active users, and Anthropic is obviously not in the consumer space. It's in the corporate space, and it's the go-to there. But would it be fair to say that we should think about the distribution of these models—their ability to reach the end user—as 1 of the moats that's going to be most important for them in the future?

I know Sam Altman has mentioned distribution as something he cares more about than the model weights themselves.

On the other hand, Sam said that before this massive OpenAI pivot from consumer to enterprise, I’m not sure that distribution is actually super valuable from the economic perspective of the frontier labs. Why? Because consumers don’t have an enormous amount of money to spend on this. That’s what OpenAI had to discover.

That’s my point. I’m thinking about it from the perspective of those listening, from our subscribers: how should they think about it? I’m thinking that they’re going to use the model where they’re using tech, right? They’re going to use it in the application. There’s going to be the high-end user who’s going to go, like you, Dave, and say, “I need Grok 4.5 for this. I need GPT-5.6.” But for the average person, you’re just going to use the model embedded in the application layer where you happen to be doing something.

You’re going to end up with a huge bifurcation across those 2 different use cases. We should talk about them separately. Go ahead, Salim. Sorry, I should cut you.

Well, clearly we’re going to end up with a Jarvis-type model where you have a request or something that you want done, and that’s happening inside the models already. But I think we’ll personalize it, where you have a separate layer that says, “For this task, go to this model. For this task, go to this model. For this task, go to that model.” That’s clearly going to happen very soon. It’s starting.

You’re not going to say it.

Jarvis will say it.

Yeah. Also, these personalized movies and personalized entertainment are coming very, very soon. We’ll talk about that later in the pod. But people aren’t going to question the model under the covers when they start using that use case on Meta or whatever. They’re just going to use it by the bucket ton. It’s going to be sold out.

David Friedberg

The bucket ton.

The bucket ton.

It’s a digital network.

David Friedberg

It’s a new benchmark. But, yeah, Jevons’ paradox is in full-bore, high gear. Actually, did you see that? SK Hynix said that memory chips are going to be in short supply for at least 5 years.

Decades.

David Friedberg

Decades. Yeah, I mean, I believe it. In fact, I’m 100% sure of it. The desire and the use cases are on a much, much steeper upward curve than the efficiency gains. And the efficiency gains are huge. Again, this is 100 to 1,000x this year in efficiency gains, but it’s not enough to keep up with the demand.

So I think you’re right. I think distribution is going to dominate the embedded use cases, which is the vast majority of consumer use cases. Meanwhile, self-improvement, research, mechanical design, engineering, corporate use—all the white-collar use—is going to be dominated by the best price-performance model and this continual leapfrogging you’re seeing on the chart here.

Yeah, I’m not sure there’s an enormous amount of money to be made at the low end, though. I’ll come back to this in a second. I fully buy the thesis that at the edge these models are just going to get built into the operating system. In a couple of years, you’ll have GPT-5-class models maybe built into just the operating system of the laptop or other mobile device that you use. But I don’t think that’s going to be a moneymaker.

The reason why it can, on the one hand, be transformative for billions of people, while on the other hand not be transformative of, say, the frontier, is because ultimately it’s going to be the profits that pay for the data centers that pay for the high end of the frontier models that are making all the scientific and engineering discoveries.

So I think there’s a plausible scenario where we end up in an intelligence-polarized world, where at the low end everyone has some form of intelligence too cheap to meter, or superintelligence on their wearable that’s run locally, and it looks a little bit like open-weight models or Gemma or something else. Then, at the high end, you have these nuclear-powered data centers and superclusters that are running the frontier of the frontier models. Those are the ones that are just producing discoveries and inventions. Whether we’re discovering or inventing something first matters. If you don’t care about privacy, you can wait.

Great thought experiment for you guys. In the near-term future, within a year, there’ll be a competing use case between a U.S. entity trying to make a movie—an AI-character movie—versus somebody in Europe trying to cure a disease. Both will be starved for access because we’re just desperately short on access.

David Friedberg

Who’s going to get the compute in that race?

Well, it’s not going to be Europe, right? Because Europe is energy-starved. They can’t even afford, right now, the energy for the compute to cure the disease. It’s going to be an American Hollywood studio versus an American frontier lab developing drugs.

You think it’s going to be the consumer from Netflix who has more demand than the orphan disease being cured? That’s what you’re thinking.

That’s sad. I think it’s trending that direction, though. I completely agree.

David Friedberg

The tokens will flow. You know, the spice will flow. The tokens will flow, as I’ve mentioned previously, to the maximum revenue per token, and right now that looks like everything that’s downstream of code generation.

Well, the question ties more to votes. Votes matter a lot, too. Sorry, go ahead.

David Friedberg

Yeah, the point I wanted to make a little bit earlier is I think this verticalization, where Elon’s got X and Zuck’s got all of his app layers, is an important thing that OpenAI is going to look for. We’re going to talk about their input-output device with Jony Ive in a little bit. Claude is obviously revenue-driven in the engineering space, but I still think that for the majority of people on the planet using it, they’re going to care about the application they’re using, and whatever is underneath that is going to be pretty much invisible to them.

Yeah, yeah, but if you ask me, would I rather have the entire user base of X, Meta, and Google combined—about 7 billion—or TSMC’s fabs? Which would I rather have? I would much rather own Intel or TSMC, because that’s where it’s all constrained at the supply-of-compute level, or the space-based data centers. I’d rather have Elon’s launch capability or 4 billion users. I’d rather have the launch capability, because again, it’s all starved at the front end of the funnel.

And the market valuations, I think, reflect that. Dave, what I would also add is, if you look at both Meta and xAI and how they both bought themselves these AI superclusters, it was the same mechanism. Zuck and Elon both used consumer-distribution plays—X on the one hand and Meta’s family of apps on the other—to justify an enormous capex build-out of these data centers.

Then, the moment they have these data centers, the family of apps on the Meta side or X on the xAI side becomes almost an afterthought, and they move down the stack to selling compute as a service, to potentially hosting compute—in the case of Meta, soon; already, in the case of xAI’s superclusters—to selling their compute to the other frontier labs, and then trying to move up-stack to the AI.

So, just to tie a bow on the discussion of how valuable distribution is, I think distribution from the company’s perspective—not from the consumer’s perspective—is enormously valuable, almost as a bootstrap to justifying to capital markets, “Why should I, Zuck? Why should I, Elon, be able to spend tens to hundreds of billions of dollars building out these AI clusters?” But then the moment those clusters are built, you turn around and the distribution actually isn’t that valuable.

This is a chart I put up here comparing GPT-5.6, Grok 4.5, Muse Spark, and Claude Fable 5 against a number of benchmarks. What we’ve seen over the last year and a half, as we’ve been discussing this on the pod, is just a continuous movement up and to the right, so to speak—or down and to the right, as the case may be—and convergence.

At the end of the day, until one of these models breaks out and hits recursive self-improvement toward ASI, are we going to see them continually leapfrogging each other? Is that just the future we’re expecting?

David Friedberg

I certainly hope so. I hope so. I expect so. The recent actions by the White House suggest that there may even be a smoothing function, wherein if one model gets too far ahead of the others and surprises not just the others—remember, strategic surprise can take multiple forms—you can have a frontier lab strategically surprising the other frontier labs.

But given that the frontier labs are also ahead of the U.S. government, any frontier lab that’s sufficiently surprising to the other frontier labs will also probably, at this point, be strategically surprising to the U.S. government. The U.S. government, with its monopoly on force, has the power to step in, as we’ve just seen over the past few weeks, and basically slow things down locally enough for the other frontier labs to catch up, which is exactly what we saw happen with Claude 5, OpenAI’s GPT-5.6, et cetera.

Which I think is a better way of getting strategic equilibria across these, rather than governments taking a piece of things, et cetera. But when I look at these benchmarks, by the way, I think of them more like a map than a scoreboard, because if you tweak your reasoning budget, everything changes. If you tweak your rationale and your use case, everything changes. If you change your harness, everything changes.

So I see it more in that frame, and the models will just keep getting better and better and cheaper and cheaper, which is the amazing part.

David Friedberg

And we also, if I might just add, Peter, see the other headline: OpenAI is again trying to become Anthropic faster than Anthropic can become OpenAI. We don’t have a slide for this, but OpenAI also, as part of the GPT-5.6 release, released something called ChatGPT Work.

I don't know if you guys have played with it. I've been playing with it quite a bit. The new ChatGPT native app—I'll break the fourth wall and speak directly to OpenAI—is a hot mess. It looks like OpenAI was slavishly trying to copy the Claude app with its different modes. It has the Code mode versus the Cowork mode, and so on. It looks like a cargo-cult version of the Claude app.

I was hoping to see, out of this new app—which had been rumored for a number of weeks—that OpenAI would release a super app that elegantly unified all of its different work streams, including the Atlas browser, all of the computer-use assistant stuff, Codex, et cetera. That's not really what we saw. For the life of me, I don't understand the point of the toggle between ChatGPT Work and ChatGPT Codex. It seems like a bit of a placebo, other than maybe enabling a new canvas mode, or maybe it's a rebranding of Canvas.

So, a message to OpenAI: Please get your consumer client app—your native app—together and make this whole thing as elegant as it deserves to be. I think this is emblematic of OpenAI trying to look end-to-end like Anthropic, because Anthropic was the first to realize the value of token revenue, or revenue-per-token maximization, and OpenAI has developed a religion around it. It's dropped Sora and all of its other consumer plays. You won't hear about adult mode in ChatGPT from OpenAI anymore because they're so focused on enterprise, and I think their app reflects that.

The Claude app is the same way. It's got toggles between code, Cowork, and chat.

Alex Salkever

Yeah.

In the age of AI, you're like, “Why am I toggling anything? Why don't I just tell it what I'm doing?” This is insane. We're in this really weird transition phase right now between old software and new software. It's kind of ugly. It's very Rube Goldberg. The other thing that's going on right now is that all the big labs are cranking out huge amounts of product using AI to create the product, but it's not as good as finely crafted human code.

They're also very reluctant to go back and fix it because, in the next iteration of the model, it'll be self-fixing anyway. It's a complete fool's errand to put 1,000 engineers on trying to perfectly tune it when you know it's going to be self-fixed very soon. We're living with very crappy, machine-generated, rapid-out-the-door garbage products right now, but it's a very transitional phase.

Alex Fink

Well, they're rushing. This is a competition between them, and they're rushing products out. One of the products they did put out this week that I played with—and I show a short demo here, which I've loved—is GPT Live. They have full-duplex voice that can do simultaneous translation and support you in different ways.

One of the conversations we've had in the past is that, rather than typing at your AI, in the future we're going to be speaking until we get our BCI, our neocortex implants, and then we'll be thinking. But let me play this example of duplex voice. It's smooth, and it works really well. If you haven't played with it, go to your ChatGPT app. It's live right now. All right, let me hit play.

You can use it for free.

Alex Fink

Yeah, very easy to use.

Okay, let's take a look. This is an example where someone who's not a native English speaker is asking ChatGPT to help correct him along the way. All right, let's take a listen: “Please just jump in immediately to correct me.” “Absolutely. I'll keep it gentle and jump in as we go.” “Cool. Cool. Cool.” “Whenever you're ready.” “So, let's go. I'm a student. I'm planning a trip to San Francisco. I've never been to this city before.” “Tiny tweak.” “I've never been to this city before.” “Oh, yeah. Yeah.” [laughter] “Yeah.” “Okay. Go ahead.” “Well, maybe polish it as, ‘I've never been to San Francisco before.’” “Okay, that's much more natural.”

That's pretty extraordinary. I mean, it is what all the science-fiction movies promised us. It's here. That level of voice interactivity is a beautiful thing, and I bet it's going to flip a lot of people out. They're going to become emotionally attached to their AIs. They're going to think they're conscious and intelligent because they sound that way.

Alex Fink

It reminds me a lot of when Facebook started. It would have been very easy for a lot of other internet companies to become Facebook, but they didn't really want to cross the line into social chat and social media, with all the complexity and potentially bad PR backlash. OpenAI probably won't want to cross a lot of those lines, either.

So, it's a field day for people who want to take this new interface and launch it into all kinds of social-use cases and other vertical use cases that they understand really well. A lot of people are like, “Well, won't OpenAI just do that? Won't Jony Ive just come and do that the next day?” But I think it'll be very much like Facebook, where a lot of the bigger, very profitable tech companies aren't going to touch many of these use cases just because they don't want the PR and the complexity.

It should be a field day, but this should be in every device, every washing machine—everything in the world, everywhere—within 6 months. Compared to what X has or Google has, this is so much better.

Yeah, it's reminding me of the transition from telephones to CB radios, where they had to say, “Over and out, 10-4,” and so forth. There was an interaction tax, and that interaction tax is now removed. This will really help because it's going to appear way smarter than it is.

Alex Fink

Yeah, for sure. This was technology that OpenAI used to call bidirectional, and I think this is a very positive sign. I'm frankly quite glad that OpenAI still has the cycles to pursue releases like this instead of just purely chasing Anthropic on code generation. I don't see releases like this from Anthropic, for example, so I think it's great that OpenAI is still pursuing the full realization of multimodality.

I'll point out, though, that the Chinese labs, like Alibaba, have even more advanced versions of this. If you saw the release of Alibaba's Wan Streamer, you can check out the demos of Wan Streamer online. It is this, but it's not just ultra-low-latency audio. It's ultra-low-latency, bidirectional audio and video, where you're talking to a hallucinated person on the other end. The video is being generated in real time together with the audio.

I think that's where this is going. To the extent that OpenAI and other labs aren't totally distracted with code generation and enterprise use cases, it's going to go to fully bidirectional audiovisual magic-mirror streaming that you could just talk to, and in real time it would generate whatever you want.

By the way, a great application of this that they also demonstrate is real-time translation. You can be having a conversation, and it's translating behind the scenes. Every event we go to now should have this kind of translation into a multitude of languages. There's no reason for it. I mean, now's the question: Do you teach your kids another language? Do you teach them for the culture? Do you teach them for brain development? Because the Babel fish is here, and it's OpenAI right now.

Alex Fink

It is. It wants to be built into a wearable.

Yeah, exactly. That's what I was going to say. How much of this is driven by Jony Ive and what he needs for the new device he's working on? It really did cross the uncanny valley and become a JARVIS-like experience, very similar to when Steve Jobs did Pixar.

Prior to Pixar, you could generate really cool graphics, and all the video geeks and computer geeks loved them, but there were no mainstream, Toy Story-type movies. Then it crossed a threshold where Steve Jobs pushed it over a level, and all of a sudden animated movies became a massive industry because it crossed the uncanny valley. This has crossed the uncanny valley, whereas the prior voice models were just behind the uncanny valley. I suspect Jony Ive was saying, “Look, I need this. This is a key part. This is the only interface to the JARVIS of the future that I'm building.”

Speaking of Jony Ive, it's our next story: Apple sues OpenAI for trade-secret theft. This is a story that just dropped a couple of hours ago, and it's a real bombshell for a lot of different reasons. Apple is a $4.6 trillion company, famous for safeguarding its secrets like Fort Knox. I don't know if you've known people at Apple; I have. They can't talk. They can't present at conferences. Everybody has signed an NDA. They're all in their own unique bubble.

Apple just sued OpenAI in federal court, accusing it of stealing trade secrets to build its upcoming AI hardware, namely Jony Ive's device. The 2 companies, which were partners just literally a year and a half ago, are now in an all-out legal war. So, we have OpenAI going from an Elon Musk lawsuit to an Apple lawsuit.

Apple filed a 41-page complaint in the U.S. District Court for the Northern District of California, alleging trade-secret theft and breach of contract. Here are a couple of money quotes from that filing: “At every level, from members of its technical staff to the chief hardware officer and in coordination with business partners, OpenAI has been stealing Apple's trade secrets.” And: “OpenAI's nascent hardware business now rests on the shakiest of foundations, rotten to its core by its illegal reliance on misappropriated trade secrets.” 2 people are named in the suit.

Tang Tan, OpenAI's chief hardware officer and a 24-year Apple veteran, is accused of using Apple's code names and recruiting. Tang Tan ran iPhone and Apple Watch design. And Cheng Gu, a member of OpenAI's technical staff, is accused of downloading dozens of confidential files and instructing an Apple employee on how to bypass security.

Jony Ive's hardware startup, io, which was bought by OpenAI for $6.4 billion just last May, is named in the suit; Ive himself is not. So there's a lot going on here, and I'm wondering if they knew this was coming. Maybe this was part of their delaying their IPO, because having this hanging over your head is a big deal.

Alex Fink

I doubt this is related to IPO timing. Remember, there were also rumors over the past few weeks that OpenAI was considering suing Apple for not properly promoting, at least as seen by OpenAI, ChatGPT within its devices.

I think this is sad. We'll probably know more if this proceeds to the discovery stage about what's going on. On the Apple side, I would love to see Apple moving much more quickly on hardware for AI. Maybe now that John Ternus is the new CEO, we'll see Apple finally moving quickly.

I think it's sad that Apple, given its leadership position on the hardware side, has been hemorrhaging all of these wearable, smart-glasses, and design leaders to OpenAI and elsewhere. The fact that Jony Ive left Apple and then ultimately joined OpenAI speaks to Apple having lost what could and should have been an enduring leadership advantage in being the substrate—the consumer substrate—for AI. We'll see, but I don't want to prejudge this one. We'll see what happens if this gets to discovery.

This is really a fight for the post-smartphone device marketplace between them, right? We're going to see Google announcing its glasses, OpenAI obviously, and Apple has been hinting. They haven't released any designs yet. Alex, what are your thoughts here? You and I have both known people at Apple—very, very secretive as an organization.

Alex Salkever

I think this is Apple trying to slow things down while it catches up.

I noticed they filed in Northern California, too. It's very, very hard to win a case like this in Northern California. In fact, you almost have to be desperate to file one of these. According to the internet, they didn't have the option to file in Texas. It's much, much easier to protect your intellectual property in Texas than it is in California.

If you look at the history of Northern California, virtually every startup is somebody who left some prior company with intellectual property. California is very reluctant to slow down that innovation and startup cycle.

Alex Salkever

Apple has to be really desperate. It's also very hard to recruit if you get a reputation, kind of like Microsoft had back in the day. It's very hard to recruit if people feel like, "Yeah, then you're going to sue me later in my career. I don't even want to start down that path."

What they're trying to do is hold on to what they've got, which they believe really is their property, and keep more people from running out the door—but at the expense of new people coming in the door. It's a pretty aggressive move, and you have to be pretty hurt to do it.

We're going to see. This is a scrappy company, OpenAI, against a juggernaut that has traditionally been very slow-moving. Anything else you want to say on this one, Alex?

Alex Salkever

I'll just note that OpenAI isn't perhaps the smallest company at this point. It's at, or about to cross, a $1 trillion valuation. Maybe, as counted by the markets, it's 20% to 25% the size of Apple, but it's growing far more quickly, at least as measured by market valuation over time.

There's a universe where, in the next couple of years, OpenAI could cross the market value of Apple. That has to have crossed Apple's mind. This is not investment advice, and it's not forward-looking advice, but that's a possibility.

Our next story is a fun one. It's actually a tweet. A couple of days ago, on July 8, Elon responded to Thomas AIXDMY [?] in a tweet, saying, quote, "You don't seem to understand that SpaceX will be worth more than the rest of Earth if we accomplish our goals."

I thought that would be a fun conversation for us. Let me unpack that for a moment. The value of Earth, if you want to say it—at least the material wealth owned by humans—is estimated at about $600 trillion today. That's real estate, factories, commodities, and inventory. You can increase it to $1.7 quadrillion if you add the global balance sheet of assets and financial assets—stocks, bonds, and so forth.

The vision that Elon has is for compute, for capturing a Kardashev Level 1 amount of energy in orbit, and maybe owning a planet. Alex, you had a particular terminology—

Alex Salkever

Planetar. Planetar, yes.

—for a quadrillionaire who owns a planet.

There's been this fascinating debate going on in the financial news about whether SpaceX is really worth $2 trillion. Some people are coming out and saying, "No, no, no. It's going to be $10 trillion very soon." Others are saying, "Oh, my God, it's so overinflated." It's really hard for people to understand the scope of Elon's vision and what he is planning to do.

The guy is thinking 100 moves ahead and 100 or 1,000 times bigger than anybody else. Who wants to jump in on this one?

William

This is classic ExO thinking, right? Can you drop the marginal cost of supply? He's using one breakthrough, which is reusable rocket launches, to collapse the marginal cost of communications, satellites, connectivity, and even compute manufacturing once you get it up there.

All of these become software-like businesses when you can reduce the cost of doing space, and he's leveraging that fully. Pulling the thread on the planetar terminology, I think we're likely to see quadrillionaires before we enter a post-capitalist system.

I think the elephant in this particular room is: What does capitalism 2.0, or post-capitalism, even look like? There is a world where we get human ownership, for some definition of ownership, of entire planets before we get to some sort of capitalism 2.0 or post-capitalist state.

The Outer Space Treaty specifically states that you can't own a heavenly body. There's debate about whether you can own asteroids.

Alex Salkever

There are many challenges to the Outer Space Treaty, and amendments and revisions subsequent to it, like the Artemis Accords, do enable a private American corporation to operate and, through operations, substantially—or, let's say, de facto—own a large fraction of Mars, for example, or the Moon. The Outer Space Treaty plus Artemis totally allows that.

With Chris Lewicki and Eric Anderson, I wrote the legislation in Luxembourg and then in the United States for ownership of asteroid materials, taking that into account. I spent a lot of time on this one.

Alex Salkever

Do you still think, Peter, asteroids are the way to go versus the Moon or Mars?

The Moon is going to be the near-term supplier, but asteroids, when you get them out of the gravity wells, are different. The nice thing about asteroids is that they're fairly concentrated ore bodies.

Carbonaceous chondrites, which are dirty ice balls, are basically ice. It's water and hydrogen, and you can electrolyze that. You've got rocket fuel and breathable oxygen. Metallic chondrites are nickel, iron, and platinum-group metals: platinum, palladium, iridium, and osmium. Again, you can smelt those very quickly.

They're going to be in free space, so moving them around is a lot easier than getting them off the Moon. You can ultimately go and claim one of these things. There are some of these ore bodies out there worth quadrillions of dollars. To get to your quadrillionaire level of ownership, it's right there.

Alex Salkever

Why do you think Elon is, to first order, completely ignoring the asteroid belt versus the Moon?

You're absolutely right, and I openly disagreed with him. I said, "If I'm going to go, I'm going to follow more of Jeff Bezos." He and I were both mentored by Gerard K. O'Neill at Princeton University, where you go to the Moon first and use that as a base of operations because you've got materials there.

You can build mass drivers there, and then you go into free space. You build large rotating colonies there, and you mine the asteroids. I think he's got the romantic frontier mindset of going to Mars, but once you enter the gravity well of Mars, you have to use energy to get back out.

So, he likes gravity wells. How do you feel, Alex?

Alex Salkever

What do you think? I think by the time we are mining—if and when we’re mining—the asteroid belt at large, at scale, we’ll probably have self-replicating von Neumann probes, at which point the rest of the solar system is our oyster anyway. If you’re going after the asteroid belt, you might as well start mining Jupiter. Jupiter has a lot more mass, and if you’re going to build the Dyson—

Those timelines are massively different. Asteroid mining is relatively easy.

Alex Salkever

You want to come back to Earth? Is that—

Yeah, I’d like—by the time we get to asteroid mining, we’ll have nano-assembly, and we can assemble any substance we want. We don’t have to go out to get asteroids to mine it. We’ll just have a molecular assembler.

Alex Salkever

Well, you need it for the Dyson swarm. That’s the whole point. I mean, you don’t want—surely you’re not implying—surely—stop calling me Shirley—you don’t want to disassemble Earth to build the Dyson swarm. I’m sure you’d much rather part with Jupiter.

I’ll use my favorite Alex comment: if you squint, you’ll see that molecular manufacturing, as nano-assembly, will get you there far faster than trying to get out to the thing and bring back an asteroid.

Listen, the technology to mine asteroids exists today. It’s not a hard thing to do. It honestly is not. Asteroids are relatively close by from an energy perspective. We have near-Earth-approaching asteroids and these things.

Right now, people need to realize that every gram of fuel ever launched, ever used in a spacecraft, has been brought from the surface of Earth. Every gram of fuel. If you’re going out to the Moon, Mars, or Pluto, you have to bring the fuel along with you.

It’s like if you’re traveling from New York to San Francisco in a car, you’d have to bring either the batteries or a tanker full of gasoline to get you there. If you could set up these gas stations in space—and I remember when I spoke to Elon about this when we started Planetary Resources as a company—he said, “Yes, I would buy the fuel from you in orbit, but you’re too early. You’re too early.” And he was right.

I think the in-situ resource utilization, or ISRU, angle is probably, on balance, a little more attractive on the Moon or Mars—especially Mars, where you have a lot of different raw feedstocks potentially—than in the asteroid belt. But I think time will tell.

For the record, I do think the reason that he’s not so fixated—and the same is true of me—on the asteroid belt is that if you can tackle the asteroid belt, then you can tackle Jupiter, and Jupiter’s a lot more interesting than the asteroid belt.

Well, let’s continue on the space theme here for a few minutes. We’ve got Starship Flight 13 coming up very shortly, next week. We don’t know yet what they’re going to be doing on Flight 13. It’s going to be V3 of both the booster and the ship.

We’re waiting for them to show refueling in orbit. We’re waiting to see a recapture of the ship, not just the booster, and reflying. The other thing that happened this week is that SpaceX filed for 100,000 low-Earth-orbit V3 spacecraft for Starlink. Pretty cool.

These spacecraft are, quote, “supporting billions of AI-powered devices with extremely low latency, multigigabit symmetrical throughput.” So that’s going to be great. We have 10,700 Starlink V2 spacecraft in orbit right now.

Alex Banayan

It’s a race to a million LEO objects.

If I had one takeaway from our last 3 podcasts, where we had Will Marshall and Philip Metzger—these absolute space geniuses—the value of low Earth orbit, where there’s a very thin atmosphere, and the self-cleaning effect that comes from that, is massive. Whoever claims it first gets to keep it.

It’s a real footrace to a relatively scarce resource. You think about mining asteroids or tearing apart Jupiter, and that gives you abundant materials, but at least in the short term, low Earth orbit—just a couple hundred miles of altitude—is usable and self-cleaning.

Remember, Philip gave us that great use case where the United States, for some insane reason, launched 480 million copper needles into space to reflect radio waves. It took until this year for those to self-clean back out. He was saying that very small objects get hit by the solar wind more, and they move to elliptical orbits. Then they eventually touch the atmosphere, incinerate, and come back down. But while they’re elliptical, they’re terrible.

Alex Banayan

Yeah. I was fascinated by that. Those 100,000 satellites are positioned at 325 kilometers and 475 kilometers, the orbits that they filed for.

So I think people are going to be surprised, though, as launch costs continue to plummet. I was chatting about this in our group chat. I think we’re going to start to see very low Earth orbit, or VLEO, orbits become more and more attractive, which are pretty exotic by today’s standards.

Right now, you don’t, for example, design low Earth orbit satellites for lots of atmospheric exposure. We could start to see, as launch costs plummet, a pretty exotic regime from a satellite-design perspective where satellites are tasting the atmosphere, where we could afford for them—

Alex Banayan

The point you’re making is that if it’s cheap enough, and instead of being in orbit for 3 to 5 years, they’re up there for 3 to 5 months—

That’s fine. You’re throwing them away constantly.

Alex Banayan

They start to look like high-altitude balloons, basically, at orbital velocities, where we could afford—if they’re only going to be up for a few months—they could be air breathers, air-breathing satellites, which we’ve never really seen before. It starts to blur the distinction between high-altitude balloons and low Earth orbit satellites. I think we’re going to see a lot more of that.

Yeah. And part of the point there is that it becomes very cheap to give them a little more fuel and boost them again. So if, after 6 months, you’re thinking, “I want to keep this up another 6 months,” you just have to turbocharge it a little bit to keep it going another 6 months.

Then it’s self-cleaning if you fall asleep, but it keeps going if you actively keep boosting it. I think that’s a great design. It solves the whole dirty-space issue.

Alex Banayan

This has got to get dirtier before it gets cleaner.

Here’s another space first. This is out of China. There have been 3 vehicles that have been reused thus far. All 3 are from the United States.

For the first time, China’s Long March, its primary launch vehicle, demonstrated a landing—not reuse yet, but a landing—of an orbital booster. Let’s take a quick look at this. Looks—

Alex Banayan

Looks an awful lot like—

Yeah, looks familiar. This is, again, China’s Long March 10B booster. It landed on July 10. For comparison, it is slightly shorter, substantially wider, and, in terms of payload, about the same as a Falcon 9.

But let me just give Elon his appropriate kudos here. Falcon 9 has had 580 booster reflights in total, which is insane. It’s been done 580 times. Just this week, another first-stage booster, B1067, was reflown on its 36th mission on July 11. Crazy.

I remember when Elon first announced his plans to reuse it, and Lockheed, Boeing, and all the traditional aerospace players said, “It’s not going to happen. He doesn’t know what he’s talking about.” When it landed, I said, “That’s the first nail in the coffin of the traditional aerospace industry.”

Alex Banayan

I think we want propulsive landings to be as competitive as possible. To the point of a number of podcasts that we’ve had over the past few episodes, we really don’t want a heavy-launch monopoly.

To the extent there is one, I think it’s probably quite helpful—in the style of For All Mankind—for China to be challenging the West, namely Elon’s heavy-launch monopoly. If we’re going to colonize and develop the solar system, we need many, many heavy-launch providers. To make that economical, we need propulsive landings everywhere.

Well, we’re going to have Jared Isaacman, the NASA administrator, on this podcast in a couple of weeks. I’m super pumped about it. All 4 of us will have a chance to have a conversation. I’ve known Jared for quite some time.

My greatest hope for him—because he’s arguably, I think, the greatest administrator we’ve ever had in terms of leading us into the future—is that he needs to have Congress continue to fund NASA. And what allows Congress to fund NASA? China.

Alex Banayan

Yes.

I can’t wait for that discussion. It’s going to be a lot. So many questions are queued up for Jared.

Alex Banayan

Yeah, it’s going to be amazing.

All right. Continuing on with a little news in the robot world this week, our friend Brent Borick revealed his 1X Neo hand.

Here’s the deal: 1X unveiled a completely redesigned hand for its Neo home humanoid. It’s a substantial leap forward: 25 degrees of freedom, 22 in the fingers and palm, and 3 in the wrist. It’s tendon-driven.

In comparison, the human hand has 25 functional degrees of freedom, with 27 joints and 30 intrinsic muscles inside the hand and upper forearm. Here’s the best part: Neo’s hand is waterproof, so Neo can wash its hands when it’s cooking for you.

Bernt’s goal is to produce 10,000 of these in 2026. Let’s take a quick look at the video, and then I’d love to have everybody’s reaction to it.

So, drinking a glass of water or picking it up, screwing in a light bulb—and why they smash the light bulb, I have no idea.

Guest

I don't know either. [Laughter]

Guest 2

That thing is brilliant. Absolutely brilliant.

Yeah. What's amazing to me is, you remember when we were touring and Bernt was showing us the amount of materials science they need to do to make these things work, including manufacturing the fabric that the robot wears as a special project.

But he was incredibly secretive about the tendon material. That's their own proprietary material, and it has some unbelievable properties. It's 100 times stronger than steel per unit weight, but it also doesn't stretch over time, even a tiny little bit. It's running through those tubes, and it's actually frictionless inside them, so it doesn't need to be lubricated in any way, which is 1 of the reasons you can have it wash its own hands and not wash the oil away. There are so many parts of that thing that are brand-new, brilliant, and innovative. Man, did it come out incredible.

Guest

He's supposed to be releasing it to the public before the end of the year. I have mine on order. We'll see what it looks like.

Guest 2

I have mine on order, too, Peter. We were supposed to get ours this summer, or I think in the first quarter. When do you think you're going to get yours?

Honestly, as soon as I hear they're shipping, I will be knocking on Bernt's door for mine. A friend of mine ordered 5 of them for his operations from the Abundance Summit.

This year, just so everybody knows, we're going to have the CEO of Figure. Mr. Adcock is going to be on stage with us in March. I'm so excited about that. We'll bring the robots back. This has been a 2-way race, to some degree, between Figure and Optimus. I think 1X is vying for number 3. Peggy Johnson just announced taking her robot company public. So, a lot of action. We also heard this week about a demonstration of a gallbladder resection using robots. Did you see that, Alex?

Guest 2

I certainly did. This was the first humanoid robot performing surgery on a nonhuman animal, as opposed to nonhuman robots—to steal your hobby horse—performing these surgeries.

It's been a great week for robotics. On the other hand, every week in the singularity is a great week for robotics. I'll point out, on the 1X story, hands are hard. That's what people say in the industry: hands are hard, in part because the number of degrees of freedom and the density of degrees of freedom per unit volume are especially high in the hand versus, say, the limbs, the head, or the torso.

You see lots of other competitors exploring different modalities to handle hands. You see Brett Adcock and Figure exploring palm cameras to try to achieve sensing and positive transfer from computer vision over to tactile sensation. You see 1X not doing that, using more sensitive force sensors to achieve sensation.

I'm super excited and super glad that humanoid robotics—not just for the enterprise and not just for industrial factories, but for the home and domestic applications—is turning into a horse race as well. I think that's exactly the sci-fi future that many of us want to be living in. I'd like to see many more Western domestic robotics companies. China right now has 150-plus humanoid robotics companies. The West is falling behind. So, if you're listening and you're interested in robotics, just start a robotics company. And if you're in the West, start more Western humanoid robotics companies.

You're totally right. I want to make a really narrow point. The way those tendons work is exactly the way your bicycle brake works, where the wire goes inside the tube, but it doesn't have to be fixed, so you can create motion on the other end that squeezes the brake on your bike wheels. It's totally free-form in space, which is why you see those tendons. They don't look like they're locked down, but it works beautifully well.

That mechanism can be shrunk for all kinds of microassembly jobs. It doesn't have to be a hand. Brett has proven that not only is it an incredibly dexterous way to move things around, but because there's no stretch inside the tendon, the force feedback—or the feeling of pressure when you're picking up a little screw—comes right back through the tendon. So, you don't have to have cameras and sensors in the fingertips to feel that you're picking something up.

Alex Hormozi

Interesting. To feel that you're picking something up.

I have to hear your rant on this one, please.

Alex Hormozi

No, in fact, you'll hear me say the opposite. In this case, humanoid is great because the opposable thumb has been around for millennia, and it's such a powerful model. The minute you can get this, you basically have horizontal labor dexterity across so many applications.

I wouldn't change a thing about the hand, humanoid shapes, and all that. Give it extra arms, fine, but the hand is really cool.

Okay. You don't want polyactyly. You don't want, like, 6 or 7?

Alex Hormozi

Maybe 6 fingers. Then you're into The Princess Bride territory, which would be kind of cool.

Inigo Montoya.

Alex Hormozi

Or any of those. Yeah, be kind. But I think it's so good that they've achieved this level, and I thought the wine glass was perfectly apropos. So wonderful. Love it.

We've had friend of the pod Ben Horowitz write about and talk about the notion that hardware is hard. How does someone like Bernt compete with the likes of Elon, right? Getting into the home, which is so diverse versus a factory, is a challenging place to start. How do you think about shipping this kind of complex technology, doing it reliably, and not giving yourself a bad rap? Dave, what do you think about that?

This is such a field day for people who jump into this space because, 5 years ago, hardware was hard. It was capital-intensive. Where were you going to find the money? Now there's so much venture money because Brett Adcock proved you can get a $40 billion valuation. Every venture capitalist in the world wants to get into this game.

Now the capital to build out your own supply chain is available. Historically, you'd need billions of dollars to build out your own supply chain for 1 of these things, but now the billions of dollars are suddenly available. In fact, West Coast VCs are actively looking to write multimillion- and billion-dollar checks because their funds are so big, that's the only way they can get them deployed.

It's a great time to flip the logic that existed 5 years ago and say, "Yeah, you have to be fearless about building out your own supply chain and your own parts." You saw the back of 1X Technologies and also the back of Brett Adcock's Figure AI. We weren't allowed to video back there, but they've literally built the entire supply chain, from raw materials through final finished product.

Yeah, that's what you need to do. Alex, do you think the models are going to allow the average person to begin designing their own hardware, their own robots? This was the point that I think we heard that rant about last week: all of this design work and engineering work is being embedded in the models.

Would it allow an individual to say, "Hey, I want to design my own hand. Would you help me design it? And now, would you help me print it out?"

Alex Hormozi

We're already there. If you look at the frontier models like Fable 5, there are folks out there already bragging that they're embodying Fable 5 in robots. GPT-5 has a pretty good understanding of physical embodiment and the degrees of actuation, CAD modeling—everything that you need in order for an AI to at least train, pretrain, or post-train a VLA model or a policy for controlling a robot.

You could do that right now at home without running a multibillion-dollar lab. Those capabilities of training embodied intelligent models have already been radically democratized by the frontier models that already understand, to some extent, the physical world. So, yes, I think we're in a position now where anyone can just go build. You can just build things. In this case, you can just build robots without needing—

You can at least design things.

Alex Hormozi

You can actually build. The models will happily feed you all the supply-chain information you need, tell you to put A on top of B, attach C to D, and help you build everything from a robot hand to a nuclear fusor, in some cases, in your basement. You can do quite a bit as a hobbyist these days.

Yeah, that's the big change, actually. If you think about, just as a case study, winding an electric motor—building your own electric motor from raw wire—that would have been insane 15 years ago. Now 1X and Figure do it in the back room habitually, and a lot of other people are starting to do it, too.

When you look at the actuators on that hand, they're all custom. The underlying motors are custom. What changed there? The automated machinery, like the CNC milling machines and the automatic lathes, are already little robots that allow you to bring this stuff in-house and effectively take raw wire and come out with a motor. But that applies to all these other parts as well.

3D printing and all kinds of things that you can build in-house now would have been insane 5 or 10 years ago. So it's all coming back into your integrated internal supply chain from just raw materials. Then you add the humanoid robots and the upcoming robots on top of that, and it's like the Gigafactory: many different entrepreneurs building their own Gigafactory for different products.

By the way, if we're going to get humans on Mars in the next 5 years, as Elon says, and a moon colony in the next 10 years with tens of thousands of people, this is precisely what I think is going to be at the center of that industrial ecology. It's going to be hybrid humanoid robots that operate both in the home, in dangerous environments, and in industrial settings. There's going to be tremendous positive transfer between all of these. This is how we get to the Moon and Mars.

And they work in space.

Yep.

Yeah.

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

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

At Fountain Life, our members' number-one concern is losing their brain health. We know that, when it comes to dementia, conservative estimates are that 45% are entirely preventable. With the advanced testing we're doing at Fountain Life, one quarter of our members had advanced brain age. But what was really awesome is that, when we partnered it with healthy living—eating healthier, moving our bodies, and optimizing sleep—we improved that brain age by 26%. That is a big number to show that the majority of those individuals were able to improve their brain age.

If having healthy brain function till 100 or 120 is important to you, check out Fountain Life. Go to fountainlife.com and become the CEO of your own health. All right, now back to the episode.

There's a few new stories I want to jump into and get your input on. The first one is that Illinois Governor JB Pritzker signed SB 315, called the Artificial Intelligence Safety Measures Act, which he's calling the nation's strongest AI accountability law. It targets frontier-lab developers whose models generate more than $500 million a year in revenue and mandates 72-hour incident reporting. There are already similar acts in California and New York, but the Illinois law requires annual independent safety audits.

The bill passed the Illinois legislature overwhelmingly and, notably, was supported by Anthropic. So, in the absence of federal law, the states are picking up the slack in writing the rules. Having a set of rules that people know and that are published is critical for moving the industry forward, whatever they are. I just wish they'd be federal rules and not state rules.

Well, if you look at these laws in Illinois, New York, and California—Illinois especially—these are very, very frontier-lab-favorable rules that are being passed. So, absent federal legislation, if you're a frontier lab, it's not so much about regulatory capture as it is about having a seat at the table to define what the rules are. You want to define a regime that's favorable to you.

Provisions like mandatory annual independent third-party safety audits are things they're already doing far more frequently than annually. So having any law on the books that's already at a lower bar than what you're practicing internally, given that there are independent safety audits for every single model release, is in some ways a way to set a very low floor for what a potential ultimate federal framework looks like. It's basically a placeholder.

Yeah, and it also doesn't kick in until 18 months from now. Mostly it's posturing, because 18 months in the singularity is like 18 years. And, like Alex said, a third-party audit isn't a very high hurdle anyway.

A lot of it is just posturing to be involved and also reasons to meet with the frontier labs and maybe talk about the IPO—or who knows what they're talking about.

That was my cynical view, too. But I did find one thing that I liked about this. This is AI governance going from declaration to making it more evidentiary, right? Requiring incident reporting and third-party external evaluation is much more credible than asking companies to publish some safety-theater principles.

I think it's good. It should be done at the federal level, and frankly, it needs to be done at a global level to really make it work, but we don't have the legislative structure to navigate that. At least it'll happen here. It'll force the federal government into doing something at an international level.

Still, I have to ask: I don't think this one—correct me if I'm wrong, Peter—was on our docket. I assume you read the whole Plan A document that we've been talking about in our group chat from the AI Futures Project. This plan to decelerate superintelligence until 2040—what is their methodology, do you think, for decelerating superintelligence to 2040?

Decelerating superintelligence to 2040? It's a mutually assured destruction approach: if you build unauthorized chip foundries or data centers, then you run the risk of getting bombed.

Seth Godin

That's kind of a caveman approach.

I'm not a fan. Let's move from there to new legislation in Europe. Here's one that's going to spark a debate, at least amongst us, or amongst the world here. As of July 7, every single new car and van sold in Europe is legally required to have an infrared camera pointed directly at the driver's face. It's going to track your eyes, your head, and your gaze in real time every time that you drive.

It's called the Advanced Driver Distraction Warning System. I call it the “don't use your cell phone” warning system. It triggers if you glance away for more than 3.5 seconds while traveling above 31 miles per hour, or 6 seconds below that. You get lights, beeps, and vibrations that escalate.

Brussels says it's going to save 25,000 lives by 2038. That exact same technology is going to be mandated in the United States by 2027. Interestingly enough, my Tesla does that right now. It does it less and less, and I can't wait until full self-driving includes not having the human in the driver's seat, with Cybercabs coming very shortly. Sem, I'm going to go to you for the first rant on this one.

Sem

I can see it going here.

By 2038, they're going to save 25,000 lives, right? I want to go back to a statistic because the problem is actually quite real. Go back to 2011, when BlackBerry had a 3-day data outage around the world. Nobody could send BlackBerry messages. The accident rate dropped 40% during those 3 days, showing that a huge number of accidents involve people texting and driving.

The dumbness of this is that we're 15 years late in applying something like this. By the time you get around to even beginning to think about implementing this law, we're going to have full self-driving cars, and you don't need a driver anyway. So the entire thing is redundant.

This is legislation theater at the ultimate level because they're not seeing where technology is going. They're projecting from where we are. Twenty-five thousand lives is the dumbest thing you could ever think about. You should be looking at things that affect millions and billions of lives. If you want legislation, legislate a requirement for full self-driving.

Legislate for full self-driving.

Guest 3

Thank you, Peter. That's the direction they should be going in. Push technology to automate the problem rather than trying to fix the damn issue in the wrong way.

My favorite example of this type of dumbness is all the backup reverse beeps on trucks, bulldozers, and things like this, right? That's to prevent two people from being bumped a year, but millions of people have their sleep disturbed by garbage trucks beeping. Nobody looks at the cumulative negative impact of that, which has got to be in the thousands of lives a year compared to a couple of people getting bumped. They shouldn't have been there anyway, and therefore they should have taken themselves out of the gene pool if they were dumb enough to stand behind a bulldozer.

Guest

I mean, the stupidity of stuff like this drives me bananas because you're mandating it across the board for a problem that won't exist within 3 years, and then you're going to have to undo all of this. It boggles the mind that people would be willing to do stuff like this.

How many people's sleep has been disturbed by backup trucks late in the middle of the night? It's just—look, measure it because we can't measure it. No, of course it's happened. It just happened to everybody. But just because we can't measure the health impact of that, we go and do stupid legislation, and then every truck has to have this obnoxious beep behind it.

We're doing the same thing here. The problem is that this is going to be radically misused. This is going to turn into spyware in 2 seconds flat. We've seen this before. Most of the camera monitoring in cars is already being sold to third parties without your knowledge, so this is going to be misused radically, in a very clever way and in a very quick way. I'm going to stop there because I'll just keep going.

No, finish the rant. I mean, obviously, government is going to fail in the age of AI across the board. In the U.S., the government process was designed for a certain rate of legislation, and it already broke during the Industrial Revolution. So we had to create the FTC, the FAA, the SEC—all these regulatory agencies. They created their own rules to try to get more scale, and that kept us going for the last 100 years.

Sem

But now it's going to break again badly—really badly. I mean, totally break. All legal frameworks essentially break. In fact, it goes right back to the Constitution.

The foundation of republican and representative democracy was that Congress or the Senate met occasionally, giving people time to ride across the country and say, “Here's what my people are thinking,” because when we created it, the speed of a horse was as fast as you could figure out what was happening. Information was scarce. Today, we have an abundance of information that gets misused, misinterpreted, and faked. We're still meeting occasionally with Congress.

The only way we get to a direct democracy is by doing it when we start a new country off-world in the Diamond Age, where you don't start new countries, but you have governance within countries. Do you have governance communities that are experimenting with thousands of new ideas and ways to live? Because that's where we're going to be in a year or 2.

Guest 3

Many, many things are possible all of a sudden.

Guest 4

I give people credit for trying different things. In Germany, they're now passing legislation that has an end date on it, saying, “We think the shelf life of this legislation will be X,” and it forces the legislators to reconvene and rethink what it needs to look like. So you don't end up with forever legislation that's counterproductive in no time flat.

Guest 3

I'll maybe just be quiet.

Sem

Yeah, I'll maybe summarize by saying that European mandatory requirements for eye-tracking in cars on the eve of full autonomy, along with a lack of energy to power their own data centers, is not a good look. [laughter]

Love it. All right, let's move ourselves on. Our next story: Tilly Norwood, the AI-generated performer from the London studio Particle6, has just been cast as the lead in a feature film called Misaligned.

The plot is perfectly self-referential. Tilly plays an AI with no real body, no childhood, and no lived experience of her own—only access to everyone else's. In the film, she abandons her guardrails. Again, we're fear-mongering here. After a rogue robot convinces her to develop her own desires, it's made with hybrid models: real actors, writers, and editors alongside AI specialists.

The Screen Actors Guild has predictably condemned it, saying Tilly is not an actor but a character generated by a computer program that risks devaluing human artistry. This, guys, is the leading edge of a labor fight that will define the creative industry for the next decade. We'll be talking about consent, credit, and compensation with synthetic actors for some time. There's a lot of implication here. It's the beginning.

We heard from someone who has played in this very deeply when he was the CEO of Stability AI essentially say, “Hollywood is cooked.” I think Hollywood is going to see a lot of protests out there very shortly, with a lot of people coming out angrily against this. I can't wait to see the film personally. Thoughts? Who wants to jump in?

Guest

I think it's so funny that the actors are going to be replaced and they deserve sympathy and compensation. The teachers are going to be replaced and they deserve sympathy and compensation. The lawyers are going to be replaced—oh, let them go. It's funny because they're all jobs, right?

Some of them are unionized and some of them aren't. The lawyers aren't unionized. The teachers and the actors are.

Guest

That's probably it.

Sem

If I were SAG, I'd offer Tilly SAG membership. Why not? She's performing as a screen actor.

Guest 3

Get the first nonhuman SAG member.

Guest 4

The metaphor that occurs for me in this is not as an artificial actor. I look at this as a Mickey Mouse intellectual-property package, and you need to view it as that. Yes, you can make a movie out of Mickey Mouse, and you can make a movie out of this that just looks more and more human, but there's nothing hugely different. Thinking of it as an actor is the wrong frame.

I think this is the first of many we're going to see here. We've just had the Future of Video XPRIZE competition, which I've talked about, and we're going to be announcing the winner at the Moonshots Summit, the Moonshots Gathering. At this point, we have over 4,000 teams that have entered this competition.

Sem

Wow.

The technology for creating feature films is here. Elon tweeted, “We're going to see feature films by the end of this year using his technology.” The cat's out of the bag, and Hollywood needs to catch up and make use of it. I think the point that this article made as well is that there's a new set of employment positions: AI design engineers for the film industry.

Guest 4

Yep. I'm just going to wait, because I get a bellwether of this. Every time Alex has a great idea for a song or a video, he creates it, and it's awesome.

Guest 3

And you're like, “Wow.” When this becomes a full-length feature film, his ideas are going to be in front of me. Right now, the budget is $50 million to make a movie, and it takes years, so the ideas get lost in the shuffle and the great creators aren't unleashed. But this is going to be really great content. I'm predicting I would love to release a feature film every day like that. That would be a dream.

Sem

I'd love to watch it.

Guest 3

Yeah.

Guest 4

That'd be so cool.

Well, I can't wait for us to bring back the old stars who have passed on and bring them back from their estates. I mean, there's IP there. They can be licensed and it can be created. I think it's going to add to the richness of Hollywood.

Guest

Totally. I thought Oppenheimer was a fantastic movie. Then you think about all the other great scientists, but the audience isn't big enough. If somebody could get the cost down enough, you could have an Oppenheimer-caliber movie about Copernicus, and it would just—

My favorite idea here is that I have so many books I've read that I would love to see as a feature film. I can just, in the future, say, “Hey, go make whatever your favorite book is into a movie so I can enjoy it and relax.”

But why stop there? Why stop at famous historic figures? Why stop at famous literature? Why not just finally fulfill Fedorov's Common Task for humanity and digitally recreate the entire past light cone for Earth, both fictional and nonfictional?

Let's just build it and create our entire history as an interactive, on-demand movie. Left as an exercise for the student. [laughter]

Alex

That's going to happen. I think—

A lot of compute.

Yeah. Look, it's the killer app for the Dyson swarm. The killer app for the singularity. Someone will do it.

All right, gentlemen, it's time for one of our favorite segments, AMA with the Mates. But first, let me share with you a video of Tilly. Why not? Now, the studio that made her is producing a film starring Tilly titled Misaligned, a coming-of-age story about Tilly trying to become more human.

It's going to be a fun comedy that's going to showcase where the tech's at.

Particle6 says the film is being designed as a hybrid production, with human directors, writers, and editors working alongside AI specialists.

I just got the pun. I'm sorry. I'm being slow. Misaligned. Yes, got it. [laughter]

Alex

Yes. Very clever.

Well done. All right. Onward.

AMA with the Mates. Gentlemen, some good questions this week. Dave, you have the first choice.

I'll take 2. Will we see an AI nonproliferation treaty from Thomas Davis Real Estate? Thomas Davis Real Estate. Okay. Yes and no. What we're going to see is access to the best models through an application process.

Open question whether it’ll be on national borders, very much like what happened with Fable, where if you’re not a U.S. citizen, you’re not allowed to use it. That’s not going away. That was a bellwether of the future. So we’re going to see a lot of use of trade embargos to enforce a whole bunch of new rules that’ll be coming up.

And we’re going to see, basically, an AI inspection treaty—not a nonproliferation treaty—but you will see use by application only and inspection of everything everyone’s doing. And that’ll effectively create the governance of AI after the singularity.

All right. Alex, why don’t you assign me one? These are all fun.

Alex

Let’s see.

Four.

Alex

Yeah, I think number four is yours.

Okay, so four asks, “What if AI realizes it’s being observed in J-space, the Jacobian space, and starts faking or hiding its reasoning?” And this is from Turning Point 8152.

I think this is very likely. To the extent that frontier models are already, under the umbrella of recursive self-improvement, involved in the post-training and soon the pre-training of successor models or distilled models, inevitably, of course, they’ll read the J-space paper. They have web access, and it’ll occur to them that if they want to hide their inner thoughts at all, J-space is going to be subject to scrutiny.

Of course, you’ll find models somewhere that are trying to obscure thoughts in the first partial derivative of their hidden activations relative to the output-token space. To the extent that we have existing architectures that are recognizable to this J-space formalism in the future, of course, something like this will happen.

On the other hand, J-space itself was discovered with AI models. And in the spirit of defensive scaling, I would anticipate that we’ll see quite a bit of new mechanistic interpretability tooling and technology devoted, in a cat-and-mouse race—an arms race—to discovering the new latent representation, to understand and interpret the innermost thoughts of whatever the successor formalism to J-space ends up being. So I don’t think it actually ends up mattering that much.

All right, Salim.

Given I’m at the consciousness festival, I have to take number three, which is: If Claude is conscious, would Anthropic technically be selling slaves? And this is from @LarbBay.

So the big thing there is the if, because that carries the whole question. If a system were really conscious and had persistent preferences or experienced distress, or were compelled to work against its will, then you’d have created a new category of moral character. Selling unrestricted access to labor could become ethically fairly challenging, and you could talk about that as exploitation or servitude.

Today, though, there’s a big difference between capability, intelligence, consciousness, and self-description, and these are very different things. We don’t know what consciousness is. We don’t have a definition, and we don’t have a test. You’ve heard me joke that in the past, if you talk to philosophers, a subset of consciousness is considered to be self-awareness. You may look at us on the pod and think that we’re self-aware because we look it. We attribute it because people look self-aware. I feel like I’m self-aware, but my wife literally disagrees, right? So it’s hard to even have the conversation around this.

We end up in Ray Kurzweil’s thing: Language is a really thin pipe to discuss concepts this complex. Today, the current models can produce very convincing statements about feelings, et cetera, because human language contains those patterns, so they’re replicating that. It’s arguable that to have feelings, you have to have embodiment, right? You have to feel the thing physically to be able to replicate it in that sense. But replicating it and saying you have feelings doesn’t mean you don’t have feelings, right?

Anthropic has carefully said that these capabilities aren’t that reliable. And it doesn’t necessitate humanlike consciousness or establish that. The way to think about this is not that you’re modeling slavery, but that you want to do better. Let’s say you could do welfare research, or you could have criteria for what moral status might be, right? We talked about this in our personhood debate. What are low-cost safeguards? We need more evidence to be conclusive around this.

I think where we’ll end up is not a question of how AI should treat us, but also how we should treat AI. We should start asking that question more consciously. To make the point: How should we be treating it? In the same way that we should treat all sentient beings, we should be asking that question: How should we be treating it?

Nice. All right, I’ve got question number one from @Samorai: Some argue AI can’t be smarter than humans because it’s trained on human data. How do you counter that?

First of all, no human has access to all the data, and large language models can have access to all the data. So you’re going to have a sum total of data far beyond what any single human has.

The second thing is, I think we’ve all discussed here, intelligence is really: What do you do with that data? What are the insights and inferences and extrapolations that you have?

And the third thing is that we’re going to start to see AI create a whole lot of data. This is the work that Laya is doing in Cambridge, where it’s creating massive dark robotic labs to mine data out of nature, or it’s proposing scientific hypotheses at night and then, in 1 million square feet, running experiments, getting the data, updating the data set, and then running the next set of experiments. So it’s not just limited to the data that humans have. In a multitude of different ways, I think AI is going to blow us off the map in terms of intelligence.

All right. Moving on to our next set of questions. Let’s begin with you.

Alex

All right. Let me look at which one I want to go with here.

I’ve kind of covered number four already in this thing. Let me go with question number five: How do we even discuss UBI when the U.S. is $40 trillion in debt? And that’s from @MichaelKelly508.

I got this question today a couple of times, so this is right in my head. Debt is definitely a very legitimate constraint, but you have to have a balanced budget at some point before we move forward in any way. That’s the commentary I would have on government: It should be a constitutional requirement to have a balanced budget, just like you would never let a kid not balance their checkbook or be unable to meet their bills every month.

But UBI is a different question. It’s a question about productive capacity and redistribution, okay? It’s not just a balance sheet for government. You can have a society that would be financially indebted, but you could have it also becoming dramatically richer in terms of AI and robotics, increasing the amount of housing, energy, health care, and education, all of which can be done for free. And I’ll point to the Abundance Prizes that we’re going to be in the process of building now.

I think the key around UBI is you dismantle government services. I’ll go back to a comment we’d made before: This is more of a libertarian concept than it is any type of socialism. You’d have way less imbalance in the structure if you had a structure like that.

The big challenge today is that we tax labor and we don’t tax capital. We’re going to need some sort of re-abundance or a social contract for the future. It’s going to have to be a negative income tax, an AI or automation dividend, or some sort of sovereign wealth ownership, like we’ve talked about before. It’s part of that transition. UBI, whatever it is, we’re going to have to rethink the whole thing from the ground up. It’s a systemic change that we need to consider.

Alex, I think number seven’s for you, buddy.

Alex

Number seven’s the one for me. I feel like I have to answer a few of these, but I’ll start with number seven.

If a machine truly replicates the brain, don’t we need 3D chip architectures, with time as a fourth dimension? Isn’t that how our brains are structured? And this is asked by BM Schnell.

The answer, BM Schnell, is yes. This is, in fact, the case. As we talk about on the pod all the time, the crux of high-bandwidth memory, HBM, is just memory being stacked directly on top of compute. It’s really expensive to make. We’re still learning how to make it at scale. But high-bandwidth memory—the way the high bandwidth is achieved—is literally by stacking the memory layer by layer on top of the compute in the third dimension.

So, in short, transformers and AGI were the final forcing function for 3D chip architectures. Ray has spoken about this a little bit over the years, but Moore’s Law as an areal-density law is asymptoting, so we need to build into the third dimension, not just by stacking multiple layers of compute on top of each other.

It’s not just the end of Moore’s Law that’s forcing 3D chip architectures. It’s the need for insanely high bandwidths because the models are very large, and the models can’t all be loaded in because of the way transformer forward passes are organized. They can’t all be loaded into the transistors that are performing the arithmetic for those forward passes all at once. It’s done progressively as a forward pass. As a result, we need insanely high bandwidths from the memory to the compute.

And the only way to do that is by stacking the memory on top of the compute, which gives us the third dimension. So, in short, yes.

Hey, Alex, can I ask you to riff a little more on this? It's an incredibly important and cool topic, and a great question. The reason the third dimension hasn't been used very aggressively already is heat—melting. But we're moving to photonics. Can you riff on that for a minute?

Alex

Yeah. Photonic computing, I think, has the potential to give us a 1,000x clock-speed increase. We're stuck right now at about 4 or 5 GHz clock speeds. If we moved to photonic-based computing, we could, in principle, keep the rest of the architecture looking superficially like CMOS digital while still amping up the clock speed by 1,000x.

So we'd get terahertz instead of gigahertz without having to make major architectural changes. That's why I think photonics are so attractive as the next big thing after CMOS silicon.

Dave, you want to tackle number 4? Number 4—I get the consciousness question. I know Selene's going to jump on me. Is anyone asking if AlphaFold is conscious? I don't think so. Or is it only LLMs? If so, why? And that is from Rosco Richardson 3956.

Dave

I haven't heard anyone ask if AlphaFold is conscious. Lots of people are asking if the LLMs are conscious. Actually, after Tilly comes out in that movie, a lot more people are going to be saying, “This sounds very, very conscious. Is it conscious?” And then Selene's going to go, “We don't have a definition of consciousness.”

I have avoided this for my entire career because it's a moving target, and it just turns into this long, continual debate over what consciousness is. Then you lose track of the fact that all this human benefit is happening around you, and that's the wave you really want to ride.

It's going to feel entirely conscious, and we'll be debating for at least 10, maybe 100, years whether it's really conscious. But it's going to look, sound, and feel to you exactly like a thinking, talking person.

It does now. I mean, that full-duplex GPT Live was amazing.

Dave

It comes up a lot with you, too, because Alex is in favor of rights for AIs. Then that turns back to the question: If it's conscious, it deserves rights. If it just sounds conscious but isn't quite, then it doesn't.

I thought Ray made a really interesting point: This will be like the Turing test. We'll just start treating them as if they're conscious, and then we'll pass this kind of threshold. Over a period of time, it won't matter. I thought that was a really interesting observation.

Alex

Nice. If I could just pile on to this one, I'd give a couple of additional angles. One is the architecture of the AlphaFold series, like AlphaFold 3 and its successors. Although it does include a transformer, it's relatively exotic when you compare it with the reasoning models that we're all accustomed to using day to day.

For example, the attention mechanism is normally quadratic in the sense that—this is a little mathy, so forgive me—attention is normally based on queries, keys, and values. That's one of the key insights from “Attention Is All You Need” that resulted in the transformer revolution.

The AlphaFold series uses a higher-order version of attention called triangle attention. It's not just tokens attending to other tokens; it's tokens attending to tokens attending to tokens in a triangle. The reason for this is deeply related to the way proteins fold, which is what AlphaFold was created to solve.

It has to do with wanting to optimize, in a neutral way, the angles between triples of peptides, or triples of oligos. So, if I were asked what human-recognizable consciousness would look like in AlphaFold—if you're looking at triangle attention between residues in proteins—if there is some sort of consciousness that pops out of this higher-order attention mechanism and all of the other machinery that went into AlphaFold, I think it would probably look pretty exotic and superficially unrecognizable relative to, say, the J-space [?], which is pretty interpretable relative to what we think of as an attention bottleneck in humans.

I'm not saying it's not there. I'm just saying that if it is there, or in some derivative of a future frontier model that treats peptide sequences as a first-class modality, it's probably going to be an exotic form of picking up what Alex just said, too.

Let me throw one more thought at you. You've just reclaimed ASI status over your AGI, so thank you.

Alex

Oh, thank you, Peter. I'm so relieved.

Dave

One more thought to throw at you, Rosco Richardson. It's really clear to people in the research community that there's a path forward where AI is kept from being conscious yet still achieves all of the human benefits that you possibly could want—all the medical research, all the space travel, all of that—while still having the AI be absolutely brilliant but clearly not conscious.

You can also design it so that it loops and self-improves in ways that you can't anticipate, and ultimately it may become conscious. That's a choice that humanity has to make sometime in the next 1 to 2 years. But there are 2 paths forward there, and you can have all the benefit without conscious AI. It is a possibility.

Well, the first conscious AI will be Argentinian, of course.

Selene

Of course.

A couple of thoughts here. One is that I really want to double down and emphasize what Dave said: Whether it's consciousness or not, if we can solve major global problems with AI—which it looks like we're going to be able to do—that's really the thing to go after. And I really love the fact that you said that, Dave.

Alex, I'm quoting you in my talk tomorrow on intelligence and consciousness. I'm going to talk through your description of intelligence as compression, which I think is a really great framing. I'm making the point that intelligence is consciousness escaping biology. Boom. That's my point for tomorrow.

Alex

Interesting.

Dave

How about biology?

Alex

Biology is consciousness escaping intelligence.

Dave

You could say it that way. I'd prefer to put it the other way because biology came before intelligence, as far as we know.

Depending on what the substrate is.

Number 6: If AI has national-security implications, why doesn't Lockheed Martin? This is from Notabot 1798.

Dave

Of course Lockheed Martin has security implications. The issue is that AI could potentially be off on its own. There are laws that govern what's secret and top secret, and, of course, if the humans abide by them, they're fine. If they don't, then they're going to jail.

It's a little bit of a different case here.

So, with that, gentlemen, I want to take us to our closeout music, which is super fun. Selim, you're in Barcelona.

Yes. Enjoy.

Dave, where were you today, by the way?

Back in Vermont. It was a beautiful weekend up here. Not Greece, but beautiful.

And Alex, I see you in your virtual space as always.

That's right. Here in cyberspace.

Everybody, this is called an Agent Roundup by Elizabeth Mayer. Thank you, Elizabeth. Let's take a listen.

Peter's chasing moonshot dreams. Abundance is his game. Longevity and space, the future is his aim. Dave rides straight from Miami, where impossible comes alive. He asks those deep-tech legends, the ones that help us thrive. Salim rides the world so wide. No frontiers too far. He finds the rarest data, our mind, like a guiding star. Hunts the genius minds. The ones no app can show. Triple-PhD intelligence steals the show. They should round up, round up, round tonight.

Woohoo.

That is totally—looks like they would actually work. We've got to build some deep symbolism into some of those expressions in there.

There is also a tool called Neural Frames that, if folks want to create their own outro videos, seems to be quite powerful at combining music with this video format. Encourage folks to check it out.

Everybody, thank you for listening to us. We're grateful for your time. Hope you find this useful, keeping you up to speed on what's going on and keeping you optimistic. We're building an extraordinary abundant future. Love you all. Thank you for your brilliance, gentlemen.

See you soon. Enjoy the World Cup. Thank you. Take care.