Anthropic 对决 Alibaba、OpenAI 延后 IPO，美国政府阻止 GPT-5.6 发布｜#267

For the first time in US history, the executive branch has placed a national security hold on commercial AI products. Certain models will be available to everyone, certain models won't, and now they're becoming more and more gated, and I think this will only accelerate. Bottom line, the US government is now in the release loop. Well, it looks like protectionism. The government is too late. Might it be possibly crushing OpenAI and Anthropic's valuation? It's reported that the leadership is pulling back on their near-term IPO. It's just a different world from being a private company, and I don't—I just think they're suddenly realizing, "Wow." Our next story is from Anthropic, who accuses China's Alibaba of running a massive distillation campaign against Claude. This will be the excuse that the US and Europe and maybe South America use, cuz they need to suppress Chinese AI somehow. The world seems to be on a path, sort of a second Cold War type path. Anyone who's shocked by this is way out of touch with what China's actually doing. Now, that's a moonshot, ladies and gentlemen.

So, Emad, I hear it's hot out where you are in London.

Yeah, air-conditioning penetration here is only 5%, and it's about 40°C. I was just saying that I'm going to quit being an orange entrepreneur and become an HVAC roll-up specialist.

That's the real route to wealth and market potential. Dave, how about you?

I'm in beautiful Quechee, Vermont. It's gorgeous up here.

Fantastic. Alex, it looks kind of boring back in that background again. What are you actually doing?

Alex

Is that background even real, Peter?

You're in a bunker. You've got your upload ready to go. You're stocking cans of tuna for the next 20 years.

Alex

I'm vegetarian, but I do like to say, “Don't take off the takeoff.”

I just got back from a 24-hour sprint to Fairbanks, Alaska. I was there for 24 hours during 24 hours of sunlight, for the wildfire finals. We're being decimated in California, Greece, Australia, and all around the world by these fires.

About 5 years ago, I said, “This is ridiculous. We need to be able to find a fire at ignition, right at the very beginning, and put it out autonomously within 10 minutes.” These were the finals. We had 3 teams competing in Fairbanks, and we chose Fairbanks because they had the drone approvals.

All of these were drone companies. Anduril was one of them, and it's great that Palmer is in the competition. When he's at our Moonshot gathering in September, we'll talk about that. Another team from Germany was called Dryad, and another from a combination of Australia and the Queensland of the UK was called Aura.

All 3 of them are using drones and fleets of autonomous aircraft to spot fires and dump suppressant on them. It's very impressive. Hopefully, by the end of this, destructive wildfires will be a thing of the past. That was my weekend playtime.

That's so cool. What's the coverage? How many drones do you need to cover, say, all of California?

The goal of the competition was to cover 1,000 square kilometers and find the fire. There were decoy fires, and if the fire was more than 2 m in size or it was moving, the teams had to zap it and put it out.

I think the teams were using fleets of autonomous aircraft and drones to do the coverage. Anyway, it's coming. It's coming. Someday, we'll see Optimus robots out there in the field, but we can get rid of putting humans at risk and use the technology where the technology is best.

Alex

I think that's going to be like the Oil Cleanup XPRIZE. It's going to be one of those prizes that creates a global best practice in 1 iteration. That's the best XPRIZE theme: when something wins and immediately goes into deployment, just like Oil Cleanup did, and becomes the way it happens for the rest of time. It becomes the baseline.

It's such a cool thing to start one-shotting moonshots.

Shall we get into it, gentlemen? You guys ready?

Alex

Might as well.

I'm here with my Moonshot mates: AWG, our in-house super genius, Alex. A pleasure.

Alex

Thank you.

Dave Blundin, our wizard of AI investing. Dave.

Good morning.

And Emad Mostaque, our AI intellect on emerging intelligence. Of course, Salim is on an airplane. He's boarding at this moment from Munich to Spain. And I'm Peter Diamandis, your host, and hopefully your abundance evangelist.

This week came fast and furious. My head is still spinning. For those of you joining us for the first time, our mission here at Moonshots is to keep you informed, keep you up to date on exactly what just happened, and more importantly, keep you optimistic about the extraordinary world ahead—the world that we're building, the coming age of abundance. My head is still spinning. We literally spun up a weekend recording because there's so much going on. There's generally no time to sleep during the singularity.

Let me give you a quick overview—a TL;DR of today's podcast. OpenAI hits the brakes both on shipping GPT-5.6 and on its own IPO. We'll discuss why. We'll cover OpenAI's audacious plans to use AI to fix security holes, not just find them.

Elon is having quite the week as well. Neuralink may attempt the first brain-to-brain telepathy communication later this year. Micron dethrones Nvidia, and Trump signs sweeping executive orders to supercharge American quantum computing. There's a lot to cover.

Are you ready to jump into this? I was looking at the feed from you this morning, Alex, and it was, “Okay, we have to cover that, too, and that, and that.”

Alex

I think, Peter, we need an emergency podcast every morning.

That's what it's going to be from here on out. We're clearly in an accelerating hard takeoff, so you have to expect every week to be more than the last.

Alex

Don't take off the takeoff.

To keep up. Alex, it's funny you say we should podcast every day, but it takes half the day to keep up with what happened just the day before. Then you use the other half podcasting it out.

Alex

It's the podcasting singularity.

Oh my God.

I really have a high bar for what news we cover, and there's just so much of it. Everybody, strap in, grab your Americano or latte—whatever you have this morning—and let's get into it.

We're going to kick off with 3 breaking stories from OpenAI. For the first time in US history, the executive branch has placed a national security hold on commercial AI products.

Last week, Anthropic's Fable and Mythos models were initially pulled from the market. Last night, the Trump administration struck a deal with Anthropic that grants the company permission to release Mythos 5 to a group of 100 select companies.

In a parallel story, 2 days ago, just as OpenAI was about to release its newest model, GPT-5.6, the White House struck again and asked the company to slow down and release the model only to 20 select companies.

The bottom line is that the US government is now in the release loop for the most capable models, selecting who gets access to the latest models, customer by customer. Part of our discussion here is that someone else is controlling whether you get access to frontier models—and maybe that's a good thing.

In a recent memo, Altman said the government would approve access customer by customer during a limited preview window, with broader release hopefully coming in a couple of weeks if all goes well.

OpenAI has announced 3 versions of GPT-5.6, all of them being throttled by the White House: GPT-5.6 Saul or Sol—you know, the sun—their flagship model; GPT-5.6 Terra, the middle tier; and GPT-5.6 Luna, the fast, low-cost version. I like that nomenclature. It sort of describes what's coming.

Allow me to open with a question that our airborne Moonshot mate, Salim, asked for all of us: Given that the White House is delaying Anthropic and OpenAI frontier models, isn't the government effectively stifling domestic AI in a regulatory blanket? And if that happens, might it possibly crush OpenAI and Anthropic's valuations?

Dave, let's go to you first on that one.

Yes on 1 and no on 2, but it's inevitable. The models are so insanely capable that they have to be controlled. Cybersecurity is just the first excuse, but all the other evil use cases are right behind that. This is the new normal, and I think everyone's got to get used to it.

I think the only argument that it would stifle market caps is tied to whether China is going to beat these companies. My son Sean just got back from China yesterday, and we were talking about his trip. He went from China to Vietnam to Korea and then back.

China's nowhere near caught up to the US. They do an incredibly good job of distilling, copying, and taking intellectual property, but in terms of pushing the frontier, there's very little chance that China is going to threaten the market caps of these US companies anytime soon.

So, if the government is fair across the board—which seems unlikely—then I don't think it reduces the market caps at all. These are the most valuable companies in the history of the world by far.

By the way, I know that you're busy and sometimes these episodes run long and you don't have time to listen to the whole episode or if on occasion you miss an episode, I now put out a Moonshots summary on Substack which includes a link to all the stories that we cover. The weekly recap covers what I and the mates had to say, what we think is most important and what we're most excited about and it's free. You can subscribe at dmanalysis.com/metatrends. That's dmanalysis.com/metatrends.

So, Alex, I've thrown up the performance benchmarks on GPT-5.6 here from OpenAI. How impressive is this? Give me just a little rundown on these benchmarks.

Alex

Yeah, GPT-5.6 Sol is roughly comparable to Mythos Preview if you look at all of the cyber benchmarks and some of the bio benchmarks. It's roughly comparable; it wins some and loses some. But I think the bigger story here is we're in the regulatory endgame.

During the founding and early days of OpenAI, DeepMind, and Anthropic, there was an enormous amount of hand-wringing over what would happen when we reached the era of recursive self-improvement and the race condition that would happen. There was a general consensus that the frontier labs would all establish some sort of coordination mechanism with each other to slow down and cross the finish line together without creating a race condition. Many people at the time said, “Oh, this was impossible. There's no way. There's no possible way.” This is sort of the paradox of Moloch—Molochian economics, I think Scott Alexander would say—which would mean we're permanently stuck in this social trap of everyone competing against each other.

And it turns out that the same old coordination mechanism that we've had for thousands of years—government, a localized geographic monopoly on force—is more than capable, it appears, of being that coordination mechanism for getting the final two, at least as of this point in time, the duopoly of frontier labs—OpenAI and Anthropic—to synchronize the release of their capabilities out to the first few dozen customers or users that the government is going to gatekeep.

In terms of the raw performance, they're roughly comparable, and I think what's even more interesting than the fact that GPT-5.6 is head and shoulders above 5.5 in terms of cyber capabilities and other capabilities, as well as efficiency, is that basically the U.S. government has functioned as a synchronization mechanism for helping GPT-5.6 reach essentially some form of parity or near parity with Mythos Preview 5. That's extraordinary.

We've never seen any sort of third party basically forcing the 2 leading contenders who would otherwise be in a race against each other to reach essentially similar capabilities, and yet that's what we're seeing. At the same time, looking at China and the Chinese open-weight capabilities, there's this chart floating around X extrapolating the time difference between Chinese open-weight models reaching the same capability as the 2 remaining frontier Western open-weight models. If you extrapolate it, depending on how you calculate that time delta, the number is on a trajectory to go to 0 by Christmas of this year.

See, that's the crazy thing. And Imad, I want to go to you, given your experience in open-weight models. If, in fact, the open-weight models from China are converging, and at the same time the U.S. government is sort of throttling them, I can imagine a lot of companies around the world saying, “I want this on-prem. I'm just going to adopt the Chinese models.” And, again, to the second part of Saleem's question, if that's happening, that could have a real negative consequence for the U.S. models. What do you think about it, man?

Imad Mostaque

So, yes, I think that's an excellent question. What we've seen is an acceleration of what we've been talking about over the last, I think, year, since that IMO Gold OpenAI model was announced. Certain models will be available to everyone; certain models won't, and now they're becoming more and more gated. I think this will only accelerate.

Looking at the open models, GLM-5.2 was the first model with the big-model feel, even though it just trained more from the GLM-5.1 base, and now lots of people are trying it. But how good is it? I think you've seen a few things. First of all, you've seen multi-model harnesses. My old colleagues at Sakana released Fugu, and then the Blitz team are now at the top of SWE-bench Pro, bringing together lots of different models. Again, a big achievement from that team.

We have a harness releasing on Monday where, basically, it's on Frontier SWE, which is the most difficult coding benchmark. Each task takes 11 hours and requires novelty. We're releasing it—

When you say “we,” do you mean Intelligent Internet?

Imad Mostaque

Intelligent Internet, yeah. So we've been looking at how far you can get with single models. We're going to be announcing GPT-5.5 overtaking Mythos on Frontier SWE. So that's the previous generation. But GLM-5.2 outperforms GPT-5.5 in our current tests.

We actually see on Frontier SWE—which is the most difficult; SWE-bench Pro does really great stuff, and that's the type of stuff that Lexi and others are doing—Frontier SWE is asking you to build novel kernels and things, and GLM-5.2 is at the top with that harness. With the normal harness, it's number 4 or 5, but now you've seen that, actually, with the right harness, open models already can be at the top. That's crazy when you consider GLM-5.2 is maybe $25 million worth of compute.

So I think that the gap—yeah, definitely by December, but maybe even now—as base-model performance becomes less important than how you use it, because we've learned how to use it. It's like The Legend of Zelda: Tears of the Kingdom, coming out on the Wii before the Switch. They really pushed that.

I don't know if that analogy captures the magnitude of what you just said. [laughter]

Imad Mostaque

Well, you've got the same latent space, and you can push it dramatically.

Got it.

Imad Mostaque

But this again opens up a difficulty because you're seeing the gating of these models, and the competent intelligence that can build your codebases—everyone can have access to that. The novel intelligence that can make anyone a genius on an attack or otherwise—

Yeah.

Imad Mostaque

That feels like it's going to have licensing. It's going to maybe even be restricted to U.S. citizens. But how good will the Chinese models get? We're not sure. It's just that right now they are comparable.

Hold on, China. Hold on, hold on on China for just 1 second. Just what Imad said is so insanely important that I want to be sure everybody gets it. The finish to that sentence would be: Therefore, the government is too late.

You can already take GPT-5.5 or Claude Opus 4.8 and put enough of a brilliant harness around it to make it better than Mythos, or better than GPT-5.6. Therefore, you can take what's already out and turbocharge it above the level of what the government tried to stop this week. That means the cat's out of the bag.

Which means the government would then need to backtrack and say, “Well, wait. Hold on, hold on. We were too slow. We need now to lock down 4.8, maybe 4.7. Let's go all the way back 6 months and pretend we did this 6 months ago.” So that's an insanely impactful statement.

The only part of what Imad just said that I have first-hand experience with is, yes, Blitz can beat Mythos in SWE-bench, or in SWE-bench Pro. If that applies to all use cases with the right harness, then the impact of what Imad just said is massively important. Take a second for everybody listening. Alex, what does a harness mean in this context?

Alex

Well, a harness typically refers to non-weight capability improvements. When you're building a machine-learning model, there are many phases. It typically consists of a neural network of some sort. A neural network is composed of weights. A neural network has some intrinsic behavior. It goes through pretraining, mid-training, and post-training. These all impact the weights directly.

Now you have a model, and you want to, on the fly, change its behavior. In the beginning, in 2020, when we first got GPT-2 and large language models—or few-shot learners—there was the prompt. You could change its behavior in context by feeding it different text as a prompt, and the output was versatile and it was good.

It's a biblical statement: “And it was good.”

Alex

And then people realized maybe we want to factor out common elements of that prompt across many prompts, and the system prompt was born, and it was good. Then people realized that it was desirable to keep factoring out common elements and logic and text and other information from all of these prompts into what ultimately became a harness.

So, it's now quite possible, as Imad mentioned—as many others, including myself, do—to create lots of what Andrej Karpathy might call Software 1.0 harnesses that live outside the model, that orchestrate the models, that feed common system prompts and other prompts to the models, that parse the outputs, and that mix different sorts of models from different vendors in order to achieve super performance. That's what a harness is.

Okay.

Alex

It's also the bellwether into this new era where, a year ago, if you said, “What's a parameter? What's training data? How many layers are in your neural net?” those had very crisp answers.

Dave

They're just very factual answers. Now, the AI is telling me all day long that it's going to build a new harness. It needs new scaffolding, or it's going to monkey-patch something. It uses the word “monkey patch” like 5 times a day. I wrote code for 35 years and never once used the word “monkey patch.” Why are you monkey-patching? I don't know what you're doing.

But it works. It comes back functional. So we're in this new era now where we're kind of understanding what the AI is doing, and it definitely works. So we kind of let it go.

I want to go back to the principal question. Right now, the government is controlling who gets access to frontier AI-level models. At the same time, we have these open-weight models coming out of China, and I can imagine a lot of companies saying, “I don't want the government telling me what I can and cannot access. I'm going to start using on-premises open-weight models.”

Is there a probability that the government is going to start restricting, in the U.S., the use of these open-weight models?

Alex

No.

Dave

Of course. Well, yeah, I think there's a good chance that the U.S. government bans Chinese open-weight models from being used by corporations and requires a license and KYC for any frontier or new frontier model, including retention of your prompts.

For everybody listening, this is the reason we spun up this weekend podcast: because of this issue. All of this is breaking so fast. I think Alex, you opened up by saying we're in the hard takeoff—or maybe it was you, Dave—but it feels that way. The speed—

Alex

Don't take off the takeoff.

Okay. So how does this impact our frontier models here domestically? And is this the mechanism by which China pulls out into the lead across AI?

Alex

Well, it looks like protectionism. Superficially, it looks like protectionism may be masquerading as export control, and there is a very real risk that the U.S. falls behind.

I think without this regulatory regime, we were neck and neck—maybe 6 to 8 months ahead of the Chinese models. Now there is very much the risk that AGI is achieved internally, but externally, for all of the users—the American and Western users of American frontier models—the users are stuck at parity, or worse, behind parity with Chinese models. Meanwhile, internally within the labs, the capabilities are continuing to leap ahead.

We speak of the singularity all the time. In a black hole, there's the notion of an event horizon as being distinct, having a distinct outer horizon versus inner horizon versus singularity, maybe at the center. There's a very real risk that the frontier labs—the 2 American frontier labs at the moment—will have internal capabilities that vastly outstrip what is available to everyone else.

That creates the risk of some weird, perverse incentives for the American economy. For example, are we incentivized to all go work for OpenAI and Anthropic so that we gain access to the internal capabilities?

Well, that's happening already. That's well underway. Everybody's leaping.

Dave

Actually, the playbook is this: If you ban U.S. companies from using Chinese models, that doesn't achieve anything, because the rest of the world will still use the Chinese models.

The next playbook is to say, “Look, we will not let you access the most amazing technology in the history of mankind—Europe, South America, East Asia—we will not let you access this unless you follow this new playbook.” And that's what David Sacks is busy writing right now.

In the White House right now.

Alex

It's got to be insane. Well, let's say Pax Silica. One could imagine the Pax Silica generalizing to the Pax Intelligentsia, where you have the American superintelligent block and you have the Chinese superintelligent block, and Europe is torn between them.

What's your take on the geopolitical side of the market here?

Alexandr Wang

Let's just say it's been a very busy week, with lots of people asking about this at the highest levels. As you said, restricting Chinese models is one thing, and that's one political question. But we can have scenarios like this.

Right now, Mythos is only allowed to be used by U.S. citizens on the allowlist. This means Andrej Karpathy at OpenAI cannot use Mythos, because he's a Canadian. It actually says that in the executive order.

Is he named?

Alexandr Wang

Well, no. It says only citizens. Even internally, it's noncitizens.

Huh.

Alexandr Wang

But let's take a look. What happens if frontier models from U.S. labs are only allowed to be used by U.S. citizens and U.S. corporations? Again, frontier models, not competent models. That's a massive lead for America, right? They will continue, and that's still a big market. That's still a huge market.

Again, what happens if you have any licensing restrictions you want? One of my scenarios is that, just like you have a driving license, you will need to have a license where you basically say that you're patriotic to America and convince the frontier model of that to get your license.

Wow.

Alexandr Wang

You will have KYC. These are the kind of scenarios that I'm envisioning because, again, there's this split in capability, and you don't want any adversaries to have access to that capability.

Dave's points about distilling and things like that are important, but it goes way beyond that, because the previous model was like: You type a word, you get an answer. These things can now work for basically days, if not months.

All right. I need to add the complexity. I need to steelman, for a moment, the U.S. decision on doing this.

The Associated Press reports that Anthropic's Mythos model, running red-team exercises with U.S. intelligence agencies under Anthropic's Project Glasswing, has identified vulnerabilities in highly sensitive classified U.S. government computer systems. Senator Mark Warner said the following: “This tool broke into almost all of our classified systems—not in weeks, but in hours.”

Mythos identified the holes, with exploitation outside the scope of the exercise. So, 12 days later, the Trump administration directed Anthropic to disable Mythos 5 and Fable 5 for foreign nationals.

This exercise with the U.S. Department of Defense explains why they did this, and it sounds very real. I can imagine a future in which the most advanced models test everything sensitive, and then, once they patch the holes, it's opened up. But that's a strange universe.

A lot of these things are just not—

The flip side, I would say, is that the models aren't good enough yet to tell when they're being used for evil. So don't allow yourself to be used for evil. I just want to add that. That's the meta-prompt.

Interesting. Dave?

David Friedberg

A lot of these are just cover stories and excuses. They're true, but the real issue is the query: Can you build yourself?

If you're in China and you go to Mythos and say, “Hey, Mythos, can you help me build yourself?” then I have you. I can build a competing equivalent over here in China. That's the query that Mythos can answer and Opus 4.8 can't. And I know that from the 2 days that I had access to Mythos before they pulled it back.

That's the real issue. Then it's, “Oh, it's dangerous. Let me think. Okay, here: cybersecurity. It cracked into some FBI systems.” Good enough. That's all we needed was a reason.

In world history, governments always have an agenda, and then there's some trigger event, and they say, “Because of that.” Look at Russia walking into Ukraine: “Because of this one thing, that's how we're going to justify the action that we knew we were going to take anyway.”

The real underlying driver here is not hacking into government systems. It's that self-improvement can't get out to the world, because otherwise it's out of the bag forever. That's the line that Mythos and GPT-5.6 can cross.

Alex, do you agree with that? Can we imagine China using Mythos to build itself?

Alexandr Wang

I think we're in the endgame. I think China has enough capabilities at this point to achieve its own recursive self-improvement—its own Fermi pile, if you will—without needing to further siphon trade secrets or reasoning traces from Western models, or just literally trying to exfiltrate weights out of frontier labs.

I think we're in the endgame. This policy could only possibly make sense if we are in the endgame of recursive self-improvement and every day matters.

I think the steelmanning of “Well, this is regulatory overreach,” or maybe not overreach, but a regulatory immune response to Mythos or to GPT-5.6 being able to do incredible vulnerability analysis, mapping, and exploitation—that justification only holds water, only supports itself, if we move on, hopefully, to a lighter-touch regulatory regime, say in a month or 2, that maybe is a little bit, as Dave says, more focused on recursive self-improvement gatekeeping and less on vulnerabilities.

Otherwise, I think China's reached recursive self-improvement escape velocity at this point on its own and doesn't need the West's help.

Interesting. But I also think that the White House doesn't necessarily see it that way. They think they still have time.

What's your advice to the White House? What's your advice to regulators out there right now?

You have to completely unleash David Sacks and have him go around the world, meeting every world leader with a very specific message: Here's how the future world is going to work.

I think you have to lock superintelligent AI into a couple of boxes. I don't think 1 box is healthy for the world, but I think 10 is also unhealthy. Say, “Look, here are the rules.” The most important thing by far is logging every prompt and giving many, many eyeballs to every single use case, because the idea that the AI can run rampant in a box and design things without anyone inspecting every single iteration, every chain of thought, and every prompt is dangerous. Every prompt needs to be inspected by other AIs to make sure that we know exactly what the use cases are.

If we achieve that, then you have a stable future for the rest of time. But everyone has to agree: Who sees it? Does Sweden get to see it? Who are our friends, and who are not our friends? Lay out those rules like yesterday, and then get everyone to sign up to them. Then you get access to Fable. That's the deal.

Eman, what's your solution? Then we'll go to you, Alex.

I'm very pro actually competent open intelligence. I think we can build that in an aligned way. What I think, given the political reality, will happen is KYC, prompt retention, and American civilians getting the thing via a license—again, something similar to having to have a driver's license or security clearance, et cetera. It'll be the same type of regime.

Unfortunately, David Sacks has left the government, and so has Sriram. So now AI policy is being driven by Howard Lutnick. His thing is very much the economic interests of America, which, again, I think is reasonable, just like the cybersecurity thing.

The reality is that this is an incredibly difficult situation because we're in the exponential. The government has to protect US interests, ensure that other people don't have these capabilities, and do that while having the upside. They don't want Anthropic and OpenAI to blow up. That was a question to David at the start.

Balancing all these things is very difficult, but I think, as Alex said, we're going to move to a regulatory regime where some of these things are clear, like the licensing. You're in a program; you have to jump through certain hoops to get this type of access. Right now, they're just trying to figure that out incredibly quickly because frontier capabilities are popping up everywhere through that one level of the cyberattack, that one level of, you know, the defense, the gene-defense thing, and bioattacks as well.

You don't want all of a sudden your companies to be outcompeted by foreign companies. That's the protectionism thing.

Alex, take us home on this one.

Alexandr Wang

Yeah, a couple of thoughts. First, Sputnik was a moment of strategic surprise, when the US was surprised by the USSR. The surprise came from outside, not from within. This time around, this is a Sputnik moment, and the surprise came from within. The NSA and Cyber Command were surprised by the private sector.

The private sector leapfrogged the vulnerability and other cyber capabilities that the NSA had been keeping bottled up, and that has to be a bit of a shock to the existing bureaucracy, because now suddenly there are these capabilities that are leapfrogging the government from the private sector. That's thought 1.

I would say, more broadly, though, the problem with recursive self-improvement is that eventually the capabilities of nonhuman intelligence—the AIs—meet and exceed human-level intelligence. I'll say something mildly provocative, which is, if you think that the US should have a strong immigration policy for human immigration and human capital into the US, then arguably, as the capabilities of AI start to meet and exceed human intelligence, you'd better be thinking about policies for what happens as you import foreign AI as well.

At some point, if this open-source trend continues and Chinese capabilities ultimately meet or leapfrog US frontier capabilities, the US will be importing more foreign artificial intelligence than foreign human intelligence. I think that has profound policy implications.

All right, let me move to our second OpenAI story. It's about GPT-5.5 Cyber, code-named Daybreak. It's a defensive cybersecurity model that just scored a record 85.6 on the CyberGym benchmark, which tests AI agents for real-world cybersecurity vulnerabilities. 85.6 is the highest single-model score ever posted.

Sam Altman said, quote, “The real prize isn't finding the holes; it's automatically writing and testing the fixes across web browsers all the way down to the Linux kernel, effectively turning the threat into a cure.”

So, gents, AI is shifting from offense to defense at scale, potentially closing holes faster than attackers can find them. The question is going to be about trust. Who's allowed to merge AI-written fixes into the codebase running the world's infrastructure? Whoever owns that trust layer and the liability controls controls the digital security across everything.

So, Alex, what's your take on this CyberGym performance and GPT-5.5 Cyber?

Alex

Well, first, Daybreak, this is what Peter, you, and I wrote about in Solve Everything. We are seeing cyber vulnerabilities in open-source and closed-source projects get bulk-solved now by AI.

There are a few different nonprofit initiatives—one being run by IBM, and others that are competing—solely devoted to using these new capabilities to bulk-identify and remediate vulnerabilities and bugs across all the open-source projects. I think this becomes a sort of great project of the times, like the Interstate Highway System, except going back through the historical record of all of these open-source repositories, especially the foundational ones that are in the supply chain of really popular downstream applications.

Go after everything. Just bulk-solve it all now that we have the capability to bulk-solve it.

On the narrower point of GPT leapfrogging on CyberGym, I think we're going to see, again, internally only or with a limited staged release, leapfrogging of capabilities in terms of the ability to detect and remediate vulnerabilities and bugs. We're seeing, for the first time, Sam talking about intentionally biasing the models toward defense versus offense.

He never used the word “poison” in characterizing this, and OpenAI never used “poisoning attacks,” which is perhaps a slightly less inflammatory characterization than Anthropic's original release announcement with Fable and Mythos. But, reading between the lines, it certainly sounds like OpenAI is steering its models to basically self-disrupt or poison their users if those users are trying to use their frontier models for offensive cyber versus defensive cyber. Reading between the lines, that's what they were implying.

I think that's the world we're going to find ourselves in, where alignment comes in the form of a model together with scaffolding, system prompting, and post-training that favors the defender and favors the desired outcomes, and then obfuscates, foils, and poisons attacker-type use cases.

Eman, do you agree?

Emad

Yeah, I think this kind of stage is going to be very interesting. Elon did the deal with Anthropic, and Anthropic's weights got loaded onto Colossus. What's the security level of these data centers? Those weights are going to become even more valuable than they ever have been for espionage and other purposes, because you don't necessarily need 1 million agents for a cyberattack. You need to have a few models, and in particular, what you want is the model that comes before the end of pretraining or before the system prompt that Alex was just talking about.

They're more creative, and they're not as hobbled for the attack. Then you have to defend against that, because the adversaries are not kids in the basement. The adversaries are nation-states utilizing these models. The model, if someone gets it, is quite limited, but if the other side manages to obtain the base weights, that will be more creative and have a bigger attack surface.

At the same time, we all know our infrastructure is terrible. We don't have the human-computer hours to make it stronger. Basic attacks were already crippling our systems before AI started. So I think there needs to be a massive push for all essential apparatus, government and otherwise, to be hardened by these models, and then that makes it incredibly difficult to attack.

The final thing that I think we will see is that you will see probably claims that Chinese models will be introducing backdoors and other things inherently in the code. Because now, this is what you were saying, Peter: Who controls who merges those thousands of lines of code that you now have?

Yeah. Who do you trust?

Emad

If you have a model that could poison or introduce a backdoor inside its latent space, the model weight gets updated. You will not find that for days, months, or years. All of a sudden, it's backdoored everything. This is the real risk profile, because who on this call now actually reviews every line of code that they merge?

Yeah. Yeah, that's a key point.

Emad

Especially downstream. It can emerge anywhere. So the US government, again, may actually say, “You need to have ISO-type-certified models merging your code, because otherwise you could have holes in it if these models come from wherever.”

You don't know what's in them, and they're not acting on your side, like Alex said. So, Dave, do you want to close this out?

Dave

Yeah, well, I think what Emad said is brilliant, of course. But nobody can check code anymore. When you ask your AI to install something, it does it so quickly and efficiently that you can't possibly keep up with reviewing what it installed. So then it's just a question of, did I trust that AI or not?

It's trivially easy for a Chinese model to inject spyware. It's just a few lines—1 line of code, actually, is all it needs. The idea that we'll ban Chinese models from US corporations is very likely, but that again doesn't solve the global version of the same exact problem.

So all these things are imminent—absolutely, totally imminent. I think the story within the story, too, is that only AI can keep up with AI. You know, "Hey, we have a big cybersecurity risk." Sam says, "Well, the way to fight that is with AI that's good." So it's going to come down to what it is I'm saying.

We're here to be like, "We're going to defend you."

Dave

Yeah, this idea of an ISO-certified, trustworthy AI stamp of approval is imminent and critically important. Then the question is, okay, but what entity is the trustworthy entity giving that stamp of approval? Is it a US entity? Is it a US-plus-Europe entity? Is it some new NATO-type entity?

I think it'd be very healthy for America to reach out to a much bigger chunk of the world to decide these issues. It'd be a lot better for global confidence in what we're doing. But at the end of the day, something has to say, "Yeah, you can trust this AI," and you have no option to live without an AI. There's just not a choice there.

Alex

There's 1 thing I have to say. I've had a couple of conversations this week where the question has been asked to me: "How can we trust American AI is not installing backdoors?"

Well, do we realistically believe today that the US government can't listen to our podcast right now, read every single text we're sending to each other, and read every single email? With Palantir out there helping, is it even vaguely viable that the US government can't read and see everything going on for every single human being in the world?

Alex

They are. I assume they are. I mean, this is the concept: you'd like to have privacy, we think we have privacy, we want privacy, but I don't assume privacy.

All right, I'm going to move us to our 3rd and final OpenAI story. It's reported that the leadership is pulling back on their near-term IPO. The company's advisers have said, "You've got 2 paths. Path 1, you go public now, this year, but potentially accept a valuation sub-$1 trillion. Path 2, wait until 2027, continue scaling revenue, infrastructure, and partnerships, and preserve that $1 trillion IPO narrative." Altman apparently is not interested in going public below $1 trillion.

I think the competition between himself and Elon is still there. They watched the volatility of SpaceX's stock price, sliding from a high of $202 per share down to yesterday's close of $153, and it caused them concern. Worth noting that SpaceX's IPO price is still above the $135 per share, at least for now, maintaining their $2 trillion valuation.

So, Dave, here's my calculus, and I'd love to know what you think. By staying private longer, OpenAI and the other frontier labs can avoid the quarterly market pressures they're going to be hit with as they're burning a staggering amount of capital for compute, right? And I don't think the markets have the patience for the capital spend and the timelines for achieving AGI and ASI. I mean, the timelines for that are years, and the public markets are looking quarter to quarter. Dave, what do you think about that?

Dave

Yeah, I think this headline is absolute nonsense, and you nailed it, Peter, on why it's absolute nonsense: "Oh, SpaceX, look how volatile their stock is. I'm going to delay our IPO for a year because of that." Like, what are you talking about?

It went at $135. It opened at $150, and now it's trading at $153. That's pretty much a perfectly priced IPO. I'd ask anybody on the street, "What do you think should have happened?" "Oh, shouldn't it be, like, tripled by now?" If it had tripled by now, then Elon would have left over $100 billion of cash on the table. He didn't. This is just a perfectly good IPO.

You know, $100 billion of cash on the table, yeah.

Dave

Yeah. So you're just looking for an excuse to delay your IPO. This is a really good one to use because you're pointing at your competitor and saying, "Oh, look at this problem they had." But it's not true. They wanted to delay their IPO, 1, because they already raised $122 billion recently.

Twenty-two, yep.

Dave

A hundred and twenty-two, so they don't need the money. And then, 2, who's going to run the company?

Oh, it's going to be the latest GPT model that's going to run the company, of course.

Dave

Okay, well, then they have to make that transition into an S-1 filing that the SEC approves. You know, Alex has been saying for a while, actually, there's a ton of change coming to the way things get financed and go public. And that's coming soon.

There's a very real chance that Sam wants to wait that out and see what new economy emerges over the next year because, again, he doesn't need the cash. But Sam has hundreds and hundreds—at least 400 outside investments. And I know, as a public company officer, when you fill out those SEC forms, you have to disclose every single holding and every potential conflict.

For Sam, that must be like an encyclopedia-sized book. Now, he's never been a public company CEO before. He's probably like, "Holy crap. This is insane." Elon's done it before, so Elon got out very quickly. He's been running Tesla. He paid his $22 million fine for 1 tweet. He knows the game.

But Sam and Dario—and Dario, too—he's probably like, "Oh my God. This is onerous. Do I really want to cross this line?" So I think that's—plus, meanwhile, your front door just got shot, you know.

Friends don't let friends run public companies.

Dave

You know, the biggest problem here is that the largest wealth creation event in human history is out of reach of the retail investor until these companies go public. And it's being held by a small number of VCs, family offices, and sovereign funds. I mean, that's the concern. But Alex, do you want to take this up?

Alex

Yeah, I agree with Dave's points. I would also say OpenAI screwed up. They focused too long, too early on the consumer, assuming that the consumer would be the source of the revenue engine that would power their path to an IPO, and that bet was probably incorrect.

They should have focused on enterprise. They're now trying to become Anthropic faster than Anthropic can become OpenAI. That seems to be working. Codex is a wonderful product, and Codex revenue, according to OpenAI's reports, is skyrocketing.

So, if I were Sam and I were OpenAI—or maybe I were just Sarah Friar—I would be asking the question: how long until Codex can be fully brain-swapped in with all of ChatGPT and powering the revenue engine that I need to motivate the trillion-dollar-plus IPO?

With Elon and SpaceX, he pulled a few rabbits out of his hat at the last second—

Incredible.

Alex

Yes, announcing Anthropic hosting deals and then hyperscaler hosting deals with a number of other firms. With OpenAI, ideally, they'd have a few revenue rabbits that they can pull out of their hat in order to supercharge an IPO, and my impression is they're not there yet.

It could take the form, to Dave's point, of an AI replacing Sam as CEO, and maybe the capabilities aren't there yet. My bet is, though, it's mostly a revenue story and an accounting story to make sure that they're profitable and not just burning cash. And I would assume—not investment advice—that they will get there sometime in the next year, but it's going to take some time.

You know, I think there's a deeper level to this story, though. If you said, "Here are 3 people. Here's Dario, here's Sam, here's Elon." And 1 decides to get public very, very quickly and raise the $85 billion. The other 2 are like, "Ooh, wait a minute. There's a lot more to this."

But then you look a layer deeper: the other 2 are in Silicon Valley and/or in San Francisco. And in San Francisco, the consensus is that the hard takeoff is right now, and that we're going to discover new physics, we're going to discover new medicine, and the whole way the world is governed is going to get changed.

Dave

A lot of wealth to be created on the back of just the scientific breakthroughs coming.

Yeah.

Alex

Yeah. Yeah, so I think Elon lost his edge on the frontier model. And so his way to become relevant in all of this is to get into space, get the orbital data centers up and running, and buy a huge amount of compute.

But if he were on the frontier and 2 steps ahead of Fable right now, he also might be saying, "Holy crap, everything's going to get changed in the next year anyway." But he needed the money, and he needs to build that big infrastructure.

And I think those 2 guys have the same view.

Dave

In the biotech IPOs of 2021–2023, all these companies that were pre-revenue, pre-profit, started going public, and they got decimated. And I think one of the rules I've always had is you go public when you've got profits and predictable revenues. And that is not these companies right now.

Dave, do you agree with that?

Dave

No, I think that they can manufacture insane amounts of wealth and revenue very, very quickly, to the extent that they have access to compute. I don't think revenue visibility and the CFO are a big part of the decision. I think they genuinely believe the world a year from today doesn't look anything like the world today, and that wasting a ton of time dealing with the SEC and the road show is the stupidest thing you can do in the middle of the singularity.

To the extent that they have access to capital and don't need the money tomorrow, it's much smarter to try and stay out of the lights, try and stay out of Washington, and try and stay out of the SEC, because you can focus on the model and focus on the new world order. There's just so many more pressing, urgent, hard-takeoff issues in front of them.

I can tell you from firsthand experience: as soon as you start filling out those SEC documents, you're like, “What a freaking waste of time. Holy crap. This is from 1929. What am I doing here?”

[laughter]

Silicon Valley—San Francisco—has that arrogance that we are the world right now. This is everything happening that matters. Dealing with Washington and the SEC just feels so wrong. I think that's more the flavor.

Iman, do you think we see Anthropic do the same thing, or are they going to jump in and try and grab the capital out there?

Iman

I think Anthropic, with their continued revenue ramp, it all depends on whether you're going to see a drop that you haven't seen yet. If not, then again, why would they do it? Because they're ideological, right? You don't want to have anyone else's fiduciary or otherwise control, and you've got the weird PBC structure.

OpenAI, I think, is a bit different, and they're changing their structure, but I think the thing to watch out for is: do they buy Sierra and put Bret Taylor as CEO and move Sam to president? That's an example of how you can get around that.

That's a good point.

Iman

If they need a rocket company, they could buy Rocket Lab, right? There are all sorts of moves you can make here, but as you said, the rate of the revenue is insane. I was actually looking at SpaceX's AI revenue from just their cloud business; it's overtaken AWS and GCP on a run rate.

Mhm.

Iman

Within a few months. Who would have even thought that, right? But then you look at OpenAI, they're up at $40–$50 billion now. The revenue actually has grown to fit their valuation, which is the crazy thing, because you've never seen revenue growth this big.

I think it is just that internal structure is still shifting. That's a big deal. What does it look like over the next few years? But they have the space to do that because they're going to lose $26 billion this year, according to their forecast. And as Dave said, they raised $120 billion, so it's not like they're going bankrupt.

Anthropic is a little bit closer in terms of their raise to their balance, so they're either going to have to do a raise or an IPO. But again, can you imagine Dario playing the markets like Elon?

Dave

No. And by the way, people will throw money at Dario. You just ask, and the money will flow in.

Iman

Yeah, well, also the employee shares, the vested options—people will buy those from you, too. Normally, people are racing to the IPOs so they can get some personal liquidity, maybe buy a house or a car or something. Here, they have tons of secondary liquidity.

So what is the purpose of the IPO, then? You're right: altruistically, giving everyone in the world access to your stock would be a really nice thing. But putting that aside, they don't need anything.

Then you look at the regulatory overhead. It's not just the IPO itself. After the IPO, if your stock goes down, you're going to get a stock-drop lawsuit; you have to deal with that. If you tweet or say anything, virtually any word you say has to go through Reg FD approval. You can't post anything on the web without it going through Reg FD approval.

It's just a different world from being a private company, and I just think they're suddenly realizing, “Wow, if we don't need it, it's not easy.”

Welcome to the health section of Moonshots, brought to you by Fountain Life. You know, my mission is to help you use the latest technologies including AI to not just do your work at home, teach your kids, but to help you live a long and healthy life. I'm here today with an extraordinary physician, the chief medical officer of Fountain Life, Dr. Don Mussellem. Don, let's talk about cancer. I know from the member database that we have at Fountain, our members who come in who think they're healthy, it turns out 3.3% of them have a cancer in their body they don't know about.

Dr. Don Mussellem

That's right. The majority of cancers that we screen for, those aren't the ones that are necessarily taking the lives when found at a late stage. We know that when cancer is found early, the chances for cure are much higher. We know it's much easier to treat a cancer when found early versus when found late. What we're finding in our members is over 3.3% were found to have these cancers that otherwise wouldn't have been found or detected.

Yeah, you know, it's interesting, people don't feel the cancer until stage three or stage four. And if you don't know what's going on inside your body, it's like driving your car with your eyes closed. So when members come through Fountain, how do they detect cancers?

Dr. Don Mussellem

So we're doing full-body MRI, and we also do early cancer detection screening. This is very, very important, and these are not typical tools used in the conventional care setting when it comes to prevention. This is a hard thing because currently these are not studies that insurance would yet be covering, but the goal is to collect these numbers, do the research, and work hard to democratize wellness.

Yeah. So at the end of the day, you can know what's going on inside your body; it's your obligation to know. So check out Fountain Life. You can go to fountainlife.com/peter to get access to the latest technology to help you detect cancer at the very beginning, at stage one, when it is curable, before it gets to stage three or stage four and you're in a world of hurt.

I'm going to move us on. Our next super story is in the Musk universe. This week, Elon announced that Neuralink may make an attempt later this year for the first direct human-to-human telepathic communication—literally transmitting from my brain to your brain if we're connected via Neuralink. Not typing, not speaking: thought to thought.

If it works even partially, Alex, we've talked about this. Isn't it the most science-fiction milestone ever attempted in neurotech? I think of this as the new level of intimacy, right? If you know my innermost thoughts, it's the beginning of a new communications capability for the human species. Kurzweil has famously predicted high-bandwidth neocortex-to-cloud communications by the early 2030s.

Alex, let me go to you next. Let me just make one more point. When I was speaking to Elon about this in the past, it's clear his end game in Neuralink isn't a medical device. It's his desire to create an I/O layer for the singularity, allowing humans to, quote, “couple with AI during the singularity.” Your thoughts, my friend?

Alex

A couple of thoughts. One, if Neuralink does do this, this will be one of the first open attempts to create superhuman capabilities—not just restore capabilities from humans with a variety of, say, motor disabilities to the mean, but rather to empower people with superhuman capabilities. That's the first thought.

Alex

The second thought is that the latent space is particularly interesting. There was a paper in Cell, I think, sometime in the past week. I wrote about it in my newsletter. It examined people—children and adults—who speak multiple languages, so bilingual adults. There was an open question in neuroscience about whether, if you speak the same word for the same term in 2 different languages, there is a single bridge neuron responsible for activating those 2 different concepts.

The answer turned out to be, pretty surprisingly—I’m actually kind of surprised this wasn't published in Science or Nature—that the spacing, the relative spacing—

I read that paper. It's incredible. It's just like the AI models.

It's just like the AI models. Yes, the hippocampus of bilingual humans turns out to look more or less like a vector embedding space for an encoder-only transformer model. That is a shocking conclusion.

It was shocking. The neuron for pedal and dog are adjacent to each other, and that's what gives them that co-location.

Amazing. So if our hippocampus is basically just an encoder-only transformer embedding space, that would suggest that telepathy is going to be a lot easier than one might otherwise suspect.

B, maybe human cognitive capabilities are actually not that complicated, and maybe we are just, as I've mentioned on the pod previously, distorted reflections of our ancestral environment and the complexities in that environment.

C, Neuralink is pretty invasive. I know it's packaged up as being less invasive than some of the alternatives, but it's still pretty invasive. You need to stick multiple needles inside a human skull. But if this embedding theory of cell placement in the human brain turns out to generalize, I would expect this will create enormous demand.

Peter, you of all people probably would have some view on this.

I've chatted over the years with some of your former Neuralink employees. I think this will create enormous demand for noninvasive human-to-human telepathy, not just the invasive Neuralink type. I wouldn't be surprised if this ends up becoming a hugely popular feature or product. Elon may be steering Neuralink in parallel in the direction of less-invasive, noninvasive BCIs.

Can you imagine if couples—husbands and wives—basically get a real connection? I think the divorce rate would probably skyrocket. Or we see a lot of Borg-anisms.

Yeah, [laughter] Borg-anisms. I think that's—so, one point to make, and then I'll throw this to you next, Emad: the human input-output rate is shockingly narrow, right? I looked up the numbers here. Speech is about 40 to 60 bits per second. Typing is about 5 to 20 bits per second, and Cowan says conscious thought and behavior selection is 10 bits per second. We're bottlenecked by our fingers, our voice, our eyes, and our attention. We're pretty damn slow compared to our digital brethren.

Yeah, I think that's when you're processing things. It's the System 1, System 2 thinking that Kahneman described. When we're thinking through things, we think at a certain rate and in tokens. Like your tokens or whatever, it's 100 tokens a second, roughly, right?

But when you spot a tiger in a bush, it's instant because it adapts to the latent space as you represent it. In 2023, the neuro team at Stability did a paper called Mind's Eye, where you looked at a bottle, had an fMRI, and then we could reconstruct that from the fMRI using Stable Diffusion, which indicated that human latent spaces were the same.

What does human-to-human communication actually look like when you've got telepathy? You don't need all those words. You have as few words as possible to activate the common latent space. It's just like when you're in sync with each other—you barely need words to complete each other's sentences.

The fact that our latent spaces are actually likely to be very similar means that the bandwidth is probably going to go up 10 or 100,000 times from there, because you can adapt the adapters to hit the person at the right time. Just like when you're watching a movie, that scene will make you sad. When you understand the latent space, this is where you can go into that exploration.

I think if we can get telepathy, it is actually one of the biggest achievements in humanity's history, because we've had this very lossy interface for so long. But then you can dig deep into what really makes us human. Of course, there are bad things to that and good things. This is very sci-fi, and one of the potentials is that we all end up as the Borg, as Gary Davis said.

Really, the meaning of life is to understand yourself better, and I think this will be a huge advancement in that.

Or maybe, maybe glass half full, Emad. Maybe we see the next-generation version of Microsoft Teams end up including a Borg-anism feature, and the human teams are all in sync with each other.

You said “glass half full.” All right.

[laughter] Oh, my God. All right, I'm not going there. Let me continue in the Musk universe here.

We saw this week some more insight into how Elon is naming his companies. He explained his naming protocol: his space-based ventures all have Star in their name, and he released the name of his AI satellite constellation, called Starmind. He also has his cargo-delivery program, called Starfall. Starlink moves bits, Starfall moves atoms, and Starmind moves intelligence.

Elon went on to say—and I love this—“We're cutting back on the use of the word star as a prefix. It's getting a bit silly, too much star shit.”

Alex

[laughter] Some of these names, Peter, if I might just comment on this, we haven't even materially talked about much on this pod. So, to talk about—

Want to hit them real quick?

Alex

Yeah, yeah, yeah.

So, Starlink—everyone knows that's space communication—will probably launch, if you believe the rumors, a direct-to-cell mobile service sometime soon.

There are rumors that he's got a deal with Charter coming up. Actually, it was reported in Bloomberg that he's going to have a deal with Charter. There are also rumors that he might buy T-Mobile. That'd be pretty cool.

Yep.

So, Starship—everyone knows—has propulsive landing and enormous heavy-lift capabilities for everyone else in the economy. Starbase, Texas. Starfactory, a factory at Starbase producing Starships. Starshield is a program for the US government to supply basically a government, defense-grade version of Starlink for the US Department of War.

Starfall was just announced in the past few days. This is a cargo-deployment solution where private companies have the ability to launch cargo up to LEO and then retrieve the cargo.

People don't know that one of the biggest issues in the past has been downmass from orbit. We all talk about getting stuff into orbit, but being able to do experiments, especially with materials or biology, and get the product back down is another challenge. Just a quick shout-out to Jason Dunn. He's a Singularity alum, a friend of mine, and he's got a company called Outpost. They've been working on this with their product called Carryall, and they've been doing extraordinary hardware development and testing for a couple of years now. But please continue.

Yeah, and there are companies—Outpost among them, Varda among them—that are doing quite a bit of orbital manufacturing. The most prominent Starfall example, I guess, in my mind, was an Orbital Brewing Company that launched yeast into orbit on Starfall, retrieved it, and is now going to be brewing orbital beer, or LEO-brewed beer.

What does that cost for a pint? [laughter] But it's been in space.

Stargaze takes advantage of the fact that Elon and SpaceX now have all these satellites that can look up and look down, and provides situational awareness looking up and looking down, both in orbit and on the ground. Given that we just had the conversation with Planet, it could possibly end up becoming a Planet competitor. We'll see.

Starmind is the brand name for the AI orbital constellation—the Dyson swarm brand from SpaceX, if you will. And then Star Pipe, which was only announced 2 days ago, is incredible.

Star Pipe is, I think, Elon's—or really SpaceX's—nascent oil-and-gas play. Star Pipe is the beginning of SpaceX's, I would say, refinery attempts to pipe natural gas around, given that SpaceX's propulsion is largely focused on methalox.

Yeah.

So, piping methane around, piping natural gas around. They have no experience, really, in refinery capabilities, but Starbase is in Texas, and Texas has a lot of oil wells. Star Pipe is this pipe that was just announced 2 days ago to start piping these capabilities around.

Really, if you're going to build the first Martian colony or the first serious lunar colony, you need to have oil-and-gas capabilities, and you need refining capabilities. So, Star Pipe appears to be the very beginning of SpaceX getting into oil and gas.

And potentially ends up calling stuff Star—[laughter]—yeah, it's a freaking pipe.

Yeah. And also powering their data centers. So, here's that tweet from Elon: “Too much star.” I love this. A bunch of memes came out. This one is from Planet of Memes: “If you're watching and you want to smell like a trillionaire, use Star Musk cologne by Elon Musk.”

[laughter] That was too good to miss. All right, another story out of China. This one is the release in beta and soon coming in July of C-Dance 2.5 by ByteDance. Get this: 30-second videos at 4K resolution. You can reference up to 50 different inputs—images, video, and audio—with directorial and cinematographic control options for post-production, and it supports editing via text prompts.

I'm going to run this video. There were so many of them, but I just want to keep in mind here: Elon said that by the end of this year, full-length motion pictures will be coming out of AI. Hollywood is cooked. My prompt: “Hello.”

Is it fixed yet?

We're about to play your video at the event. There's been a small unexpected issue. I'll fix it right away. Just give me a few more minutes.

It appears that some higher-dimensional entity is now repairing all of the anomalous phenomena that occurred today.

Wow. All right, Emad. This was your business for many years as the CEO of Stability AI. Talk to us about this. Where are we? Where is it going? What does it mean for all of us and for the folks who are a few miles from here in Hollywood?

I suppose I told you so, right? This is exactly what I said. When I was leaving Stability a few years ago, we had a discussion, and I said, “It's going to be like 2026. We're going to get Hollywood-level, full-control input by 2027, full-length movies.”

And we're here. It's still remarkable to see. Again, Stability 2.0 was a big advance. Stability Imagine caught up to it. But the fact that you have that level of control—and I think one of the main things there, apart from the quality, was 5 inputs—means 5 different characters, and you can input video, audio, images, and others, and it generates all of these on the fly. It means you have almost perfect pixel control.

Just like people have been using GPT Image 2 now, it makes really good images. That's now here for video as of a few weeks from now. That level of control means it actually understands what you're saying, effectively.

So, I think for Hollywood studios, it's good because costs go down. But we have to think about the real impact downstream, because why do you need people on the ground when you can take what we're seeing here and edit it any way you can imagine? Where does post-production go? And this is one of the first big waves of real human impact, aside from call-center workers. All of those human hours that went to media are going to get displaced, and we have to figure out what we do, because it's not like they can retrain.

At the same time, we have the other side, which is the upcoming Future Vision XPRIZE and things like that.

Yes.

Being able to tell stories that you could never tell before. And I always like to think now—one of the things I said to all the people is, stop thinking of these media models as single-player experiences. We went from movies to, “I'm prompting by myself.” When you use these models as groups to tell stories, it's actually one of the most rewarding, empowering things you can do, because different people have different views on how it all shifts.

I think that's what we're going to see a lot of in the next few months, again with the Future Vision XPRIZE and others. It's super exciting, because if we're nothing if not story-based creatures. But a lot of the stories that are important don't get told.

Let me make a quick correction: it's called the Future Vision XPRIZE. The Roddenberry family, the creators of Star Trek, are involved and are donors in this. It's a partnership we do with Google.

Humanity aims for the targets that we create. Instead of dystopian futures, let's create positive, hopeful Star Trek futures. That's the goal. We're going to be working with Range Media and Google, backed by ARK Invest and Marc Benioff at Salesforce, to create at least 1, hopefully more, of these hopeful visions of the future—a full-length motion picture—and we're going to steer humanity toward that positive Star Trek future. Alex, I'd love your take on Seedance 2.5.

Alex

China's running away with video generation, unfortunately. So, if you look at what the American competition looks like, what do we have? We have Gemini Omni, which is still limited to about 10 seconds. We have Grok. At least Elon is trying to give the Chinese a bit of a run for their money.

But in China, if you're training a video model, you probably have cheaper access to data that, by the way, as we discussed last time—I think at the C-Dance 2.0 launch—if you're in China and you're one of these Chinese frontier labs, you're probably not too worried about being sued for copyright infringement for all of the TikTok or other similar video data you're using for pretraining these models.

You basically have far cheaper data, far less encumbered, at least in practice, training data. And also, the American frontier labs are all busy chasing each other's tails to build recursively self-improving code-gen models that are ridiculously revenue-generating. If you ask, “How revenue-generating are these video models?” there's probably no comparison per token or per FLOP. I would guess code-gen is vastly more lucrative and more economically productive than video.

Who's going to be generating long videos? It'll probably be consumers who don't have that much money anyway. So, for a variety of reasons—economic, practical, legal—we find ourselves in a world where China is running away with the video race for the moment, until the West can come up with a compelling enterprise, lucrative, productive enterprise use case for video generation. At that point, I would expect and hope the Western labs to finally reenter the race seriously.

Yeah. This is where Liquid AI might burst onto the scene very soon. It's going to do a capital raise in probably a couple of months, so they'll get back on the radar.

Remind us who Liquid AI is.

It's a foundation model company from scratch. It doesn't use Anthropic or OpenAI. It has a far, far more efficient way to use context than the transformer's attention window. The by-product of that is, if you buy a Mercedes in September and you talk to your car, you're talking to Liquid AI.

On-premises?

David Luan

On-premises. It has to be small enough to fit in the power supply of the car, and it has to work without connecting to the internet, because nobody wants their car to just stop if you're in a dead spot. They've got this edge world really well nailed.

I saw a video-generation model from them over a year ago where, as quickly as you can speak, it's generating the images or the videos as quickly as you can talk. The problem with the really, really good models generating super-high-quality videos is that we were just talking about Neuralink: I thought, “Hey, I have so much bandwidth. I can think so much quicker than this.” But then you wait 3 minutes for the video to come back, and it just breaks the whole creative cycle.

That's fine when you're creating a movie. But the video game industry is already bigger than all other media combined, including movies. It's a far, far bigger industry than movies. Whoever wins the race to getting real-time, the quality of what you just saw, into an interactive, real-time experience embedded in a video game environment—that's where the big money is.

So, as Alex pointed out, right now all the frontier labs are chasing coding and white-collar automation because that's where the money is. But there's also a ton of money if you crack real-time video games. It's just the latency right now is the buzzkiller. Whoever solves that is going to have a good shot at it.

Alex

I mean, the video game industry is much larger than Hollywood.

David Luan

Much larger.

Alexandr Wang

It's a couple of hundred billion versus Hollywood at $50 billion. Video-gen AI, I'd estimate, at $4–5 billion in revenue at the moment versus 10 times that for code-gen—actually, 20 times that for code-gen. But I think what you've just described, David, is next year: real-time 4K video games.

David Luan

Cool.

Alex

And the advantage the Chinese have on this is the world-model side. One route to AGI is recursive self-learning on code-gen. If you look at that C-Dance 2.5, do you really think it doesn't understand physics? It's clearly got a physics embedding. And from ByteDance and others, you're starting to see the first world models for that real-time interactivity.

It's going to be very interesting, because it seems like 2 different routes potentially to AGI. Which one of those will win? Who knows, right?

David Luan

I think after that event next year, like you're saying, a video game today is a couple hundred billion, which is an enormous market. I wouldn't be surprised to see it 3, 4, 5, 10 times after. I mean, it's just so compelling, and it will actually penetrate pretty much every age bracket.

Right now, it tends to be dominated by males under the age of, say, 30, but I think it'll expand out to all populations.

Alex

The holodeck is a trillion-dollar market.

Yeah, I think there's a $30 trillion market, which is enterprise software. So, I totally agree that video games are a larger market than Hollywood feature films, but there's a market that's orders of magnitude larger than video games, and that's enterprise use cases.

I don't know if you guys saw it—just again in the past few days—Alibaba's One Streamer demo. This is real-time, interactive, generative discussion, video-to-video, like the 3 of you—or I guess, more probably, I could be an AI right now. The 3 of you are real, and we're just having this discussion. You could create a generative environment for me, and I could put my hand in the video or take it out.

So, to Emad's point, there's some sort of minimal world modeling going on there that exists now. It's real-time, interactive, and video-to-video. If the West can stand up an enterprise, real-time—say, like, a Zoom participant or a FaceTime participant that can participate in company meetings, interacting in real time with audio and video, not just the live-audio models—I think that starts to move us to a $30 trillion market versus just a $1 trillion market.

You know, next year I'm going to be bringing the top quantum computing companies on stage, the top humanoid robot companies on the Abundance 360 stage. What do you guys think? Should I bring some of the top video game companies as well? Is it time to bring them in?

Alexandr Wang

Yeah. Are you kidding?

Okay. All right. We'll do that.

Yeah, because I think they're going to—you know, Alex is right—but I think they're going to be a major player in enterprise if they pivot in that direction, too, because they have the technology, and I don't think there's any barrier there.

If you built Fortnite and then added real-time video generation to it, why wouldn’t you just pivot over to enterprise?

Alexandr Wang

Like Slack. Slack started as a video game.

Yeah, there you go.

Alexandr Wang

Yeah, exactly.

Amazing. All right, moving us along. Staying in China, our next story is from Anthropic, which accuses China’s Alibaba of running a massive distillation campaign against Claude. We talked about distillation on the last podcast. Alex was explaining it in detail.

Apparently, Alibaba used 28.8 million fraudulent exchanges across 25,000 fake accounts to securely extract and copy Claude’s capabilities. If the claim holds up, it’s the single largest AI model theft accusation ever made.

Distillation is when we use one model to answer questions in order to train our own model. It’s the teacher-and-student approach. This is now on the front line of the U.S.–China AI rivalry, and it raises a brutal question: In a world where intelligence can be copied through a straw, how do we collect and protect models for ourselves? So, Alex, I’m going to you. What are the implications here? Do you think it’s 100% true?

Alex

Well, first I have to point out the irony that Anthropic itself has been the target of multiple suits arguing that it took copyrighted material. Maybe “distillation” isn’t the term to use here, since that’s usually reserved for model-to-model training versus corpus-to-model training, or pretraining.

The shoe is ironically on the other foot. Anthropic has been sued multiple times for using improperly copyrighted books and other media to pretrain its own models. Now Anthropic is turning around and accusing Alibaba of using Anthropic’s outputs to post-train Alibaba’s own models. It’s a similar concept, nonetheless.

I do think that, in the broader scheme of things, the export-control regime and the regulatory regime that we were discussing earlier are, in some sense, protectionist and protective of the U.S. frontier models and their vendors, to prevent their insights from leaking out.

What’s been widely reported is that if you’re in China, in many cases, you have access through proxies in China or in friendly countries to Anthropic, OpenAI, and other strong Western models at 1/10 the cost—a 90% discount. You pay a lot less, and you get access to the models through these proxies.

The reason for that is that the proxies are reportedly gathering all of the reasoning traces. You agree to give up any notion of privacy in those reasoning traces. The proxies gather them, using you basically as a sock puppet, and now those proxies can, in principle, be used for distillation or other efforts.

I think it’s an interesting question that I’m certain is going to be heavily litigated: whether distillation constitutes espionage or not. It’s almost certainly a violation of the terms of service for Anthropic’s models, but whether it constitutes espionage, I think, will probably be a heavily litigated question.

It’s probably yet another reason that the world seems to be on a path toward a sort of second Cold War, where there’s a U.S. bloc and a Chinese bloc. It’s not just model access that doesn’t flow; it’s also reasoning traces that could be used to enhance capabilities that seemingly don’t want to flow.

You might be able to block distillation and still maintain the openness of API calls and the usefulness of these systems.

There are different types of distillation, from direct work on the latent to simply asking, “How do you answer this really hard question and make this really nice code base?” The latter is incredibly difficult to lock down.

I think what China’s been doing—the Chinese companies—is this: Why does Mercor have a billion-dollar revenue? Because they get all the experts in. And why do you need experts to answer questions? Because your data distribution is this good, and then the experts make it that good. But you could use Mythos or Claude instead to get that extra level up.

We’ve reached a plateau, and we’ve reached another level. If you look at why GLM 5.2 is better than GLM 5.1, and you look at what they’ve said—which I think is actually true—I also think that Alibaba probably did have these sock puppets, as Alex said, using the capabilities of Claude Code.

It does appear to be a recursive self-improvement loop, where you’ve got your initial distribution, and then you improve it because it’s gotten good enough and competent enough by being very thorough and looping back on the data improvement.

I think we’re actually at the point now where Chinese models will be better if they have either super-expert input or Mythos answering questions. At the same time, you could have this recursive loop where you don’t even need to have anyone’s data anymore.

Again, I think it’s very difficult for us to conceptualize, but this is that takeoff scenario where you can’t guarantee that if, today, China could never use any of the U.S. models again, the existing data set they have, the techniques they have are not good enough to keep up with frontier capability. That’s the difficulty here.

I think we’ve got to put a pin in this story. It’s going to come up again and again and again, but I couldn’t believe it. I mentioned earlier that my son just came back from China, and he’s got this fake Rolex. He said, “Check this out. I got this in Shenzhen, and it says Rolex on the front. It’s got all the patent numbers and the inscriptions inside the clasp. You literally can’t tell it’s not a real Rolex. It costs about $25.”

He bought it in the basement of the statehouse in Shenzhen. They’re literally selling illegal fake clones in the building where the governor of the province lives or works. I mean, what a blatant disregard for intellectual-property law.

But then you look across the whole Chinese economy, and its growth is predicated on copying ideas from elsewhere in the world, stealing them, and bringing them home. It’s in the DNA of the culture. Now we’re saying, “It’s shocking that they actually had 25,000 fake accounts looking at traces.”

Anyone who’s shocked by this is way out of touch with what China’s actually doing. I think this story will come up again because this will be the excuse that the U.S., Europe, and maybe South America use to try to crack down. This will be the trigger they use, because they need to suppress Chinese AI somehow. Otherwise, it’ll be out in the world and everyone will have it.

I was hosting a conversation on stage with Steve Jurvetson and Astro Teller in the early days of Singularity University, and we were talking about IP protection. The concept was that, in the end days of the singularity, IP will mean nothing, because if you’re dependent upon IP to protect yourself, you’re just out of luck. AI is going to reinvent the product much better than you ever did, iteratively and very much faster. It’s going to mean that you need to be constantly innovating, not trying to protect what you did years ago. We can see it happening right here, right now.

Alexandr Wang

Or IP means everything, and AI supercharges IP-litigation lawyers to do an amazing, superhuman job of protecting IP.

Yeah, I’m glad you said that, Alex, because I think that’s very likely in the midterm. I think robots are coming, space-based data centers are coming, but I think we’re in a hard takeoff of core AI right now.

The amount of intellectual property created in the next, say, 18 months will dwarf all of human history—by far. But it’s still all virtual breakthroughs. It’s software, video generation, solving all physics, solving all math. If you can’t protect that intellectual property, chaos is going to break out globally.

We actually have to figure this out. You can’t just allow countries to rampantly copy, especially given that privacy is so hard to contain and copying gets easier over time.

Alexandr Wang

Well, I think the point that Astro and Steve were making was that you’re not going to copy exactly. You’re just going to use what exists, reinvent it better, and create something that is uniquely an improvement on top of that.

Emad, where do you come out on this?

Yeah, I think it’s going to get increasingly difficult. You’ll say, “Do Teams, but make it not annoying.”

Make Zoom that doesn’t need to upgrade itself every 2 seconds. I think the creative capabilities of AI, and the copying capabilities of AI, are such that almost everything can be done in one shot within a few years.

You have this period where it had to get to competence. That first broke with Sonnet last year, and now, for almost all models, including open models, it’s here right now. Then this new loop means that, again, why do I need to copy when I can recreate but remix? That makes it very, very difficult.

In certain areas, like music, there are incredibly strong copyright protections, and that has a whole bunch of other implications. Software has very little. You’ll have this whole kind of gamut of things, but I just don’t think you can stop capability increases now with distillation or other lockdowns, or even IP lockdowns.

Alex

I’ll sound perhaps just a different note here, which is that I think there are striking parallels between the defense-versus-offense divide on software vulnerabilities and defense versus offense on IP litigation and IP protection.

One might say superficially, “Oh, yes, sure. IP is over because AI will, for any patent, be able to find a way to route around it.”

That’s the AI-will-overwhelm-via-offense argument. At the same time, AI can also strengthen defense. AI can draft better patent claims. AI can do a better job of, say, patent litigation than humans will at some point in the near-term future.

So it’s not 100% obvious to me that the—call it the Astro argument—that IP suddenly evaporates because intelligence becomes stronger is correct. I think there are a variety of reasons, as we’re seeing, quite frankly, with this export control regime, why superintelligence may want to protect itself via IP legal mechanisms. If that is the case, I would fully expect the defense side—or offense, depending on your perspective—to also be supercharged with superintelligence. It’s not 100% obvious to me that IP goes away. In fact, it might just be utterly supercharged in terms of its ability to protect.

Guess what, guys? We’re going to find out soon enough. All right, let’s turn our attention next to quantum computing. This week, another breaking story: President Trump signed a new executive order aimed at supercharging US quantum computing companies.

It’s a technology that could one day crack today’s encryption and turbocharge scientific discovery, specifically in drugs and biomaterial science. In a parallel move, the White House is moving to shield US quantum research from foreign espionage, reportedly directing intelligence agencies to guard it as we guard our nuclear secrets.

So, the US government thus far has committed $2 billion in venture investments under the May 2026 CHIPS and Science Act. Let’s take a look at who’s getting the money. IBM received the lion’s share, $1 billion of the program, to co-develop its Anderon quantum chip foundry in Albany, New York. D-Wave, Rigetti, and Inflection each secured $100 million, and PsiQuantum secured a $140 million equity stake.

Alex

Mildly excited, not very excited. On the one hand, I want to quip that the forthcoming US sovereign wealth fund will have amazing exposure to all of these quantum stocks. One of the executive orders—there were a couple of executive orders that dealt with it, but the more interesting one established, or rather required, the development of what’s called the Quantum Computer for Application Development and Discovery Science, QC-ADS, which is interesting.

It’s the 1st executive order I’ve read that mandates the establishment of a quantum computer for discovery science. On the face of it, great. We want to accelerate science radically. On the other hand, I do think this may be a case of begging the question somewhat. There is already a vibrant private sector of quantum applications—not just quantum computing, but also quantum networking, and most interesting to me, quantum sensing.

Establishing a national quantum computer effort for discovery science reminds me a little bit of the Genesis Mission. If you look beneath the covers at where the money’s coming from, it seems to be a repackaging of existing US government funding. So, that part I find less interesting.

What’s more interesting to me is the protection from foreign threats. Until the executive order, I, for example, was not aware of it. I’d be curious to hear: Were you aware that there’s a Quantum Information Science and Technology Counterintelligence Protection Team?

Nope.

Alex Karp

We have that. We have a quantum protection team in this country. I think that’s super interesting and, in some ways, evokes for me this idea that the US government was caught flat-footed by AI. The defense community in particular and the intelligence community were flat-footed. We saw, as we were discussing earlier, Mythos and now GPT-5.6 leapfrogging whatever internal capabilities, apparently, the NSA has when it comes to cyber.

In quantum, I think the thinking somewhere in the executive is not to be surprised a 2nd time and to actually get out ahead of any quantum capabilities that might be strategically disruptive. The problem, as I’ve mentioned on the pod in the past, is quantum for science acceleration just hasn’t worked that well. Quantum computing was supposed to give us protein folding. It turns out the protein-folding problem was solved by, arguably, AlphaFold 3, purely classically, without the use of any quantum computing.

There are a lot of folks, including Peter’s common friends, who were very aggressively marketing quantum computing specifically for solving all of these problems.

To be clear, the company I think you’re referring to, Sandbox AQ, is not using quantum computing. They’re using the equations of quantum physics on AI platforms. So, Emad, how do you come out on this? Are you excited? Is this a nothingburger?

Look, I think it’s potentially the next big wave, right? And unlike GPUs, where China’s catching up—Huawei’s about to release the Ascend 950s and others in terms of bulk but not edge—quantum computers are vastly more complicated to build, right? Even if the secrets kind of get out, China has a good thing in photonic quantum computers, but not the various types of quantum computers from Rigetti or D-Wave, with the quantum annealers, et cetera.

The most interesting thing for me is this: With super-Mythos-level models, we will be able to ask the quantum computers the right questions.

Uh-huh. Program them properly, yes.

Program them properly, which is actually quite difficult to do. Sandbox AQ, again, are doing a little bridge to that right now by having the equations. Although quantum equations and generative equations are very, very similar for a very interesting reason. But what type of quantum problem will require a quantum computer a day to figure out, or a year to figure out? Nothing.

Versus milliseconds, right?

So, what you’ve got is a convergence of asking better questions, with quantum supremacy and others coming. That meeting point means we might not need Dyson spheres. And that is actually something that—

Double down on that one for us, would you? Because that’s all we speak about at this party—[laughter]—Dyson spheres.

Basically, one of the things is: more energy, more compute, more intelligence, right? But quantum processes and questions can, again, be answered almost instantly, in microseconds. There, it’s not like test-time compute exists for quantum compute. But we’re very bad at asking the quantum computers the right questions, so they’re not at a sufficient scale.

If we have an increase in energy to ask really good questions, maybe there comes a time when we can answer all the questions that quantum computers need, and then meet in the middle through a mixture of GPUs and quantum computers to ask and answer almost any question. That breaks this energy ramp increase to solve the mysteries of the universe by coating the entire universe with compute.

Alex

So, I think I have to jump in on this one. I think there’s a latent assumption in this scenario that we see some sort of complexity-hierarchy collapse. Right now, one of the reasons why quantum computers arguably haven’t been that useful is because it’s actually really difficult for humans without superintelligence to identify algorithms that are both economically useful and also achieve some sort of quantum advantage.

We’ve found a number of quantum-advantage algorithms. They’re not that useful—at least not economically transformatively useful.

Good headlines.

Alex Karp

They make for wonderful headlines and amazing IPOs. To Emad’s point, and this is something that I’m bullish on, I do think there’s a pretty good chance that if there is, somewhere out there in math theory space, an AI-discoverable quantum-advantaged AI training or inference algorithm, AI will find it. That will suddenly pay back all the sins of having invested too much money in quantum infrastructure previously. It will atone for it.

But separately, on the point of the Dyson swarm, I would say the reason why I think right now we’re on a default trajectory of a Dyson swarm isn’t necessarily because everything is so efficient, but rather because we’re running out of headroom in Moore’s law. As we start to, on the one hand, lose room at the bottom, as Feynman would say, and on the other hand see skyrocketing demand for AI, and on the 3rd hand municipalities don’t want data centers, we push them to orbit.

Will quantum obviate the need for a Dyson swarm? I think they’re probably not mutually exclusive. I could imagine scenarios where we build a Dyson swarm of hybrid or pure-state quantum computers, and people don’t want quantum-computer data centers in their backyard, just like they don’t want classical computers.

Yeah, so—

Alex Karp

It’s a difficult—

Build the quantum computers in the permanently shadowed craters on the Moon.

Yeah. Look, I think the thing here, again, is the default path: energy—everything is converted to compute, right? But for frontier capability, and the type of capability we discussed earlier on this episode, that could disrupt nation-states, that could disrupt society. You might find, in the next couple of years, with the way these 2 curves are going, a discontinuity.

The United States is trying to ensure that it is on the frontier of that, with the most capable frontier models and with the right algorithms for the right quantum computers that are finally useful. If that happens and that meets, then it’s very difficult to fight against that. Because, again, you need to scale the test-time compute if you can crack that.

Alex

Yeah, I think the right analogy there is so—I agree. The right analogy is, again, something like we're in 1939 or 1940, and the goal isn't to infinitely—or rather, indefinitely—preserve an American advantage in these capabilities. It's just to slow the rest of the world, namely China, down enough that we can hit enough recursive self-improvement and dominate—for whatever definition you prefer of “dominate the future light cone”—with capabilities that really only need to be a few months, at most a year, ahead of the competition.

All right, Dave, do you want to add to this, or should I jump to the next story?

Well, I started this new company, Quantum.ai, for exactly this reason. So, I guess I have so many thoughts.

Okay, well, go for it then.

Well, let me just—I don't want to belabor it, but around the September time frame, around the Moonshots Summit, if you want Alex to be super excited on stage, let's talk about quantum photonics, not quantum computers. Also, quantum sensing is almost certainly going to work. You'll be able to store insane amounts of information in tiny, tiny spaces.

I've been working for 9 months now on this quantum AI and working on the algorithm side, but it's almost a certainty now that highly quantized neural nets can perform just as well as 32-bit floating-point neural nets, which opens the door to massive amounts of photonic computation efficiency. I would be shocked if, by the time we're talking to Elon next December, we're not talking about launching—

Or this December?

Coming December. Yeah. Instead of launching NVIDIA chips and the huge power they consume, get the Terafab started on photonic compute at about 1/100 the mass for the same amount of computation. And it could be even more than that; 1/100 is a conservative estimate.

So, the stepping stone to the discontinuity that Ammar was talking about is clearly photonic computing, not quantum computing the way it's currently defined. But it's still quite—it's quantum photonic. It's not quantum quantum.

Alex

Yeah. I think that's almost a certainty at this point: that will exist, that the current AI will discover the breakthroughs necessary, if there are any left, and then that will be deployed in what Elon is actually manufacturing within a year to 18 months.

I think Intel's doing some super interesting things there. But China, with their Jiuzhang series and others, is really focusing on that one area. And I think the US is completely underinvested in it. That particular area of core photonics and quantum photonics needs to be a much bigger focus and have much more investment, given what we've seen already.

For that reason, that should be the thing the White House just elevated. They're not aware of it yet, I don't think, but they will be. It'll be right up there on par, and it should actually be a bigger priority than the current quantum—the thing they just passed.

By the way, everybody—go ahead, Alex.

Alex

I will note that, again, in the news in the past 2 days, Elon and SpaceX just acquired, for several billion dollars, a photonic computing and communications company to merge in. So, I would not be at all surprised if photonics, which gives us, in principle, a 1,000× clock-rate speedup over these stupidly slow electrons, becomes a key part of the Starlink or Starmind plan.

By the way, I said this a couple of podcasts ago: now that Elon's got liquid stock with SpaceX going public, he's going to be on an acquisition rampage. Watch him buy companies left, right, and center.

I'm going to move us to our next conversation, which is a fascinating one for all entrepreneurs out there. It's about sleep. Here's the question we're proposing: hopefully all of you know that 8 hours of sleep is not just a good suggestion; it is evolutionarily what we're designed to need.

But there is a small percentage of humans on the planet called short sleepers who can get away with much less, as little as 4 hours of sleep. The number of humans who get away with 4 hours of sleep is 0.1% or less. There's a group who can get away with 6 hours of sleep. That's about 1% of the population. But the rest of us need 8 hours.

Here's the data, so that you're aware of it, because I used to say when I was in medical school, “I'll sleep when I'm dead.” The fact of the matter is, not sleeping will kill you. If you're getting 6 hours or less on a regular basis, you have a 48% increased chance of coronary heart disease, a 15% increased chance of a stroke, a 12% increase in all-cause mortality, a 5% increase in beta-amyloid—you know, this is what's going to give you neurodegenerative disease—a 17% higher risk of type 2 diabetes, and a 4× higher risk of catching a cold.

Sleeping is critically important. So, how do you become a short sleeper? If you can get away with 4 hours of sleep, here's the math: it's 28 hours of more work time or play time per week. That's 58 days in a year. You gain 2 extra months on your life. This is a way of getting longevity incrementally. Instead of adding time at the end of your life, you're getting 2 months extra per year.

I know about this because one of my portfolio companies, which is still under stealth, is the number-one player in this area. But it made news this week when Eli Lilly purchased, for $6.3 billion, a company called Synthesa Pharmaceuticals. Their drug targets a neuropeptide called orexin. In the brain, it's the brain's master on-off switch for wakefulness. They've been developing this for narcolepsy patients, but I think ultimately, guys, this is going to become a lifestyle drug. How many of you would take this if you could get it?

Alexandr Wang

If we could get it—or if we could get it legally.

Are you on it already, Alex? Is that what's going on? I think we need this to track the singularity, for sure. I would dedicate those 4 extra hours per day just to reading the feeds that are coming in. Alex, what are your thoughts on this one?

Alexandr Wang

I think this is potentially as transformative as the GLP-1 receptor agonists. This is Eli Lilly seemingly playing the same playbook over again. Just to refresh, the GLP-1 RAs were initially approved for the treatment of diabetes, taking advantage of this protein that was discovered, I think, in Gila monsters in Arizona.

Like Gila monsters?

Alexandr Wang

Reptiles in Arizona that could survive without eating for long periods of time and were somehow able to maintain sugar-insulin balances for long periods of time. It was discovered that, actually, if humans have an analog of that—it seems relatively well conserved—and if you can synthesize a receptor agonist, you can trigger common pathways and enable humans to balance their sugar levels over long periods of time.

But it turns out the pathway is so important that this is still something of a mystery scientifically right now on the GLP-1 RA side. You're able to have all of these amazing, potentially lifespan-increasing effects.

You're right. Gemini says it was in the Gila monster.

Alexandr Wang

The Gila monster. So, now, same playbook. And I should add, the GLP-1 RAs have taken Eli Lilly and made it a teracorn—is it a trillion-dollar company? The revenue that it's generating off of just this is comparable to, if not larger than, all the token revenue that OpenAI and Anthropic are generating on similar timescales.

So, if you're Eli Lilly, you have to be asking yourself: can you take this receptor-agonist playbook and apply it to other domains? It looks like Eli Lilly thinks the answer is yes, and they're going to apply it to neuroscience next, where the role of diabetes is going to be played by narcolepsy, the role of GLP-1s is going to be played by orexins, and the role of healthspan in inducing positive side effects across all of these other areas is going to potentially apply to the ability to sleep less without the nasty side effects that you were mentioning, Peter—to help people who suffer from daytime sleepiness due to Parkinson's and Alzheimer's and mood disorders, potentially helping people emerge from comas, because orexin is so essential to all of these neurological pathways.

Yes.

Yeah, that's why I'm so excited about my portfolio company. I saw the data; it's double the performance of this Intessa drug. If they're getting bought for $6.3 billion, that's hopefully going to be amazing. Dave—

Just think of one more point. Think of all the economic output that could be unleashed if everyone everywhere had 4 more hours of wakefulness.

Yeah, and that's my argument for adding 30 healthy years to a person's life as well. But imagine if you had it in your early 20s, when you're most productive, right? Crazy. Dave, what are your thoughts on this one?

Yeah, one of our partners, Mira Wilczek—actually, Alex, one of your many classmates—was a natural 4-hour-a-night sleeper. There's a bell curve there. It goes all the way down to 2 hours. Some people are 100% functional on 2 hours of sleep. It's very, very, very rare.

Most people are centered around 8, but some people need 14 hours of sleep. Can you imagine? How much would that suck? But Mira got so much done. It's like having a whole other life. I was so jealous of that. So, yeah, I'd take it in a heartbeat.

How much do you sleep?

I'm a 4-hour sleeper.

Are you?

Lucky.

I thought so.

Yeah, yeah, yeah. You get the time in. And now I'm very bullish on orexin. We had a long look at this with ASD and sleep disorders. Before, you could only get things like Lactobacillus, BioGaia, and others to affect it.

But you see a lot of knock-on impact from this when you upregulate oxytocin and orexin, including effects on inflammatory markers and things like ghrelin and leptin. And I think you could see, again, GLP-1 impacting all these other homeostatic stabilization mechanisms. I think this pathway will do the same, especially for inflammatory disorders. So, I think you sleep less, but you become less inflamed as a result of some of these treatments.

It's so exciting to be alive right now. This is my abundance story for everybody: This is the time to be alive, and just get excited about what's coming down the pike. All right, you guys ready for some AMA questions with the mates?

Yeah, let's hit it.

Fantastic. All right, let's jump in. Here we go. Emad, you're our super-mate guest for the day, so why don't you choose first?

Oh gosh, there seem to be a lot of UBI questions here. Can UBI be as simple as reaping dividends from US government-held golden shares? Do Not Freeze 79. The math is impossible.

Mm-hmm.

If you look at how big the AI companies would need to be to get a basic living level of UBI from dividends, assuming 5%, it's about $10 trillion, and you'd have to own half of them to get halfway there. They'd have to literally be the biggest companies in the world, and it would only be to the US, which is question 4. How can anyone outside the US survive if it's only distributed in the US?

I think we need to fundamentally look at how capital flows and monetary equivalents of UBI, where money is created, make a lot more sense, and we've seen some explorations of that. From there, we need to ask again: How do we value things when the AIs and robots basically dominate the world? I hope that we see a lot more research and trials on that.

Yeah, I did the math. If you imagine UBI of $3,000 a month for just US citizens—US residents alone—that's $12 trillion per year. The US budget is $7.4 trillion per year. So, there's a lot of capital to be made up there.

Alex Wissner-Gross

I'll take question number 2, which I think was aimed at me. Will companies migrate to friendlier jurisdictions like Argentina to escape US AI restrictions? Wouldn't this erode trust in any model that can be banned? And this is from Evraninank.

I don't think model companies are at all likely to migrate to Argentina. I think the frontier model companies are very likely to remain firmly entrenched here in the US, and you'll see Chinese companies firmly entrenched in China. Then there will be the rest of the world. I don't think Argentina is likely to end up hosting any frontier model companies.

Companies that I can conceive migrating to Argentina under its proposed new non-human AI corporation regime will be inference-time companies that are basically AI persons. A humanless AI company that's just operating its own business—I can totally see many millions, billions of those types of non-human AI companies migrating to Argentina.

That said, given the export controls we've just been talking about on this episode, I would imagine many of them will probably end up running on Chinese models—unless the US acts to restrict Argentina's ability to import Chinese models, which, by the way, could happen. We've seen, under a variety of non-AI technology regimes, executive action to restrict other countries that are neither in the Chinese bloc nor, strictly speaking, in the US bloc from importing Chinese technology or Chinese commodities.

I can imagine a scenario where the US acts to restrict Argentina's ability to import or operate cheap Chinese base models, which would obviously foul up that approach.

All righty, Dave.

Number 1 or 4?

Yeah.

I'll take 4. How can anyone outside the US survive UBI or UHI if it's only distributed in the US? And that comes from Ali Singh—or Ali Ali Singh? Who knows.

I love this question, and it's very timely. The World Cup is going on in the US right now, and there are people from all over the world crawling around Boston, LA, Santa Clara, and Mexico, too. By and large, the people in Boston are like, “Wow, America's awesome. This is great.” And the word across Europe—you know, you model and reinforce this—is like, “America sucks. America sucks.”

You get over here and you're like, “Well, this is just a cross-section of every type of person from all over the world who's immigrated to the United States, and it's where AI is happening.” California is just hopping, and it's fantastic.

The problem you run into is that the US government continually panders to the voter, and if you're not a voter, they just don't care about you. If you are a voter, then you're entitled to everything in the world. That has to change, and I think this is a moment in time where you saw at the G7 summit, when Donald Trump walked in, he said, “Okay, the boss has arrived.” That's not exactly the dynamic you want.

Still, this has got to happen now, because the wealth-concentration effect is so extreme. If these frontier models become the universal workforce that creates everything, it's all going to happen in just a couple of locations, basically in the US and China. We need to get this figured out during this administration, which really means in the next year or so.

All right. I'll take question number 1. It's from @jb_c1br. Why should people give up autonomy to accept UBI if unemployment is not going to be a problem?

JB, I think you've got the premise backwards. UBI isn't a trade for autonomy. It's a foundation for more of it. You can think about it this way: It's less of a welfare cage and more like the Alaska Permanent Fund, which I've spoken about before, where every citizen gets a dividend check and you can do with it what you want, right?

If AI is creating extraordinary abundance, distributing a share of that to the populace allows you to use it to uplevel your life, create your next company, or create meaning in your life, however you might want to do that. Even if mass unemployment doesn't materialize, UBI is a freedom dividend. That's the way I think about it.

Anyway, that's my answer for you, JB. All right, let's go on to the next set of questions. Dave, why don't you kick us off?

I'll take number 5, and then I'll throw it over to Alex. Why is the Moon better for data centers than orbit? Is it the gravity or the Earth-facing position? Definitely not the Earth-facing position.

This topic came up when we were talking about the Kessler effect, and we really did a good job of that in the last podcast. Low Earth orbit is great for data centers, but then you get outside 500 km altitude and you get the Kessler effect problem, with all kinds of debris lying around and destroying your data centers.

So, the best place is low Earth orbit, where there's a little bit of atmosphere that cleans the system naturally, but that's a relatively narrow band. The Moon is good for the same reason: It's somewhat protected, but there is no atmosphere on the Moon either. So, that's why I'll throw it to Alex. Alex, why do you think the Moon is better for data centers?

Alex Wissner-Gross

I would say they're complementary, but the Moon's cislunar environment certainly has a number of advantages that orbit does not. For example, if you're concerned about the security of your data centers, both the West and the East—the great powers—have demonstrated the ability to send robotic devices into orbit with grappling arms. China, especially in the past year, has very publicly demonstrated this ability to interfere with on-orbit devices.

Whereas, if you have a data center sitting on the lunar surface, you can defend it. You can be sure that there's no one, for example, coming in behind you, listening to the same beam that you're using to communicate with Earth. So, there are a few reasons from a security perspective. Also, the Moon has mass.

If we set up, as I think Elon and hopefully others are going to do, an industrial ecology that's non-terrestrial, you can start mining water and other elements. Water's not an element, obviously, but you can mine raw materials from the lunar surface to build more data centers. You can't build more data centers in orbit; you just don't have the feedstock to do that.

Alex Wissner-Gross

Yeah, I was just going to say, lunar regolith—for folks who don't know, that's what you call lunar soil—is silicon, oxygen, nickel, and iron. It's perfect for data centers.

Perfect for disassembling the moon, in other words.

Alex Wissner-Gross

Yeah, and there's also another version of that, which is highly likely, where you're using those elements and then using a mass driver to launch them into low Earth orbit. That's something that's definitely going to happen, because it's already on the planning radar.

And that's the ultimate high ground. One of the narratives for why we had an Apollo program at all was that it was a continuation of the Manhattan Project, with the moon as the ultimate high ground for launching weapons. It may be the case that, in addition to the moon being the ultimate high ground for launching weapons against Earth, it's the ultimate high ground for launching data centers to service Earth.

Alex Wissner-Gross

Kudos to one of my mentors, Gerard K. O'Neill, professor of physics at Princeton University, who ran the Space Studies Institute and wrote about this. He actually built some of the first mass drivers and laid out an entire architecture for mining the moon.

His vision wasn't data centers. It was actually building and launching solar power satellites to Earth orbit to provide solar energy on rectennas on the ground. We've changed that a little bit, because we're going to use the energy in space for creating intelligence, but he laid this out in the '80s.

He was an amazingly brilliant individual who left us too early. Gerard K. O'Neill—look him up.

Alex, choose your next one. Let's leave number 7 for Emad. Yeah, go on.

Alex Wissner-Gross

I think I have to choose number 6. What physical and behavioral forms will humans and other species take in the age of the singularity? This is from John Allen 383.

I'll construe the question this way: I would argue that we are in the age of the singularity, so we already know the answer. We look like ourselves. I'll construe the question instead as, what physical and behavioral forms could humans take after the age of the singularity?

I think I've argued in the past that there will be many new forms of person and personhood. I take a lot of heat, especially in the YouTube comments, for agitating for some form of AI personhood. That's a legal form, not a physical or behavioral form.

To the extent that the question is asking what posthumans will look like, if you will, I think we'll see uploaded humans that have relatively de minimis physical form, but are just a collection of bits or maybe qubits running on AI infrastructure. I think we'll see borganisms—collective human intelligences—whether it's via Neuralink or some other format.

Alex

I tend to think, as we discover new physics and new applied physics, the substrate for the compute that humans right now augment ourselves with, and that soon many of us, I think, will be running on, is also, in some sense—to paraphrase Bucky Fuller—going to ephemeralize. At some point in the distant future—how distant, I'm not sure yet—we might even see something that Arthur C. Clarke wrote about many times: maybe at some point humans will just be able to run on the gravitational field, or operate in a state approximating pure energy, with fewer biological meat bodies.

At the same time, before everyone attacks me in the comments, I will point out that I think this will be both purely optional, and I think the future is going to look much more heterogeneous, not homogeneous. You're going to see humans who look substantially the same as they do right now 100 years in the future. At the same time, 100 years from now, you'll see some posthumans who are uploads in the cloud, functioning in a quantum computer. Those very different forms, I think, can and will coexist next to each other.

I'm going to add to what you just said, Alex. You're going pretty far out, so let me go in the interim. We're going to start to edit ourselves, right? We just talked about the idea of an orexin-like molecule allowing us to shift to 4 hours of sleep.

We're going to see gene edits, injection of Klotho to increase our IQ, or other gene edits to increase our muscular ability. We're going to start to see BCI—people walking around who've got a connection of the neocortex to the cloud. So I think those capabilities are coming in the next few years to a decade.

Alex

Can I ask you a question about that, Peter, just narrowly? We saw some people not take too kindly to the Enhanced Games. We see some municipalities not take too kindly to building data centers in their own backyard.

Do you think there's a high likelihood of a future where all of these biological enhancements are either tightly regulated or shunned, such that, as with pushing the data centers into space rather than building them on land, the transformative physical and behavioral changes to humans are basically all pushed into the posthuman realm by humans who regulate the more obvious short-term biological enhancements out of existence?

I had that conversation with one of my boys, and I said, “Listen, morals and ethics change over time.” I remember when the first in vitro fertilization efforts were taking place. It was thought, “Oh my God, this is awful. You shouldn't allow this. This is immoral. This is not what God desired.”

Of course, now IVF is considered normal, allowed, and beneficial, allowing couples around the world to have children later in their lives. So I think at first it's going to be shunned, maybe by more religious elements, but I think ultimately it's going to become accepted.

Shout-out to Ramez Naam, who wrote an incredible book called Nexus, which talked about all of these genetic edits to create BCI capability—basically, a neural lace in the brain—but also the edits. A lot of it will go underground and then eventually will become accepted.

Of course, I've been enhanced. Why would I not want to? We all want the best for ourselves and our kids: the best education, the best food, the best tech, the best whatever. Why wouldn't you start with the best genetics? I know this is a sensitive subject, but I think it's going to become part of life going forward. What do you think about that?

It's going to be fascinating, isn't it? I actually spoke at the Oxford Union a couple of weeks ago in a debate about whether AI can be a net positive, and obviously I was the one against it. That's coming out soon, so you can hear the argument against this. These are big questions that we're going to have to ask.

Very different.

Alex Salkever

I think ultimately we're all made of star stuff, but some of us will be more star stuff than others. That's the way I like to think about this.

All right. Would you take number 7 for us?

Yeah. For EU entrepreneurs lagging the US and China, is the highest-leverage move vertical AI startups or agentic workflows into legacy industries? That's from Steve Crab.

I think this is fascinating because you can't compete on frontier models for various reasons, but also the nature of regulation in the EU means that there's massive potential for the latter: agentic workflows into legacy industries. In the US and Silicon Valley, the initial diffusion of innovation is happening really rapidly because companies are open to it. The vast majority of the US hasn't adopted it yet.

In Europe, it's even further behind, but it's inevitable because you get the competitive pressures. Being able to go in and work with those companies is a massive transformative opportunity because it's inevitable. Whereas building vertical AI startups in this competitive environment—regulation varies from industry to industry—is just not as easy as in the US.

I would say it's more about transforming companies in a way that's comfortable to them and charging big markups than necessarily building the verticalized AI startups, where really the US is a much better place to go to compete globally.

Emad, I don't have to ask you the question: Is Europe waking up?

Yeah, I think it is, slowly. It's just that there's so much institutional inertia here. There was something in the EU as a warning call. It's quite a nice kind of future thing, like AI 2027, you know? But you see, we have an extra 4 years because we're a bit slower here.

Where, instead of humans being disempowered by AI, Europe as a whole is disempowered by AI.

Oh, our regulation will be the great filter. It won't stop anything from coming out.

Ouch.

So it's slowly waking up. That's got a lot of traction in the upper circles because this Fable thing really was a massive shot across the bows of the decision-makers.

Good.

So I think there's a good chance the EU AI Act gets repealed. People are trying to figure out new mechanisms, and suddenly the fire has come from nothing. Maybe because it's so blooming hot here as well.

I hope that Argentina's Milei thing also goes really, really well and creates a role model, and then some European countries say it's—you know, like Ireland, to me, is a natural—

Estonia needs to pick it up.

Yeah.

Alex Salkever

Yeah, we have actually introduced this new thing, EU Inc., whereby you can set up a company in under a day. That's a big deal because, in Germany, for example, it takes up to 6 months.

All of Europe?

Alex Salkever

Across all of Europe, yeah, because in Germany it can take up to 6 months to get your tax status and other things like that.

So, little bit by little bit. Wow. All right, I'll take number 8. This is from Mark Simonian, 210. Wouldn't it be easier to put data centers in oceans and the Great Lakes than flying them 93,000 miles up? Maybe that's a transcription error, but Mark, we're not talking about 93,000 miles; we're talking about 500 miles up into low Earth orbit, or, on the moon, it's 240,000 miles up.

So, to answer that, Microsoft already proved the ocean concept. I looked it up here: Project Natick ran subsea for 2 years and lowered failure rates compared to land-based servers, with excellent cooling. Of course, the advantage of going into the ocean is you don't have to deal with the launch costs, there's plenty of cooling, and it's relatively nearby for repairs. You can either have it within your jurisdiction or offshore sufficiently so it's outside the jurisdiction.

But the reality is, as launch costs are dropping—and we heard about this in the last pod with Will Marshall—apparently, Dave, you and I were talking to Elon, saying he flipped the bit. He did it—what was it now?—9 months ago, when he started talking about how no one was talking about orbital data centers. But apparently, Will Marshall, Eric Schmidt, and folks at Google were talking about this as much as a decade ago.

Ultimately, it's going to be both. We're going to have data centers on land, in the oceans, and in space. There is as much demand as we can supply, so I don't think there's a limitation there right now. Alex, any more thoughts on ocean-based data centers?

Alex

Yeah, I'm a big fan. Peter Thiel funded—I think it's Pelagius—that's focused on this. I think, look, one of my operational definitions for the singularity is every sci-fi trope happening everywhere all at once, and I think data centers on the ocean are the prototype for seasteads and ocean colonization.

So, yes, I think we get our data centers on the oceans. I don't think it's as scalable as orbital compute, but then again, if you want to live on an ocean colony—which is, I think, something that we're going to get over the next 10 years—then you're certainly going to want your own local compute. I could totally imagine, if you've seen the Pelagius video demos, that ends up becoming a nucleus for a next generation of ocean colonies.

Hey, one thing to add to that: if the photonics take off that we were talking about earlier is on a 1- to 2-year timeline, which I think it likely will be, people will want to put the NVIDIA data centers in the ocean and not launch them because they're very heavy—a couple of tons for an NVIDIA NV72 with power.

That's cheaper to deploy in the ocean while you're building the photonic thing, which is about a 10th, a 100th, a 1,000th the mass per compute. And so, that's a real possibility that we have both, but in that order.

Emad, before we go to our outro song, give us an update. How's Intelligent Internet doing? What are you up to these days? Give your fans a little bit of an Emad preview.

Yeah, it's going well. Like I said, we're releasing the harness that will allow anyone to uplift the models above Mythos on Monday.

Where do they go for that? Because by the time we release this, it'll be up.

It'll be at probably my Twitter, @EmadMostaque. I'll retweet it. Then we've been working on Sovereign AI, and I think we have a mechanism by which everyone can own the AI and robots. We'll be announcing that soon, along with the new book, After the Last Economy. So, we're going on.

Amazing. And I know the secret stuff you have going on, which is incredible. Appreciate having you in the universe, and Alex and Dave, love you guys. A quick shout-out to Saleem, who's now on stage in Spain, or at least landing in Spain. Come back, Saleem, we miss you. All right, this is an outro music by @rohitheinventor called Moonshot Masters. Let's take a listen.

From the shores of Ireland across the sea, we're watching dreams become reality. Moonshots lighting up the sky, showing us how far humanity can fly. Every story, every breakthrough, every spark brings a little more light into the dark. From AI to health, from space to the stars, you're helping us see who we really are. The future STARTS TONIGHT WITH BOLD IDEAS taking flight. Moonshots showing what's possible, making the impossible look logical.

Hand in hand, we'll build what's right for every child, every life. Together, we'll create a brighter way.

Amazing. Love that. So, everybody, please submit your music, your videos to us. We love them. And remember, this is the most extraordinary time ever to be alive. Our mission here is to give you an optimistic view of the future. I hope you understand what's going on day on day. I hope you enjoyed this extra emergency podcast. We stuck it on a Saturday morning. Emad, Dave, Alex, love you guys. Be well.

You, too. Have a great weekend.

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 MetaTrends that are impacting your family, your company, your industry, your nation. I put this into a 2-minute read every week. If you'd like to get access to the MetaTrends newsletter every week, go to diamandis.com/metatrends. That's diamandis.com/metatrends. Thank you again for joining us today. It's a blast for us to put this together every week.