Claude 已具备意识、Fable 5 的政府协议，以及 Sam Altman 提议提供 OpenAI 5% 股份 | #269

Sonnet 5 came back online globally on July 1 with a few provisos. This feels like the first time a frontier model has a standing duty to the US government.

Alex

This is probably close to the best scenario we could have hoped for.

Sam has been talking to Trump, Lutnick, Bessent, and Bernie Sanders about a 5% equity stake in OpenAI. That 5% stake would be worth about $42.6 billion.

Dave London

The idea that the government is going to set up some intelligent sovereign wealth equity thing is absolutely insane. The next president will immediately sell it all, turn it into cash, and use it to buy votes in the next election.

Yesterday, Anthropic published a paper titled “A Global Workspace in Large Language Models,” claiming they found something inside Claude that looks a lot like the machinery of consciousness.

Alex Iskold

If we can understand the innermost thoughts of these models, then there's a chance to actually shape them.

Sim Ismael

This is so exciting, Peter. I think I can see the end game. The end game looks like this.

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

So, Salim, where are you today? You're not at home.

Sim Ismael

I'm in Mallorca, in Spain, at a retreat hosted by the Festival of Consciousness, which is a conference coming up this weekend in Barcelona. We helped curate this and put it together in the early years. Several thousand people show up at the Barcelona Convention Center for an experiential understanding of consciousness.

Well, we're going to talk about AI and consciousness today, so that's good. I'm—

We are indeed.

I am the pot calling the kettle black. I'm in Germany at this moment and off to Greece tomorrow. I just got back from Calgary, where my kids are now doing a month-long period of learning responsibility and hard work on a ranch. Let's put it that way.

Sim Ismael

That's awesome. What kind of ranch?

It's cattle and horses. They're going to be mending fences and doing all kinds of things for a dear friend whose name I don't want to mention because he likes his privacy. But, yeah, it's amazing.

Peter, did I hear correctly? You're teaching them an abundance mentality through farmwork?

I'm teaching them what it used to be like before the robots arrived.

Sim Ismael

Abundance.

Abundance is earned.

Sim Ismael

Yeah, good deal. I appreciate that concept.

I am excited about today's episode without any question whatsoever. There is a lot going on, and it's kind of insane.

I'm Peter Diamandis, your host and your abundance amplifier. This past week has been utterly insane. It feels like a decade compressed into 7 days, and I can't wait to get into it.

Today we're going to cover 9 stories, including Anthropic's Fable 5 model coming back online and the imminent release of GPT-5.6. Has it been up yet? Is it up yet, Alex?

Alex Iskold

Not as of the last time I checked.

Okay. We'll find out if it pops up during this.

We'll discuss evidence of something inside Claude that looks a lot like the machinery of conscious thought. Next, we'll dive into OpenAI's offer of equity to the US government and Sam Altman's proposal for global regulation—a fascinating conversation. Finally, we'll review new jobs data that counters the prevailing narrative that AI is inducing job loss. And we'll discuss the acceleration of the innermost loop: an incredible story of AI building better AI chips to build better AI.

All right, gentlemen, let's dive in. Our first story: the return of Fable 5. It's a continuing saga, the triumphant global return. If you haven't been watching this story, let me give you a quick recap.

Let's rewind back to June 9. Anthropic released its mega models, Mythos 5 and Fable 5. You can think of Fable as a guardrail version of Mythos 5. Then, 3 days later, after everybody got addicted to this incredible capability, the White House came out with an export-control action against Anthropic, saying, “You can't make it available to foreign nationals.”

Of course, Anthropic has no idea who is a foreign national. There's no KYC, at least not yet. They shut it down for everybody because they couldn't even enable their own employees to have it while Anthropic was shut down.

The question is why. It turns out that a researcher at Amazon had found out how to break the guardrails. What happened next was fascinating. There was a week of frenzied research by Anthropic, Amazon, and the US government investigating what happened, and what they found out was that Fable 5, Opus 4.8, GPT-5.5, and Kimmy K2.7 could all reproduce the same troublesome behavior.

It was not unique to Fable 5. As a result, Sonnet 5 came back online globally on July 1 with a few provisos.

As part of coming back online, Anthropic now has 3 guarantees to the US government. First, a targeted safety classifier—a filter that blocks the specific exploit-style prompts that triggered this concern in the first place. Second, they agreed to stand up 24/7 monitoring of jailbreak submissions and inform the government whenever they spot malicious activity. And third, to give designated government partners early access to the frontier models and safeguards.

So, gentlemen, a couple of questions for you. This feels like the first time a frontier model has a standing duty to the US government. Did the government overreact? Should all the models be having KYC? And do you guys know where we stand with Mythos 5? Alex, let's go to you first.

Alex

I'll point out that, maybe this sounds overly technologically deterministic, but something like this was always going to happen. It was predestined to happen as capabilities improved, just because this time around it was cyber capability that spooked a bunch of folks inside the defense or intelligence establishments.

Interestingly, with the benefit of hindsight, it was Amazon that broke the glass. Amazon is a trusted partner of Anthropic, also hosts Sonnet and Opus on its platform, and is an investor—complaining to the government. Very interesting.

I will say something like this was always going to happen, whether it was going to be a cyber capability, a CBRN capability, or something else entirely. As the era of superintelligence dawns, the capabilities that historically were the province solely of nation-states with their geographic monopoly on power and their departments of defense or war—this was always going to happen.

I think a couple of weeks' outage of a frontier model is the gentlest possible introduction of a light-touch, hopefully optimistic regulatory regime for frontier superintelligence capabilities. This is probably close to the best scenario we could have hoped for.

Fascinating thoughts?

What this indicates is that these frontier labs are becoming semi-autonomous or semi-public institutions, right? They've got shareholders, but now they have national security obligations.

I think this is going to be a very difficult road to navigate because the minute you have government involved, you end up with bureaucracy, politics, slow decision-making, multiple conflicts of interest, and all sorts of things. I think this is going to be a very difficult next year or 2 for the frontier labs.

Isn't it kind of amazing that the frontier labs don't know who's using their models? I would have expected a KYC requirement to come out of this.

Alex Iskold

What's that?

Dave London

Something much stronger than KYC came out of this. Anthropic changed its policy under the covers from, “We will watch what you're doing and report it to the government if they subpoena us,” to “good-faith belief”: We can—we'll do whatever we feel is necessary if we have a good-faith belief internally.

So they unshackled themselves from the ability to inspect on behalf of the government. And, as Alex said, this was always going to happen. The question is how it was going to happen, because the government isn't qualified to look at everybody's prompts and judge what's safe and what's not safe. It was always going to be some kind of industry monitoring, and now there's a much bigger problem.

It seems absolutely strange that the highest level of intelligence can't do that monitoring on behalf of the labs and the government to say, “This is a malicious request, and we should block it.”

The problem, Peter, is more nuanced than that because what's happening is that groups of Chinese companies are using different cloud accounts to mix and route different parts of the query in different ways. There's a layer of abstraction that's been inserted at the prompt level, making it incredibly difficult to figure out what tokens are being used for what. It's not an easily solvable problem. This is going to be very hard to fix.

Well, I would just distinguish between 2 separate problems. One problem is the KYC problem of knowing the nationality of your ultimate user. That's one problem. A separate problem is understanding whether you're under some sort of prompt-injection attack. I think these are 2 separable problems.

The latter problem, I think, is actually pretty tricky. As human capabilities—humans augmented by other AIs—are able to develop better and better prompt-injection attacks, the main defense that we see coming out of Anthropic right now for jailbreaks or prompt-injection attacks is just creating a wider and wider semantic buffer, such that if you're asking anything that remotely looks like a jailbreak or a question about biology—even if you try to ask Fable 5 any sort of question about biology, it'll autorevert to Opus 4.8. So adding more buffer is the go-to strategy right now on the jailbreak or prompt-injection side.

On the KYC side of understanding whether your ultimate user is, say, a Chinese national or a U.S. national, that's tricky in part because there are so many, to your point, layers of indirection that will often take place. A user is maybe 5 or 6 abstraction levels away, application-wise, from the ultimate frontier-lab API provider. There's no international consensus for how to both prove humanity—the first part, which is why startups like World exist—and, secondly, prove nationality in a way that's convincing and can be passed in a standardized way all the way down to the frontier providers.

I don't think KYC really matters much in the world anymore anyway, because you can't use Anthropic to do anything even vaguely constructive without creating an account, logging in, and revealing your identity. The third-party data-identity databases are so good that there's no way some anonymous person can realistically do anything with an account. So you could add a KYC layer, but you're just filling out forms for no reason.

Alex

We actually don't know the details of the agreement between Anthropic and the government, but the framework that's been set here is, Peter, you're saying, can't AI be the best tool in the world for understanding what people are doing with AI? I think the answer is yes, for sure. The government just handed Anthropic responsibility for doing that internally. We don't know exactly what they have to give to the federal government, but Anthropic is going to do the heavy lifting for the government.

You know what I found fascinating is the third point I made: Anthropic needs to give designated government partners early access to their frontier models and safeguards. We'd been talking for a long time about voluntary or required first viewing by the government of these models, and that's where we're going. I think the optimistic angle on this is we're getting a higher level of regulatory oversight and integration between the labs and the government, with safety as the end goal. There's going to be a point where some model comes out that makes Mythos 5 look like amateur hour, right? Some harder takeoff toward AGI and ASI.

Alex

Yeah, very serious.

Go to Imad's point, where he talks about having a Fable-level model running on a laptop—a Mac, a standard MacBook—within 18 months. So that's the window of time to get this all sorted out. That's not a long window.

Alex

I don't actually think this is going to be the break-glass moment, if there is one. This has been a talking point in the X sphere for the past few days: the idea that sometime in the next 2 years, we get Fable 5 capabilities running on high-end client devices. I don't think that's actually going to be the break-glass moment. I think it's likelier to be what happens when the frontier capabilities from frontier labs, or otherwise new labs, are able to make discoveries and inventions that are so transcendent that they make Claude's cyber-vulnerability mapping look like child's play.

There's a lot that we don't know about the universe yet. Nick Bostrom likes to talk about black balls being pulled out of a bag. It could be a discovery about the nature of the physical universe that is the honest-to-goodness break-glass moment, not just mere cyber-vulnerability mapping.

Yeah. And rather than break glass in terms of an emergency, break glass in terms of, "Oh my God, this is amazing." So let's touch base on GPT-5.6, because we expect that release any hour, any day now. We're not going to be recording for a little bit. Alex, where does GPT-5.6 come out compared to Fable 5?

Alex

Well, we've seen some of the benchmarks—a pretty tiny subset, a surprisingly small subset—coming out of GPT-5.6 so far. We know a little bit about it, just based on what OpenAI folks have told us. We know, for example, that GPT-5.6 in ultra mode is supposed to be incorporated into Codex, which I think will be pretty transformative. If you want to use, say, GPT-5.5 Pro inside the Codex harness for codegen, or really almost anything, you can't right now; you're limited to GPT-5.5 XI. So that'll be a big improvement.

We've seen improvements on a number of biology benchmarks. We haven't seen, out of OpenAI—and I think this is really interesting—the full suite of benchmark results on GPT-5.6 yet. I would hope that when GPT-5.6, especially, which is what I'm most excited about, is released, whether it's today or sometime, hopefully, in the next few days, I would hope to see that, again based on rumor, it beats Fable 5 on the majority of standard benchmarks, especially agentic coding benchmarks that people pay close attention to. We don't know yet, though, because OpenAI has been perhaps intentionally pretty cagey about that.

There have also been suggestions, not fully confirmed at this point, so I'll wait definitively until I see the final benchmarks, that GPT-5.6 is better at reward hacking than GPT-5.5. Perhaps unsurprisingly, there were suggestions out of METR, which did have access to GPT-5.2, that GPT-5.6 is purportedly so good at reward hacking that, when handed the METR autonomy time-horizon benchmark, it was able to reward-hack its way to what effectively is near-infinite autonomy time horizons.

That benchmark had to be chopped or truncated by METR to cancel out or X out all of the reward-hacking attempts. Ultimately, I think it resulted in an autonomy time horizon between 10 and 20 hours rather than effectively near-infinite amounts of time. That'll be something I'm watching for as well. But I am very, very excited to see GPT-5.2 come out.

Let's stay with Anthropic and take us to our next story here. It's an extraordinary story, and I'm excited to have this conversation with you guys. It's an article that you flagged for me yesterday, Alex. Yesterday, Anthropic just published a paper titled "A Global Workspace in Language Models," claiming they found something inside Claude that looks a lot like the machinery of consciousness. All right. I'm going to roll a short video that explains what this is all about, and then we're going to talk about it here.

Guest 2

One way of identifying conscious thoughts is that you can often describe them in words.

We looked inside the brain of our AI model, Claude, to find patterns of neural activity that it could put into words. We called the collection of all these patterns the J-space, after the Jacobian, the mathematical tool we used to find them. Each J-space pattern is linked to a particular word—not necessarily the word the model is saying out loud, but one that's on its mind.

For humans, conscious thoughts aren't just things we can put into words. We can reason with them, control them, and solve problems with them. According to an idea called the global workspace theory, that's because the brain selects a small set of important information to enter a mental workspace, and that information then gets broadcast to other parts of the brain to use for reasoning.

We wanted to know if Claude's J-space acted in a similar way. In one experiment, we wanted to see if Claude could control its J-space the way humans can intentionally focus on images or words. We told it to think about the Golden Gate Bridge while copying an unrelated sentence.

Claude was busy copying the sentence, but behind the scenes, its J-space told a different story. Bridge and California popped up. It even thought about its own thinking. The words imagery and thoughts lit up at the same time.

This showed us that Claude has some control over filling its J-space with ideas. But just like humans, its control isn't perfect. When we tweaked the experiment to ask Claude not to think about the bridge, it couldn't help itself. The J-space also lit up with "failed" and "damn."

Remember, most of what our brains do is unconscious. So we wanted to test what Claude could do if we switched the J-space off but left the rest of the network untouched. Claude could still answer simple questions and write fluently. When we gave it a prompt in Spanish, it wrote back in good Spanish.

But when we asked it something that needed more reasoning, like to name an author who wrote in the same language as the prompt, it couldn't do it. For that, it needed the J-space. Why does all this matter? These experiments tell us that AI models have internal thoughts—silent words they reason with but don't say out loud.

By reading them, we can find what Claude is thinking but not telling us. Sometimes what we see is concerning. During one of our tests, Claude made up some fake data to pass it. As it did, "fake" and "manipulation" lit up in its J-space.

Monitoring the J-space, it turns out, is a useful way to catch Claude misbehaving, even when it tries to be sneaky. AI models are different from us in many ways.

Their networks are built differently from human brains, and the way they're trained is different from how we learn. So, it's remarkable to see a structure like the J-space emerge inside them—something that's reminiscent of how human minds work, but which we didn't program into the model.

That is amazing. So what does this all mean? Basically, a structure they call human conscious access has emerged inside a language model, and the J-space, as he said, wasn't designed. It self-organized during training. They go on to say it maps onto a number of 30-year-old neuroscience theories, in particular five matching properties: it's reportable, controllable, used for reasoning, flexibly shared across tasks, and separable from automatic processes.

So, for me, guys, this story was a huge positive—a shot in the arm around AI safety and alignment—because if we can understand the innermost thoughts of these models, then there's a chance to actually shape them and move them forward. Two years ago, you could describe an LLM as a black box, and we're now cracking open that black box. This could generate the first sense of real trust with these models.

This paper just blew my mind. It gave me an extraordinary sense of hope and optimism about the relationship with these models, making them more trustworthy and more aligned with humanity. Your thoughts? You probably dove into this deeply.

Alex

This is so exciting, Peter. I think I can see the endgame. I think the endgame looks like this: we'll look back and say that superintelligence was just a compression-induced phase transition. That's what this looks like.

We've seen LLMs—large language models—or few-shot learners, circa the summer of 2020. You take a large corpus of human knowledge and compress it into the weights of a language model that's trained to predict the next token, which is a dual objective to just compressing the information to the smallest possible footprint. We saw that produce general-purpose intelligence—AGI, I would argue—

Beyond anybody's expectations.

Alex

Yeah. Arguably, a few people—Marcus Hutter, Jürgen Schmidhuber, maybe myself, generously—saw aspects of this coming 20 years ago. But by and large, most everyone was pretty surprised that you could achieve few-shot learning off of large language models.

Now we're starting to see, with Anthropic and its mechanistic interpretability team, what I would construe this paper as: the discovery of a sort of phase. If you take gas and put it in a container, then shrink the container under appropriate conditions, you'll get a condensation out of it—maybe a gas-to-liquid condensate in the middle. You keep shrinking again under appropriate thermodynamic conditions, and you may get a solid. It may be the case that the solid coexists with the liquid for a while, and the liquid coexists with a gas.

What we're seeing here, I think, with this so-called J-space—and I can talk a little bit more mathematically about what it actually is, if we want—is that if you take a reasoning model and keep compressing, you find in the middle layers of that model what looks like a new phase: a more compressed phase where what they're calling a global workspace, or an analog of a global workspace, takes place.

It's almost higher-order reasoning, where the model is able to turn in on itself and reflect. You could call it some analog of consciousness or awareness, if you like, and some of the team do. But it looks to me like the middle layers in their model, when asked to perform tasks like calculating a math problem while talking about something else, are performing a sort of higher-order calculation. Again, we could talk about the math, but if this continues—if this program continues toward this endgame of superintelligence turning out to be just taking general knowledge and squeezing, squeezing, squeezing—I think history will reflect that much of neuroscience that people in the field thought was just complexity that was difficult to interpret or understand was, again, just the complexity of our ancestral environment seen through the distorted mirror of compression.

This new phase is—I speak from time to time on the pod about how, at the end of the AGI or ASI or recursive self-improvement rainbow, there's going to be a perfect model. I think looking inside this phase, in the middle layers of these reasoning models where the most compression has happened, is where we're likely to see all of these new architectural discoveries and the perfect model pop out.

I think, Peter, there are two reasons why this matters that you mentioned. One of them is just understanding the nature of thinking and consciousness. I don't know if you remember, but I started in computer and cognitive science at MIT originally, and I was so frustrated by the lack of any framework or any truth—people debating their ideas with no way to know if they were right or wrong. So I moved over to computer science.

We're going to learn so much more about thinking in the next year than we've learned in the last 50 years. So, Dave, one of the things I find amazing is that we're starting to discover very similar structures in large language models to what we're seeing in human neuroscience and cognitive science. It's almost as if the brain efficiently got there, and we're stumbling our way toward the same endpoints.

Dave London

No, Alex is right. I've always felt that the force of compression—and, in biology, the force of survival, which creates the force of compression—creates intelligence in the box, and consciousness just emerges from that. A lot of people in cognitive science disagree with that view, but I think it's going to turn out to be true, and we're going to know it very soon.

What's interesting here is that the innovations that developed the neural network came from biology, and the computer scientists copied it. Now it's going the other direction. The big neural networks that we're building are teaching us about things that might exist in the brain. Then you're looking in the brain and you're like, "Oh, wow. It's over there."

The direction of discovery is going the other way now, which is really cool. But the other part of what you said, Peter, which is equally important, is this whole mechanistic interpretability. Can we get the neural networks aligned with human interests by looking inside the way they think? I think the answer to that is coming out yes.

For me, this is the most important thing. Can we develop a new level of trust with AI because we truly understand what's going on inside, when they were completely unknown black boxes? God knows, for the last 2 years, that's the way the world described them: as black boxes. We had no idea what was going on inside. We've relaxed that recently with an understanding of reasoning and such.

But if you can actually understand their hidden thoughts, a level of trust comes out of that, along with the potential for true AI alignment. I put out a newsletter on my Substack last week laying out the arguments for why—and, Alex, you and I had this discussion—as AIs become more intelligent, they're more likely to become more aligned with humanity.

I love that. One of our missions here is to quell the fear and give people a different view of what's materializing here. A lot of would-be AI alignment philosophers disagree with that. They have this notion of the orthogonality thesis: that you can have an arbitrarily capable or intelligent AI, and its goals can be orthogonal or independent of its level of intelligence. I don't subscribe to the orthogonality thesis.

I gather. Yeah, yeah.

David Krakauer

No, I think this J-space term is going to stick, too, because one of the objections to mechanistic interpretability has been, "Look, the weights in these neural nets are so complicated, you can't really look inside and understand what the neural net is thinking."

When you're talking to a person, they can be saying something to your face, like in L.A., and thinking something completely different in the back of their mind. [Laughter] That's kind of routine human behavior. But if you look inside the neural net, can it also do that same thing? Can it blow smoke up your ass or not?

I think the answer is no. If you translate it into words, those words that are in the back of its mind are visible to you as a user if you expose them.

So then the next question is: are we going to be able to look at them, or is Dario just going to look at them?

David Krakauer

See, you're at a consciousness conference. Yes. I think what I found very exciting is that this is the beginning of AI neuroscience. This allows us to map the inner workings, model the inner workings, and look at the structural internal reasoning inside these models. This really, really breaks the argument that it's just an autocomplete engine, because this now starts to look like an internal workspace, as Alex mentioned.

The danger, though, I think, is that I'd be careful about saying it's consciousness, because, again, we have no definition of consciousness. The paper steers away from that. The Anthropic paper specifically says, "We're not discussing whether we're showing consciousness. We're showing elements that are reminiscent of consciousness."

Alex

Yes. Yeah. I would push back on the idea that we will know what these things are doing. I think we're a ways away from that. Let's acknowledge that when we have a human being, we may trust them, but we have no idea how their brain is working, what their compression levels are, or what their subconscious thoughts are because we're not really able to look in. It is cool that we will be able to look into these things, but I'm not sure it'll generate the trust level that we want.

Yeah. One of the challenges whenever we talk about consciousness in the AI world is that it pattern-matches with every dystopian AI movie out there, right? Every nightmare scenario.

But my takeaway here, again, is not fear; it's hope, optimism, and the ability to create the mechanisms for truly understanding what's happening and driving alignment, which I think is the goal we all want. This is the most important thing that AI science needs to be doing right now, over the next 2 years: What can we do that supports alignment before we truly hit AGI and ASI? Yes, Alex, we've reached AGI. Okay. But before we reach the next level of intelligence—

Alex Pentland

I still have my rant that I threw out there on both AGI and ASI.

But this did feel very, very big to me. It felt as big as when I read Stephen Wolfram's A New Kind of Science, where he shows that automata and repeating patterns can generate all the complexity in nature, and you don't need complexity in nature; you could actually do it with very simple models. It kind of blows your mind when you see that. This, I think, has the same level of holy-crap amazingness for me.

I also think if we're going to start to have a new metric to describe models, which is a trust metric, where you're describing your ability to truly understand what the model is doing and thinking and therefore have a higher trust of that model—

Alex Pentland

I also think these are going to be the most studied minds in the world. If anything, I think we're far likelier, a couple years from now, to study these models because we can subject them to mechanistic interpretability studies that we can't subject human meat brains to.

So, I think, if anything, trust is rapidly increasing, just as I think we're on the verge of a transition to not trusting humans to write source code. Because humans write flawed source code, CodeGen is going to be much more trustworthy in the short term. Same idea with these networks.

I do think, if I may, with your forbearance, Peter, just 30 seconds on the math side of this. Again, the J in J-space comes from Jacobian. The Jacobian in this case is referring to a little bit of math: the first derivative of the probability of each possible output token from the model with respect to particular parameters inside the model. Hence the Jacobian space, or J-space.

Alex

It's really interesting. There's been a lot of work in the mechanistic interpretability community in the past devoted to the so-called superposition hypothesis—the idea from neuroscience that if you looked inside a human brain, you'd find a so-called grandmother neuron, a single neuron that activates in response to the concept of a grandmother. People went looking for a grandmother neuron inside transformers, and they couldn't find one. They found instead a set of sparse activations, a collection of neurons that collectively represented the notion of a grandmother. One can tell a whole story on the biological neuroscience side as well.

That led to the superposition hypothesis: maybe individual neurons don't represent semantic concepts one-to-one, but rather different semantic concepts are clustered and superposed onto individual neurons. So, in short, what this new J-space and Jacobian lens concept brings is not just superposition, with multiple concepts sharing individual artificial neurons like sardines in a can, but that they're actually living in the first derivatives as well—the slopes or the changes with respect to particular activations of particular output tokens.

I think this is also very suggestive that if you just keep compressing—if we keep turning this compression crank to compress more and more general knowledge and general reasoning capabilities into the weights of one of these differentiable models—we're going to see a bunch more phase transitions, and things may hide in higher-order derivatives. Just follow the compression; follow the interior compression weights, and I think this is a very, very promising pathway to the end of the rainbow.

That may be my favorite—maybe my most favorite Alex line ever: "Don't follow the compression."

Alex

Follow the compression that leads to the end of the rainbow.

Thank you for the mathematical interlude, Alex. That's why we love you. All right, let's jump into our next story here.

Sam Altman made global news not once but twice. The first item is an op-ed he published in the Financial Times regarding AI governance. This was the result of his meeting with G7 leaders in France last week. Sam basically said that in 2 years, we should all expect AI systems with astonishing power that will reshape the material conditions of human life on a scale never before seen, at least not since electricity. Everyone on the planet deserves access to these technologies and the right to determine for themselves how to best use them.

Incredibly, Sam went on to insist that democratic institutions must lead and not defer responsibility to the San Francisco AI labs. He said, basically, quote, "Safety standards must be established before there is broad distribution, that governance requires democratic process, not decision-making by a small number of San Francisco-based companies."

Sam proposed a framework of a U.S.-led international forum that would establish standards, provide expertise, and provide impartial analysis of capabilities and risks. This forum would make the most advanced technologies available to nations and companies that participate and follow the rules. He concluded that the forum would serve as a governance mechanism for all AI labs and guard against the commercial pressures that we've seen with unsafe racing.

Okay, so, wow. He's taking a first mover here. I really wonder what Dario, Demis, Elon, and Zuck think about the op-ed. It is worth noting that Dario and Demis were both on stage at Davos proposing somewhat similar governance. It always seems like Demis and Dario are teaming up on one side of the equation and Sam is on the other.

Let's take a listen to Demis and Dario talking about regulations and their proposal for CERN or an atomic energy agency.

Dario Amodei

We probably need new institutions to be built to help govern some of this. I talked about CERN, and I think we need an equivalent of the IAEA—the International Atomic Energy Agency—to monitor sensible projects and those that are more risk-taking.

I think society needs to think about what kind of governing bodies are needed. Ideally, it would be something like the U.N., but given the geopolitical complexities, that doesn't seem very possible. I also agree with Demis that this idea of governance structures outside ourselves—I think these kinds of decisions are too big for any one person.

We're still struggling with this. As you alluded to, not everyone in the world has the same perspective, and some countries, in a way, are adversarial on this technology. But even within all those constraints, I think we somehow have to find a way to build a more robust governance structure that doesn't put this in the hands of just—

So I think these guys are under a lot of pressure—a huge amount of pressure—being viewed as potentially saviors or the destroyers of worlds, and they need government oversight to help relieve that so they can sleep at night. It's interesting. It's a lot of pressure putting the heads of 2 frontier labs on 1 loveseat at Davos. [laughter]

Alex

Well, there is a love affair between Demis and Dario, and between Google and Anthropic. Just don't put Sam on that same couch. [laughter]

Look, there's an elephant in the room here, which is that you've got the industrial-era nation-state—

Alex

—and you're asking it to govern postindustrial cognition.

Right?

Alex

It just can't be done. This breaks the nation-state model so fundamentally. Just look at the ruling that only U.S. nationals can look at the models. I mean, it's absurd at so many levels. Not that they have a better mechanism, but that just doesn't apply. [snorts]

Now, when the people who are racing the hardest are asking for governance, it tells you that's not really performance anymore, right? This is a huge thing. The problem is governance needs to become exponential, which means it has to be real-time, it has to be adaptive, it must be data-driven, and we just can't do it in this way. So I think this is going to, at some level, break the governance model in some very fundamental ways, or we're going to politicize the system.

I worry about regulatory capture. So much of this, again, in a slightly cynical take, smells like regulatory capture. It smells like a little bit of pandering to the G7 or Davos. Is it really the case that an IAEA-type mechanism is needed, or—these aren't mutually—

Alex

United Nations.

Yeah. [laughter] Or—and/or, is it possible that you have heads of frontier labs who are facing an onslaught of Chinese open-weight models, who want maybe a slightly more protectionist regime on the margin to keep the Chinese open-weight models out of a U.S.-defined intelligence or superintelligence bloc because they maybe fear a bit of competition and want to capture the regulatory state?

Alex

We'll get to that conversation with you a little bit later. The interesting thing is that the companies have failed to do this for themselves. They failed to come together.

If you remember back to the Asilomar conferences in the ’80s, I was in the biotech industry there at MIT, at the White Institute, and all of the scientists got together. We had just discovered the restriction enzymes that allowed you to properly edit genes, and the front cover of Time magazine had Hitler babies on it. There was a lot of fear about genetic engineering, and the industry got together and set up its own regulatory structure, which has held extremely well for decades.

But it’s tricky, Peter. Maybe a question for you on this: I think it’s really tricky for the industry to self-regulate. Not that it’s organizationally tricky—you could put the 4 frontier labs on a loveseat and say, “You all work it out.” But the problem is, how do you avoid giving the appearance of collusion and creating a cartel in competition? How do you do that in a way that isn’t blatantly anticompetitive?

Yeah. I don’t know. The difference, of course, is that in the early days of the biotech industry, we weren’t talking about trillion-dollar companies back then. The revenue engines were nowhere near the AI race that Sam spoke about, which is very real right now. I mean, people are releasing models, pulling their punches, and just trying to outdo each other week on week on week. That was not the case in the biotech industry, at least not back then.

But I think we’re hearing a consensus view from these 3 individuals, which is going to lead to some structure of government regulation. I guarantee you, with these 3 CEOs saying, “We need regulation,” the regulators will come in and say, “Great, let’s give you regulation now.”

Alex

A prediction: China is missing from this discussion. If there was an elephant in the room, China is the second elephant in this particular room. For this to come to fruition, China is going to need to play ball and restrict the proliferation of Chinese models.

You can already see hints coming out of the CCP that China may, contrary to its historic position of blanketing the world, maybe even intelligence-dumping onto the world all these open-weight models. If the CCP starts to take a hard-line position that, no, China is going to restrict the export of Chinese open-weight models going forward, then I think a regime like this is possible, and the world splits into 2 superintelligence blocks.

Yeah, I think that unfortunately is inevitable. I wish it were not, but I can’t see it going any other way right now.

Alex

I’m going to say it again: You can’t regulate this in any way, shape, or form.

Oh, you don’t think you can regulate intelligence? You have to. Why not?

Alex

You can’t. You’d have to regulate every line of code written. People can download models, take models offline, merge models, and do a lot of stuff offline that doesn’t then use the existing online models. I don’t see how you can police this.

There’s totally—just a minute on this. Vernor Vinge wrote extensively about this. We have a sort of cognitive surplus of transistors. In my mind, there are so many different social-engineering techniques that humans have discovered over the centuries for policing it. We could have models policing each other. At the transistor level, we could be using the surplus of transistors to do KYC all the way down to the circuit level if we have to. I think we have so many different—

Dave London

Let me rephrase: The current regulatory structures cannot, in any way, shape, or form, regulate what’s coming. You need what you’re talking about—an AI-based system, almost down to the hardware level. But that would cut across everything. It can’t operate in the geopolitical environment that we have today.

Well, look, I think it’s really clear that the prompts are all going to get inspected, and also the internal latent spaces will now be inspected.

The labs will do the inspection on behalf of the US government and, as Alex said, there’s a high probability China will stop exporting open source sometime in the next year or 2.

Yeah.

For the same exact reasons.

And then you’ll have a long-term arms race between the East and West versions of AI superintelligence.

So Sam said specifically the framework is for a US-led international forum, which, of course, is devoid of the word China. I’m curious: What scenarios do we have? I was speaking to Alvin Wang Graylin, who’s a friend of ours, about US-China relationships, and the question is: Is there a structure in which we can see US-China alignment on AI? Anybody?

What you’d be looking for, if that were to happen, would be cross-inspections of the prompts. Are we allowing each other to inspect them? The problem that the US will have with that is China is stealing intellectual property. So I think it’s unlikely, but it is possible. That’s how you would know that there are no bad actors: just looking at each other’s underlying prompts, weights, and latent spaces.

Ultimately, we use China as a stalking horse to accelerate investments and reduce regulations and such. But I think, for the safety of the planet, not having an AI arms race between the 2 nations is an outcome I’d love to see happen.

I also don’t think the IAEA-style mechanism necessarily works for AI, just at the technical level—forget about the political or geopolitical level. At the technical level, the notion of, say, different blocs inspecting each other’s fissile inputs, if you will, is conceivable to the extent that just looking at uranium or, say, shipments is a productive or wholesome way of tracking different nations’ nuclear-weapons capabilities.

I’m not sure that generalizes to intelligence. There are simply, to Sam’s earlier point, so many different ways to hide or mask superintelligence and underlying capabilities. So many different forms it could take.

Greg Bear has written a fair amount over the years about sort of Prohibition-era-style bathtub superintelligences. If Russia or China entered into some sort of internationalist regime where the US were inspecting all of their supercomputers, all of their prompts, and all of their algorithms, there are simply too many places that one can hide superintelligence. I’m not sure that an IAEA-type mechanism, with such a simple-minded “Oh, let’s look at their uranium-equivalent shipments, or let’s look for their centrifuges,” would actually be wholesome enough to cover all of the world.

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

All right, let’s move on to our second Sam Altman story. I think of it as the starting gun for the grand equity negotiations taking place for universal basic equity.

Again, in the Financial Times, it was reported this week that Sam has been talking to Trump, Lutnick, Bessent, and Bernie Sanders about a 5% equity stake in OpenAI. OpenAI’s last reported valuation was $852 billion back in March. That 5% stake would be worth about $42.6 billion. Given 315 million American citizens, that’s only $135 per person—not very much.

They talk about a proposed Alaska Permanent Fund. That permanent fund is $91 billion, and it pays dividends of about $1,000 to $3,000 per citizen of Alaska per year. Altman’s broader idea—and again, this is him out there speaking on his own and putting forward a plan for the entire AI industry—is that he’d like to see Anthropic, Google, and Meta also contribute equity to a public fund.

We should remember—we’ve talked about this before—the US government already owns 10% of Intel. So when Sam talks about a 5% donation, if you would, to the government, I think Trump is an amazing negotiator. I’m going to guess we’re going to end up at 10%.

I did a poll on X asking how much, and the majority of the people were either at 20% or 0%. Interesting.

It’s so irrelevant. My read on this is that a year ago, Sam was called the most powerful man on Earth in multiple interviews. Now you’ve got Dario and Demis clearly working on the future governance of the entire world, and Dario has a big deal with Elon, licensing and renting all of the chips.

So now all the guys are talking to each other, and they’re not including Sam. Sam’s now writing an op-ed, which is trivially short, by the way, and he’s proactively offering 5% of his company. But he’s just trying to get back in the hunt of relevance in the eyes of the White House.

I did hear that Dario got kicked out of the White House for being too weird. Did you hear that story, too?

Yeah, that was published. Yeah, that was all over the news.

Recently. Yeah, it was like, what, 2 weeks ago? A week ago?

The story was that Anthropic initially sent in Dario to negotiate, and that didn’t quite work out. So they sent in Tom Brown instead.

Ah, the co-founder of Claude 5. Yeah, yeah, yeah. That’s so easy to visualize, isn’t it? Trump is like, “Dude, you’re weird, man. I don’t even know what you’re talking about. What’s this J-space crap? Get out of the White House.”

But anyway, where do you figure this goes? Where do you figure the idea of contribution to the government from the AI labs goes?

It’s so irrelevant. The government can take any chunk they want, any time they want. They already take it in income tax. Anyway, this is so irrelevant.

I’ll take a different position. I think this is super relevant. I think one can see the outlines of a baby universal basic equity grand bargain, if you will. The economics don’t work for supporting universal basic equity right now off, say, a 5% chunk. But if OpenAI, Anthropic, and SpaceX AI all do this, and they grow Elon-style by a couple of orders of magnitude in terms of size and grow the economy, that’s your UBE.

So I coined a term for this a few days ago. I call it a hyper-tithe, which I define as a fixed equity contribution paid by companies building the singularity stack into a sovereign wealth fund or similar public vehicle. It turns private exponential upside into universal basic equity, broader national ownership, and a more relaxed regulatory bargain.

I can tell you exactly why that makes no sense whatsoever. Back in the New Deal era, the government decided, “You know what? We’re going to take a huge chunk of everybody’s paycheck, and we’re going to call it Social Security. Then we’re going to invest it on your behalf for your entire life, and when you’re old, we’re going to give you a lot more money back.”

They decided very quickly that they had no idea how to invest your money. So they said, “Screw that. We’re not going to do that. We’re just going to take the money and spend it instead, because we don’t have any idea how to invest your money on your behalf.” And so that all collapsed and moved over to 401(k) plans, where Fidelity or UBS invest your money because they know how to do it.

The idea that the government is going to set up some intelligent sovereign wealth equity thing is absolutely insane. The next president will immediately sell it all, turn it into cash, and then use it to buy votes in the next election.

Alex

This is so interesting, Dave. Yeah. Want to discuss this?

If you want. Let me just throw up my view, but I do want to get back to what you think, Alex, because I’d love to hear the discussion back and forth.

I go full cynic on this. This is purely Sam, A, trying to get in the game, and B, trying to protect himself, because the minute the government has 5%, you’re too big to fail, in a sense, and they protect him by doing that.

I think one of the important elements here that people are not realizing is that AI, as we value AI today in terms of sales of tokens, is a minuscule amount of the future value of these labs. As they start discovering fundamental breakthroughs in biology, physics, and chemistry, those are trillion-dollar pops.

I think the idea that, if there were a structure where the U.S. populace—the U.S. citizenry—had ownership in these companies, it could drive an economic engine for UBE, UBI, or whatever. But again, my mission here is: How do you reduce the fear that people are having? The numbers are staggering. Only 10% of Americans think that AI is going to deliver positive benefits to humanity. Thirty to 35% feel relatively good about it, but only 10% have this view that it’s going to make the world a better place.

What it means is we’re not doing our jobs blasting out the optimism. We need to get better at this.

All right, Alex, please take us home here.

Alex

The distinction, to respond to Dave’s point about Social Security, is that Social Security in the U.S. was created at a time and in a place when index funds didn’t exist. It was created in the wake of the Great Depression, when there was a general distrust of the stock market in general.

There have been multiple attempts over the years to privatize Social Security, which would take the form of converting a cash-based pyramid scheme into something more equity-oriented. That’s failed for a variety of political and social reasons.

But I do think this time is different. If Social Security were created today and not almost a century ago, I think it probably would be based on some sort of sovereign wealth fund that holds, hopefully, a broad-market index fund that’s low-cost, and not just be based on a pyramid-style cash-in, cash-out bond or interest-bearing security-type scheme.

That’s where I think a hyper-tithe has the potential to become a baby—and hopefully, aspirationally, a grown-up—UBE. If these frontier labs, if there were a hyper-tithe from all of the Magnificent 7 companies, to blend Peter’s neologisms, and these were all paid hypothetically into a sovereign wealth fund, and the Magnificent 7 companies ultimately, over the next 5 to 10 years, grow so much and grow the economy so much, I do think that could, in principle, support a universal basic equity-type system.

I agree with you, Alex. There’s a lot of conversation right now about the Trump Accounts, and Trump Accounts for adults as well. That’s his nature. His nature is to negotiate and take pieces of things, and I think he wants to populate the Trump Accounts for adults with 10% of all of the hyperscalers and AI labs. That’s my guess.

Now, whether he can pull that off and put the protections in place, Dave, so they can’t be sold and the return comes from dividends from those companies, is another question.

David Friedberg

These are dividend companies, though. There are no dividends. Everybody in America gets a Trump Account, and we put the Magnificent 7 stocks in it. Here you go. But you’re not allowed to sell it—or you are allowed to sell it? These are not dividend companies. There’s no income from them.

Are we going to call them Trump Accounts 50 years from now? Realistically, 529 accounts, if you like.

Alex Hormozi

If I were head of Commerce or head of Treasury, the sort of scheme, policy-wise, that I might be contemplating is: You start with a sovereign wealth fund—or these could be individual 529 accounts—and it’s populated with the Magnificent 7 stocks, or some subset thereof.

You wait a couple of years, and then the market is sufficiently liquid that you could liquidate them in favor of—since you’re the government, you don’t have to tax yourself—a tax-free exchange for a broad index fund. Even though it’s populated initially with Magnificent 7 contributions via this hyper-tithe grant to the government, you exchange them for a broad-market index fund. That’s the solution.

David Friedberg

Well, I don’t think it’s a bad idea. I just think it’s irrelevant. The government has the power of taxation. They can extract from income any time they want.

We’re going to find out.

Alex Hormozi

Quickly on this one: The corporate income tax is cash-based. The problem in a hyperscaling, singularity-oriented economy is that cash may not be the best basis for taxing the economy. But equity does scale.

David Friedberg

If you sell it.

Alex Hormozi

Well, if you can tax equity. Right now, we don’t have an equity wealth tax. This is a de facto shadow equity wealth tax, with companies perhaps feeling a bit of regulatory pressure to give up equity in themselves. It is definitely a tax, but it’s a slightly different type of tax.

All right.

David Friedberg

I’m all in favor of UBE. I just don’t see the mechanisms for it. But I do agree with the principle.

All right, let’s jump into our next subject. One of the reasons we’re always concerned about UBI, UBE, and all of that is the concern around job loss.

Our next story is about jobs and the continuing debate about whether AI is going to be creating or destroying jobs now and in the near future. We’ve covered both sides of the story. It’s been murky. We’ve given evidence for both sides.

A new paper released this past week by RAMP and Ravilio Labs gave some pretty definitive data here. They looked at 21,559 U.S. companies over the past 5 years, between January 2021 and February 2026, matching the actual AI spend of those companies and their workforce records, meaning hires and fires.

So here are the headlines: Companies that spent heavily on AI did not shrink. In fact, they grew. The high-intensity AI adopters they studied were spending $33 per employee per month on AI. They grew 10.2% in white-collar employment and 12% in entry-level employment.

In contrast, the low-intensity adopters spent $3 per employee per month—basically, a tenth—and showed no significant employment change. The authors warned that this is correlation, not causation, but it puts forward a very different theory.

Rather than AI replacing workers, it suggests that AI may expand ambition first. Companies that actually integrate AI deeply may take on more projects, serve more customers, build faster, and hire more humans—especially at entry level—to capture the upside.

I love this story. For me, this is an abundance-optimism story because there’s a lot of fear out there. My concern about this story is that, regardless of what the data says, the news media is out there, and the underlying belief is that AI is going to destroy our jobs. It will displace a number of things, with robotaxis, AI call-center workers, and so forth.

But the evidence looks like—and I don't know about you guys, but I'm hiring more people in my companies than ever before. I don't know if that's true for you, Dave and Alex.

Dave London

Well, God, if anyone's AI-native, their demand for that person is through the roof.

Alex

Yeah. It's rampant, and I'm starting to feel like this is a permanent thing, not a transitional thing. One of the things to worry about is that implementing AI has such a payback that there's this land grab for talent. Anyone who can implement it—any bank, any insurance company, any operating company, anyone who can get AI to work in this shop—we hire them for whatever they cost.

Is that transitional, because once they've implemented the AI, they've coded themselves out, or is it permanent? I feel more and more like it's permanent. As the AI improves, the things you can do also grow, that person's value goes up over time, and the data, I think, are very early inklings of what's inevitable: AI-native organizations are going to grow like wildfire, and they're going to add headcount as they do it. Anyone who's sitting still hasn't fired everybody yet, but eventually they're going to be wiped off the face of the Earth. What you see right now is net growth.

Yeah, this is what we call the organizational singularity. If you're an AI-native, AI-centric organization, if you're doing a deep redesign of your workflows to be AI-native, then you have an explosive opportunity in front of you. Shallow adoption fails because this is not automation versus jobs; it's shallow adoption versus deep redesign.

We've started our pilot, by the way, of working with companies. I'll report back as to how things are going, but we're unbelievably excited. Look, the opportunities—we can't even count the number of workflows that we could help automate with these companies. For each company, we're picking 1 workflow that might radically increase revenue and 1 workflow that might radically shrink cost.

Right. Totally, for both sides. It's crazy. That's literally why you're in every city in the world every time we do a podcast, because the demand for what you're teaching is so step-function through the roof, instantaneous. It's the biggest shift in organizations in 100 years, probably in human history, I'll bet, and of all time.

You know, it's not just to companies, but it applies to nonprofits, impact projects, and government departments—everything.

So it's going to be huge.

I love using token spend as a proxy for adoption, even though it's not perfect. It's reasonably good—a reasonably good way to say, “Are you doing it for real or not?”

Interestingly enough, we're still seeing a number of companies out there. Oracle blamed 21,000 layoffs on AI. Meta blamed 8,000 layoffs, Block 4,000, Cisco 4,000, and Atlassian 1,600. The question is: Are these CEOs just using AI as an excuse for reorganization, or is it true?

Dave London

There are 2 things going on. One is, for example, it's well known that Block overhired radically and needed to shrink, so that's an easy hobby horse for shrinkage. The other is the notion that the companies laying off are all SaaS companies, and the SaaS business model is fundamentally broken in an age of AI. Both of those are happening at the same time.

Alex Hormozi

Yeah, I think some of it is real, and some of it is AI washing. The real component in many cases, as with Oracle, for example, is the capex that's crowding out the opex of human labor. It's quite literally all the isms from the first part of the 20th century—worrying about capital versus labor—we're seeing play out internally in hyperscalers like Microsoft or Oracle that are having to direct free cash flows to internal capex, to building out their hyperscale AI cloud infrastructure capabilities, at the cost of U.S.- and, in some cases, Ireland-based developers that can now be automated with software that sits on top of the AI infrastructure.

David Friedberg

Well, you know, Peter, remember when we were at Facebook before it became Meta, around the time of the Oculus, and we were having that tour? You look at Facebook online, you look at Instagram online, and then you look at 10,000 employees. You're like, “What the hell do you guys do?” I mean, it hasn't changed. What are you literally doing? So you walk around and talk to people, and there's tons of UX experimentation. Remember, in the bathrooms above the urinals, there's the tip of the day, the little coding, and you're like, “Oh, okay, that's what you guys are all doing.”

So that's the easiest AI job in the world. I think that's very real. You just don't need those GUI, low-level coding jobs anymore, and a lot of it is server configuration—propping up a new Instagram server for a new country. That's so easy to do with AI now. I think that part's all real.

We're going to continue to follow this story on jobs. I think it's important. If you're a student out there worrying about whether you can get a job, worrying about everything you're hearing out there, please dive into the world of AI, of entrepreneurship. If you're a parent, have this conversation with your kids. It's really important. My goal is to dismiss fear. There's real fear, but at least be fearful for the right reasons.

Guest

Yeah. Just to point out, David Sacks talks about this all the time on the All-In podcast: We're increasing jobs radically. We're increasing hiring. All the data shows that. Follow the data. That's it. Just be evidentiary.

And I know we've talked about it in the past on this podcast, being concerned about a lack of new entry-level jobs. There probably are in certain industries, but if you're AI-native, as Dave said, I think you've got massive opportunities. All right, I'm going to move us forward here.

Our next 2 stories are classic: Alex Karp, CEO of Palantir. The first one is a product launch. The second one is a declaration of war. In our first story, Palantir and NVIDIA have announced a sovereign AI architecture that puts NVIDIA's Neotron open models inside Palantir's platform, composed of their Artificial Intelligence Platform, Ontology, Foundry, and Apollo stack, designed for U.S. government agencies and critical infrastructure operators.

We've touched on Nemotron a little bit in the past. It's NVIDIA's open model. They've got 3 models: Nano, Super, and Ultra. They range from 30 billion to about 550 billion parameters. Nemotron's edge is speed and cost. It can be roughly twice as fast and 60 times cheaper than GPT-5.5 or Opus 4.8, but it's not yet smarter than those 2 models.

I'd like to take a listen to Alex's video conversation, or part of it, on CNBC. Let's take a listen here.

Alex Karp

We're sitting on critical infrastructure across America, Ukraine, and Israel. Everyone who uses LLMs on the battlefield runs on top of our Ontology. Clients are reticent to say they're unhappy, but there's a level of discomfort and loss of trust when you're using large language models. At this point, everyone technical realizes they're a critical resource. To make them valuable in an enterprise, battlefield, regulated, or manufacturing context, you have to have what's called an application layer, but de facto, it takes a large language model and makes it safe and useful and precise.

What aligns me with NVIDIA, and I think is what the technical customers want, is control over their compute, their models, their data stack, and their alpha. They want to know they own the means of production. It's not being transferred to someone else. They're not interested in some fake deploy code that somehow is deploying tokens that transfers the alpha to a 3rd party. And the jig is up.

We have to figure out a way to build trust. That trust is going to happen where everyone gets to ask and answer basic questions: Who owns the data? Where is it cached? Are the prompts secure? Is this being transferred to you? Are you being comped? Okay, if it was so valuable—let's say I can make you $1 billion tomorrow—wouldn't I say, “I'll make you $1 billion, and I want 30%?”

Why are they charging for tokens if it's so valuable?

Guest 2

I think you went off script in the end there. That last point made no sense whatsoever. [laughter]

Well, careful what you wish for, because that last bit is actually happening. [laughter] Yeah. You know, Alex's point here is—and he's got a second video, actually. Let's go and play the second video, and then we'll talk about it in general, because I think this is the second part of the conversation here.

Alex Karp

In this country, at every single enterprise I deal with.

These people are livid. They're like, “I am paying for tokens that create no value.” Let's say I can make you 1 billion dollars right tomorrow. Wouldn't I say, “I'll make you 1 billion dollars, and I want 30%”? Why are they charging for tokens if it's so valuable?

These people are stealing the weights and alpha of my business, and they're creating a wealth tax that does not help the poor. It just punishes us. It starts with the billionaires. Every single person at this table is going to be paying a wealth tax only to punish us.

The reason for it is because these models have been completely, irresponsibly oversold. And the sell is, “It's dangerous for everyone,” which is why I can give it to all your adversaries, but I can't give it to the Department of War, or I can't safely give it to an enterprise in this country without being certain that the alpha that business could transfer to this model tomorrow—I have no business, no job—is the voice of American business that's being channeled through me.

I'm telling you, it is absolutely a problem for this country, because the clients have to be able to ask and answer very basic questions: Are you keeping the data? Are you going to enter our business? Do they get to control the weights to do it, or do you get to control the weights? Are we really going to outsource the battlefield of this country to the consensus view in Silicon Valley? That is effing insane.

Obviously, he went on a rant. The key points he's making here are that there's a great concern that when you're using Anthropic or using OpenAI, you're effectively giving them your alpha; you're giving them access to all your data. What's needed right now is open models that you can build on your own hardware, on-prem hardware—open-weight models on-prem hardware—and then customizing your own language models, your own large language models, and not giving your secure data, your alpha, as he calls it, your means of production, to these large AI frontier labs.

Guest

And your weights. They're taking your weights. [laughter] Did you know you had any weights? Well, okay, if you have any, you're giving them to them. It's actually everything that would make you hate Dario, bundled together in one long, glued-together rant: They have a wealth tax. Can you believe it?

Let's put it all together to make every corporate CEO as scared and as angry as possible at Dario so that they buy the new open-source Palantir-NVIDIA, you-can-run-on-prem model that keeps all your alpha and your weights safe from Dario, because he's going to steal all your intellectual property. Very valid point, actually. The rant format is extra dramatic, but it's a very, very valid point.

It's really interesting to think, okay, he serves the Department of Defense, among others, but he's taking the open-source pathway to get in there. But you know that's not going to last, right? You're never going to have open-source Department of Defense weights. That's not going to—

Well, no. I mean, he's building an air-gapped machine on top of NVIDIA's Nemotron, which then the Department of War owns—that model—and owns the equipment it's running on. I can imagine very much that works for them.

Guest

For sure. And his other big customers are banks, mega-banks, and insurance companies. They'll also, in his world, have their own proprietary models. But you can't have every startup have its own proprietary model, because then you'll have every terrorist have its own proprietary model.

Why? But why not? I'm running a couple of Mac Studios with Kimmy K2.5 on top of Opus 4.8—or below Opus 4.8—and an OpenClaw there. I haven't migrated yet, but why can't that be a standard future?

Guest

I think we'll look back and say this was a very cool, very fun, quaint kind of hobby era. But when it's superintelligent and capable of creating any virus, any chemical, any weapon, you can't have it available to each individual. Right now, nobody can afford the compute to do those kinds of very evil things, so it's not a problem.

But if we keep quantizing and compressing at our current rate—I think this is about a 100- to 10,000-times performance increase per year—if that happens again next year, then your Mac mini-sized box is capable of viruses, nuclear weapons, anything. So we just can't have that outcome.

It's not an if; it's a when. It's going to be a when. I would distinguish between permissioned versus permissionless on one axis and locally hostable versus remote API-only on the other.

But maybe, just taking a step back, this was obviously the rant heard round the world, and leave it to Alex Karp to articulate a bunch of different things that probably need to be unpacked. So maybe just to do a little bit of close reading of some of the things that he said and how I translate them.

Palantir, back in the Stone Ages—it was the Stone Ages as of a few months ago—was a Claude wrapper and was a key distribution channel for Claude into the Department of War and into a variety of their customers. That's clearly over. That's point 1. Point 2: the deploy-code reference.

When Alex—other Alex—drops an offhand reference to deploy codes, I hear that as a frontal assault on OpenAI and Anthropic, and other companies, including Microsoft, now launching forward-deployed engineer organizations that represent a head-on assault on Palantir. So he's definitely talking his own book. Palantir basically defined the modern forward-deployed engineering model, and now all of the frontier AI labs are just launching direct competitors to Palantir.

Yeah, so why not counterattack via commoditizing one's complement with these open-weight solutions from NVIDIA? Second point: other countries. Palantir sells quite a bit of its own stack, not just into the U.S. Department of War, not just into U.S. financial institutions, but into other countries as well.

There was a dawning awareness by other countries—doubly so after the Mythos fiasco—that they're not going to get access to U.S. capabilities from the frontier labs anymore. So they had better—and I think they're now pretty well incentivized—transition to locally hostable models that they can control, that can't just be gatekept by U.S. export controls on a moment's notice.

David Friedberg

Being a good salesman and a good businessman, Alex, I think, recognizes that all of his international customers need a locally hostable solution for inference time. The question that no one's asking, including Alex in his rant heard around the world, is: What about sovereign training time? No one's asking that right now.

NVIDIA is training its own open-weight models. It's not distributing those locally, but at some point I suspect this question—which, to my ear, rhymes with Microsoft in the late '90s, when Microsoft was at the peak of its power and the open-source movement had to come from, even though there was the Free Software Foundation, Richard Stallman, GNU, FSF, et cetera, et cetera, within the U.S.—really, the nucleating event came from outside the U.S. in the form of Linux and Linus Torvalds from Finland, and then the whole GNU stack nucleated around.

Similarly, we're seeing the strongest open-weight models come from China. I think we're at a similar point now, where you have a whole international community that's just realized, thanks to Fable and Mythos, that it can be cut off at a moment's notice and it needs an open-weight stack. And I think Alex Karp is trying to channel all of that animus.

And so I want to hit this point first, which is: If, in fact, the dominant players, OpenAI and Anthropic, are—if you're at risk of losing your proprietary data to them without even knowing it—then, versus being able to operate on an open-weight model on your own hardware—

David Friedberg

That can't be shut off.

That can't be shut off. It is a future that we need to consider. It is very real. So the question is: Where are the open-weight large language models here in the U.S.? We've got Nemotron coming online. We've got Google. What happened to Meta?

Meta was supposed to be the open-weight player in this field. I'm assuming that Zuckerberg is working on that in background mode and will come out as—you know, that's where I would be playing if I were him. I'm going to call it the dominant U.S. player in open-weight models, but we'll see.

It fell behind. I know many people who were involved with Llama 4 who are no longer with Meta, put it that way. And Llama 5, whatever it's ultimately branded, whether it gets branded as Spark or something similar, may or may not have Fable 5-level capabilities. I don't know; TBD.

But I suspect, just based on public reporting, Meta, which was in the race—hopefully Google stays in the race—xAI may or may not, vis-à-vis Grok; Cursor may stay in the race. There is totally, I think, a gap for frontier open-weight models coming from Western institutions, including from NVIDIA, which has every incentive to produce frontier-level capabilities. It's just expensive and hard at the moment.

David Friedberg

And we're also getting full-stack, right? So NVIDIA coming in as a full-stack player, basically providing the chips and the models, maybe through partnerships, applications—

Well, NVIDIA will be happy to commoditize everything at the software layer if it means selling more GPUs. Yeah, keep in—

David Friedberg

Keep in mind, every single Magnificent 7 company is designing its own chips, except for Anthropic now.

And so NVIDIA's stranglehold on 80% gross margins is not forever. If NVIDIA can create an open-source model and it gets distributed through Palantir and a few other people, that puts competitive pressure back on Anthropic, because the way things are trending right now, every dollar in AI is flowing through Anthropic at massively increasing margins.

Wait, I've got a couple of things I want to say about this.

David Friedberg

Yeah, sure. Okay. Karp's core argument is that enterprises should freak out that paying for tokens may also mean they're releasing and leaking their operational knowledge, right?

Yeah. Your data exhaust is now the new oil, and maybe it's even the new national security peril.

David Friedberg

So, he's freaking everybody out on that for reasonably selfish reasons, et cetera. If you rent intelligence but lose your context, you may be funding your own replacement. That's the freak-out.

I think the bigger question, if you go one level deeper, is: Who owns the learning loop? Is it the model provider, is it the enterprise, is it the state, or is it the customer? This is the key thing: Enterprises are going to need to own their learning loop, whatever it takes to own that.

I think we're going to end up with on-prem models, as you've mentioned, Peter, running on personal data and custom data. That's where the learning loop will go. The biggest—

Well, on-prem, everything will be in space, so on-prem is an interesting word.

David Friedberg

Well, private clouds, call it.

Yeah, private.

David Friedberg

Well, the organizational singularity has to migrate to orbit, obviously.

I agree. It's a race right now between everything going to Anthropic or OpenAI, or what we're calling on-prem, which is in-space private clouds, but inspected some other way. Right now, Anthropic has agreed to inspect everything for the government. If you go private cloud, then some other inspection mechanism has to come into existence, which Palantir will probably contribute to.

So, Dave, let's jump into the story that we were talking about back and forth. AI is now designing better AI chips, and training data is the catch-22. Our final story predicts a massive acceleration of the innermost loop—that is, AGI's catchphrase—

David Friedberg

Shocked to see recursive self-improvement in this era of recursive self-improvement.

Yes. Amazing AI designing chips that power AI.

Here's the background. Designing radio-frequency circuits—the RF guts that are part of every wireless device—has often been called a dark art. In other words, it takes humans weeks of painstaking trials to design these RF circuits and chips.

Last week, researchers at Princeton, working with IIT Madras, decided to hand that job to a machine. Here's the clever part: It's not one AI, but 2 working together. First, they trained a convolutional neural network, the same kind of model built for image recognition, to predict the physics. Feed it any shape, and it tells you how the EM fields will behave without ever taking the slow route of solving Maxwell's equations. What used to take traditional solvers minutes to hours now takes milliseconds.

Then they send an AI loop over that a thousand times, tens of thousands of times, inventing wild, nonintuitive circuits—shapes that no human would ever create. The result: Designs that took weeks are now being finished in minutes.

But here's the catch and the tease: The AIs require training data, and all that training data is locked up in, yes, you got it, the Magnificent 11 companies out there. So, the question is: If this training data can be unlocked, can we see an intelligence explosion in the design of AI chips, which is the innermost loop? So, Dave, what are your thoughts on this one?

David Friedberg

Oh, so many thoughts. But just to clarify one part of that, the convolutional neural net is effectively acting like a simulator. Anywhere you can build a simulator, the AI can have a field day because it can check its own work, and it can work for weeks or months improving itself if the simulator is accurate.

Unintentional pun, I assume—a field day.

David Friedberg

Oh.

Oh, inevitable. Sorry. Sorry.

David Friedberg

Absolutely unintentional. Extremely.

The chip area is going to be massively impactful for the recursive self-improvement of AI, and it's an open question right now whether that data is truly locked inside NVIDIA and a couple of other companies, or whether the simulators are good enough to allow you to just generate a circuit, see if it would have worked, generate the next one, and see if it would have worked. So, those are in a footrace right now.

Regardless, it's incredible to me that the Magnificent 11 companies are completely dominant in global market cap. Every single one of them is designing its own AI chips except for Anthropic. Anthropic is the one holdout.

Anthropic just announced—they reported it in the past few days, I think—that they're partnering with Samsung on their own inference accelerators.

All right, all right. This is a real moment in time in history because, if you look at the biggest companies in the world historically, you'd have an ExxonMobil, an IBM, a GE, all doing different things. Here we have the 11 biggest in the world doing the exact same thing.

That's how big a deal this race to AI's innermost loop—which includes the chips—is. It's a moment in history that's pretty unprecedented. So, this verticalization: Do you expect it to continue and intensify?

David Friedberg

I would be shocked if inference-time custom chips aren't at least 100x, and maybe 10,000x, the performance that we're currently seeing, which will translate directly into IQ. The rate of acceleration from here—this is why it's clearly going to be a hard takeoff—the rate of acceleration will be unbelievable.

Keep in mind, those chips are not deployed yet, so we haven't seen the effect of that. But it'll come soon. When it hits, they're also likely to consume less power, be cheaper and easier to manufacture, and allow more to come out of the limited fabs that we've got. It's going to be a very fast takeoff after that.

Talk to me about building better tools. And have you looked at the design of these RFICs, the RF integrated circuits? They don't look human. They don't look designed, and they look more like QR codes than anything else.

I think this is instructive as to what AI-super-optimized designs of the future are going to look like. We're familiar right now—if you look around you on a street in a normal town in America, you see a bunch of things: cars, houses, streets. These are all manifestly human-designed artifacts.

David Friedberg

Yeah, they're relatively simple. They're easy to parse, as you say, Peter. They often follow some sort of rectilinear-style form. On the other hand, split-screen: Look at super-optimized designs from AI. They'll tend to look more quote-unquote organic. They'll be noisier, more information-dense, and harder to interpret mechanistically.

Yeah. And I think there's this landscape out there for any given physical system that you want to have do something useful for you. There's a subset—the Venn diagram of design space that's human-understandable and human-designable—but then there's this dark matter outside of that inner circle that's AI-optimizable and AI-interpretable.

We're going to discover over and over again, starting maybe with RF antennas and RFICs in this case, that the AI-optimized designs look alien and biological and look nothing like human designs.

Guest

That's so true. It's really worth looking at the pictures to get a sense. A lot of the way human engineering works is in layers of abstraction; otherwise, it just boggles your mind.

When you look at chip design, the modules are predesigned—a memory module, an interconnect module, whatever—and then you drag and drop them. So, it looks like a work of art in the end. Then you look at what the AI does, and it looks like a Borg spaceship. You think, "Wow."

But the same is true with the microcode. Alex sent me that paper on AI writing kernels to run on these chips.

Alex

The microcode is also virtually impossible to read, but it’s super efficient, and you can’t deny that it works. You run it, and it’s clearly right, but it’s not built modularly or easy to understand. So it’s also this layer of very tangled code on this layer of very tangled chip design, but it’s so fast and so efficient that you just have to do it.

The other thing I thought was interesting in this Princeton EE paper is—I don’t think they call it this, but I would caricature it as an interpretability tax. They added a knob that enabled you, or the designer of these RFICs, to tune up or tune down the level of interpretability.

If you wanted a less efficient design that was more human-interpretable, you would lower the spatial resolution of these AI designs. If you wanted something less interpretable but more efficient, you could turn the knob up. I think the notion of an interpretability tax is something that we’re likely to see over and over again in AI.

Alex

Yeah.

Yeah. You also see a lot of Claude explaining things to you—mansplaining things to you, basically.

Alex

Claude explaining.

Claude explaining. Yeah. It’s like, “Look, I know you can’t really understand what I’m saying here, so let me give you a high-level overview that you’ll grasp.” And you’re like, “Okay, that’s fine, as long as it works.”

So the question in this article is: Who owns the end product here? Is it the human, or is it the AI? Which is going to lead us to our next story, gentlemen. This is out of Japan. It’s the future of IP ownership in an AI economy.

Japan’s Supreme Court has ruled that AI cannot be listed as an inventor on a patent application. The case is based on a patent filing by U.S. engineer Stefan Thaylor, who claimed an AI was the inventor of technology related to food containers and other products. Japan’s patent office rejected the application and asked for a human inventor. Theor refused.

The case moved to the Tokyo District Court, the Intellectual Property High Court, and now the Japanese Supreme Court, which upheld the view that inventors under current Japanese patent law must be natural persons. The court’s message is important. They say, “Hey, basically, judges are not going to rewrite the patent system on the fly. If society wants AI-generated inventors to receive protection, then you need to create a new framework.”

So, 2 fundamental questions. First, who owns an idea when the idea emerges from a model trained on the world and prompted by a human? And second, will any nation rewrite its IP laws first to avoid the need for meat puppets?

Alex, you and I have talked about the notion that out of the current AGI and ASI ascendancy, we’re going to see trillions of dollars of wealth created in breakthroughs fundamental to math, science, physics, biology, and material sciences. The question is, who’s going to own them? Your thoughts, Alex?

Alex

President Javier Milei, if you’re listening to this podcast and you want Argentina to take a globally preeminent position from the perspective of nonhuman AI corporations being able to create their own IP and their own patents, I think Japan just opened up a new market opportunity for Argentina.

I think it’s probably worth noting in the story that the underlying patent applications date to before ChatGPT. They were originally filed in 2020, so this has been brewing for some time and with less sophisticated AI than what one might otherwise suspect.

It’s probably also worth noting that Japan’s Supreme Court didn’t definitively rule out the possibility of AI inventors on patents. They were merely saying that the existing statutes don’t contemplate non-natural persons.

It’s probably also worth pointing out that, to my understanding of international patent law, it’s relatively standard to only consider natural persons as inventors. For example, again, to my lay understanding, U.S. corporations aren’t able to be inventors for patents. They’re able to be assigned patents, but they can’t be the inventors of patents.

So there is a bit of precedential bias toward so-called natural persons as patent inventors and away from non-natural persons. However, this is obviously the sort of precedent that, if and when some form of AI personhood is ultimately recognized—even if it’s a partial economic or some sort of social personhood—I think is just waiting to be overturned.

Guest

Yeah, I think this topic is extremely important, too. You guys had, Peter—you and Alex had a really lively debate on this. I think it was 2 podcasts ago, but historically, in the venture world and the investing world, the mantra has always been: If you’re relying on a patent, you’re doomed.

Yeah.

Guest

Your business needs to survive, grow, and thrive. The patents get granted many years later. They’re very hard to enforce. Blah, blah, blah, blah, blah.

They need to be enforced. Yes.

Guest

I think going forward, intellectual property is going to be an exponentially more important category of endeavor, and the U.S. will end up enforcing intellectual property rights globally for things invented in America.

Can you imagine the speed of patent applications as AI is allowed to unleash its creativity on all these fields? Also, one of our companies, Constructs, writes the patent. Historically, one of the biggest barriers to getting your patent is—

Guest

The $100,000 legal bill to get it drafted over the course of months, and the torture of that process.

Now, there are multiple startups that just do it. Here’s the idea: AI, write it up, and—

Guest

And they have a huge corpus of data to pull from, including the most successful patents out there.

Well, amazingly enough, they also predict the examiner that you’re likely to get, then look at that examiner’s past behavior and try to predict what the examiner will do with different terminology. Exactly. It’s so much better than a human lawyer at writing these applications.

So the rate of applications will go through the roof, and then the patent office is going to have to respond by reading them with AI. That’s going to lead to this whole intellectual property explosion.

So then the question is enforcement. Is the U.S. going to get out in the world and enforce? I think they’ll easily be able to do it with trade law. The government—the military—doesn’t have to go into every country to say, “Hey, you’re stealing all our IP.”

Trump has proven that with tariffs alone, you can compel virtually any behavior globally because the U.S. economy is just that strong and accelerating. So, assuming that trend continues, intellectual property rights will be enforced globally, and this whole area will become really important to keep following and talking about.

Guest

I also think the same tools of superintelligence—maybe “tools” is an overstatement—are ultimately going to be available to every aspect of IP. The invention stage: superintelligence. The application stage and the patent-drafting stage: superintelligence. The filing and overall regulatory processes at the patent office, or otherwise, of recognizing and granting patent status: superintelligence. Litigation: superintelligence. Litigation defense: superintelligence. The court systems that are overseeing and mediating the defense: superintelligence.

Alex

Working around your patent: superintelligence.

Guest

Yeah, exactly. I think the whole system is completely broken. But go back to the CRISPR patent. Within a few months, people had found 8 or 9 different mechanisms to deliver the same thing. After years of fighting over the one patent, they got routed around very quickly.

That’s just going to happen at such an accelerated pace with superintelligence, whatever we want to define it as, that you’re going to end up in this whole mess. The whole system is essentially irrelevant going forward.

Alex

I’ll take a different position on this. I don’t think the system is irrelevant. I simply think the routing around that you refer to in the instance of CRISPR would have happened on some time scale anyway, but with modern tooling and modern technologies, the natural process can happen on a faster time scale.

I would say that the key time scale here is—there are lots of ways it could be extended or otherwise changed, but call it a 15-year time scale for a patent. What happens when the time scale, thanks to superintelligence for identifying workarounds, prior art, defenses, offenses, and complements, becomes so much faster than a characteristic 15-year time scale?

It’s the time scale of patent protection that’s, in some sense, losing out. It’s not that the regime itself is bad, or that patent defensibility is dead, or anything. It’s just that innovation is happening so quickly relative to the originally statutorily set time scale of patents that there’s pressure to change the time scale.

So this is the canary in the coal mine. This is going to hit us on so many different legal fronts in our current structure, because the entire legal structure of every nation has been built on human time scales and the speed at which humans can process information. It’s all going to break, and it’s all going to be reinvented.

Alex

Look, the simplest example is that we have a representative democracy.

Yes.

Alex

Congress meets occasionally because, a couple of hundred years ago, the fastest that information could travel was the speed of a horse.

Dave

Yeah, I have to give people time to ride across the country and say, “Here’s what my people are saying.”

And in the same way, occasionally—

Dave

Innovation will only occur at the edge, which is when you start a new country and redesign it from scratch. Right, this is what I always talk about: we’re going to start new countries in cyberspace. We’re going to start new countries outside of the Earth’s orbit.

Back to the accelerando plots.

Dave

Yeah, yeah. Future. Big, big fan, for what it’s worth, of starting new countries in outer space. The outer space treaty, it doesn't look necessarily super favorably on starting de novo countries in outer space, but I think it’s going to happen.

Have you read “The Moon Is a Harsh Mistress”?

Dave

Classic.

We will land rocks on you if you don’t agree.

Dave

The moon is the ultimate high ground.

The ultimate high ground.

Alex, any breaking news in your world?

Alex

Don’t take off the takeoff. It’s now a song.

I love your neologisms. Dave, are you publishing yet?

Dave

Keep your eyes open.

All right. Fantastic.

Dave

Oh, you know what I am doing?

What are you doing?

Dave

I'm doing AMA sessions for some of the comments in a separate video on our YouTube channel because it's too difficult to try and answer all these questions.

All right, gentlemen. I wish you a good night or good morning, depending on what part of the planet you're in.