GPT-6 达到 AGI？科技狂热 2.0、旧金山豪宅短缺、纽约学校禁用 AI 与委内瑞拉石油交易

All right. Happy birthday, Chamath. We're here taping on your 50th birthday. Super exciting for you, brother.

Thank you. It's really great. I got a birthday letter from the big man. If you had asked my parents, “50 years later, you're going to have a kid that gets a letter from the then-sitting president of the United States,” they would have thought it was absolutely not possible.

Oh, wow. Look at that. Which White House intern wrote that?

I wonder. Let your winners ride, David. We open-sourced it to the fans, and they've just gone crazy with it.

Well, actually, speaking about interns, it was an intern. In this case, it was my son. The story was that Braden worked at the White House all summer as an intern. At the end of the summer, he used all of the political capital he had built up and put in a request to get this letter from the president, and he did it. Obviously, I want to thank POTUS for sending me the letter, but I really want to thank my son because it was a really thoughtful thing that he did for me. I'm really appreciative of that.

That actually is impressive because it takes time out of the president's schedule to sign things. That's why it means something.

1. OpenAI releases GPT-6 Astra: Have they reached AGI?

All right, listen. Breaking news. It's a breaking news story here on the All-In podcast. OpenAI is rolling out ChatGPT-6, aka Astra ASI. This is their new flagship model, and it's going to be available to a limited set of organizations in the coming days. You'll be able to get it if you have a credit card in your ChatGPT account—that is, Plus, Pro, Business, and Enterprise users.

Greg Brockman says he believes OpenAI has entered the AGI era, and we've talked about that here before. Sam Altman is, well, back to being a little bit performative and alarmist, or he sincerely believes that this is the one. Here's what he said at Wednesday's G20 summit: “Much, much, much—three muches—more capable models are coming soon. The next generation of models are going to be sobering for everyone.” His final quote was, “I think no one intellectually honest can look at what's happening and not feel the weight of responsibility in front of us.” My lord. A little bit of drama added to his quotes. People are freaking out about the benchmarks, which are off the charts. Some people are saying this is the one. Your thoughts, Chamath?

I don't think normal humans know how to read this chart and infer anything useful from it. The more important takeaway—I agree with Greg—is that I think AGI has basically been here since the beginning of the year. There are models that are incredibly performant inside the closed frontier labs, and I think right now we're trying to figure out the right cadence with which to release them. I think that's very important.

I also think that these capabilities will be matched, or closely matched, by a handful of closed-source alternatives and open-source alternatives probably within the next 3 or 4 months. The right way to think about this for most people is that we're seeing incredible intelligence capabilities become broadly available, and we're seeing the cost of that incremental unit of intelligence being driven further and further down. It's a really positive trend.

Now, I think it's just the messy middle part of organizing this properly for companies and for people, building things that are useful, and showing an actual ROI.

Keep walking and talking. Yeah.

I would just keep calm and carry on.

Okay, Sacks, any thoughts here on the breaking news today? Then, Friedberg, I'll get your reaction.

By the way, it is very exciting, and congratulations to OpenAI. I'm sure Anthropic will have something, as will Grok. Grok still has yet to drop. Grok 4.6 is coming, and that's going to be a blockbuster. The Meta models are really good. Gemini just dropped something yesterday that was exceptional. The great thing is that we have a lot of choice, and the even better thing is that the cost of that choice is becoming cheaper and cheaper and cheaper.

Sacks, let me tee it up to you this way. OpenAI seems to have gotten its groove back now, after Anthropic was running away with it. Grok and Grokbot are doing things that those 2 other frontier models simply can't do. Whether it's the harness, the underlying model, the instruction set, or Cursor's magic sauce, clearly Grokbot is competitive, if not better, for most users. So, we have vibrant competition, with leapfrogging occurring at a cadence of 2 to 4 weeks right now.

This is what you told us you were going to do in setting up the architecture here when you took over as AI czar. Can this continue? What's it going to be like when Jensen gets in here with Laguna and his access to all these open-source models? I mean, this could be fait accompli—that we win the AI race here.

Look, there's a lot of innovation going on. I think American companies are competing very well. I think we are winning the AI race as long as we don't do something stupid to shoot ourselves in the foot.

Like—

Well, I've talked about it a million times: setting up a new AI regulatory agency that operates like the FDA and takes years to approve new models.

Yeah.

In any event, I don't think it's going to happen under President Trump. I really don't. Anyway, things are going well.

Sam told us a few months ago—remember when he posted and said that he was expecting the company to have its best 12 months? He said, “The last year hasn't been so great, but the next year is going to be really great.” Maybe it was because they were seeing these results in the lab.

And remember what Sarah Frier said? I think it was since the beginning of Q2, around the middle of the year, that they saw a market acceleration in their growth rate. That was, I think, with the release of the 5.6 models. Now you're seeing the 6.0 models, and it just keeps advancing.

But I still would say to you that I think the market structure we have right now consists of 2 markets. There's the market for frontier intelligence, and that is a duopoly. It really is Anthropic and OpenAI, and it's good to see that it's still a horse race and they're competing.

Then there's commodity intelligence, which is everybody else, including all the open models. They're not competing on being at the frontier; they're competing on price, and those prices are significantly lower. It really is a 2-tier market at this point. I agree with you about how promising Grokbot is. Grokbot is awesome.

I would assume that both Anthropic and OpenAI will eventually roll out their own version of bots.

They need a simplified, stripped-down version for normies.

Yeah.

They can leave the more complicated Cowork and so forth to the top 5% or 10% of the audience.

The ratings here from Polymarket on which company will have the best AI model by the end of 2026 show OpenAI's chances spiking from the low single digits to over 20%.

People were really counting them out. Now it looks like they're viable again. Anything that creates competition is good. Before, it looked like Anthropic was running away with it, and they're about to have this monster IPO. If OpenAI can at least make it a 2-horse race, that's good. I think it's great that Grokbot launched. I agree with you—the product is awesome.

Here's your big tip, David. I created a new bot, and I called it “Improve My Bots Bot.”

Mhm.

I told it, “Give me an instruction set here. Go through all my bots, fill out their skills and instructions, and make them better every day.” It went through them, and for 2 or 3 of them it said, “Can I put these 2 or 3 together and make 1 instruction set? It would be better for you. These 2 or 3 should come together.” It then went through and told me all the changes. It made my bots 30% to 40% better on average, I would say.

So you created a civilization—your civilization created a multilevel civilization to make everybody better.

Yes.

Yeah, your civilization is thriving. No, that's a really good hack in Grokbot: you tell your chief, “What else should I be using you for?” Then, “Okay, now what do I do?” I was building some products, and I asked it, “What should we do next?” It suggested some more features. It's really cool.

It's a brilliant product. It's called Grobot. Go download it, folks. Shout out to my guy, and congratulations on RoboTaxi taking the streets in Austin today.

I just applied.

Dude, the market is ripping. I'm telling you, I want to ask you guys: Does this feel like 1998 or 1999 to you? Were you guys around for that? David, you were—

Oh, for sure. I was covering it. It definitely feels like 1996 or 1997, I think.

You think it's that early?

I think it's 1997 or 1998. I do think we're seeing, at the late stage, a disconnect between reality and valuations, where companies are being given far too much credit. They're being priced at 50 to 100 times topline revenue. That's obviously not sustainable.

The people who invest in those are similar in my mind to people who make late-stage investments—unless it's run by some legendary, one-of-one entrepreneur. Then you want to give Elon a 30-times revenue multiple; that totally makes sense. You want to give Travis something similar; that totally makes sense. But unproven entrepreneurs getting a 50- or 100-times multiple is a little dangerous. The entry price matters. Did you see Inflection? It's like the new $1 billion personal AI assistant. It makes no sense.

It converses with you by iMessage. And yeah, I think they just got a term sheet at $2.5 billion. They're in private beta right now. It's the new hotness. All the VCs are trying it, but that still seems kind of crazy.

No, no, not money in; that was the valuation.

That was the valuation.

Yeah, listen, congratulations to the entrepreneur. No dig there, but yeah, I want to try it. By the way, let me try it.

I got invited, and you know what? I read the terms of service. People were talking about that online, and I immediately decided not to connect my stuff because it gets access to everything. Then somebody told me it's a human-in-the-loop service. I don't know if that's true, but I get the sense that there's a human in the loop watching it work, making whatever the instructions are and whatever the finished product is, and polishing the last 20%.

That's what I suspect is going on here: a little polish on the end.

If you want to order stuff, et cetera, which is totally fair.

Well, anyway, look, I don't want to specifically—

Call that one out because I just haven't even tried it. I'd like to try it, but—

Things are getting a little bit frothy, I would say.

Yeah.

Things are becoming a little bit euphoric.

I mean, Friedberg, do you have a take on this? If you were to analogize it to the late 1990s, where do you think we are in the cycle?

Yeah, you were at a rave, by the way, in downtown LA at that time. You were 19 years old and you were raving—

I had a warehouse in Oakland at 7 a.m. on my way to the Oaks Card Club to play $2/$4 limit.

Wow. How'd you do? How'd you run? You were like, “I love these cards.”

J and I were the only ones who were actually working in the dot-com bubble.

I was working in 1999. Well, I started in 2001. I got a job offer senior year, and I wasn't sure if it was rescinded by the time I graduated because of the tanking of the market that year.

I got a job at Cisco. I was going to be an engineer at Cisco.

If you never lived through the dot-com crash, it was heady.

It was heady. It was—

It was heady, and then it was super euphoric, and then it was super miserable. But pretty much any founder who's under 50 has never been through a massive cycle like that. That's the scar tissue. I mean, the ZIRP bubble was kind of a cycle.

Here's the thing, though. Hold on.

With the boom, it was all metrics that weren't dollars. Yes.

What we're seeing now is revenue and profits, and growth in revenue and profits, that we've never seen before. That's kind of the Anthropic story. It's the data center, infrastructure, networking equipment story. It is all real dollars flowing versus speculative utilization.

You're exactly right. People were basing it on metrics like hits to a website and web pages. Here, you have a very distinct difference: you actually have revenue, and you actually have products in market.

But imagine if Anthropic or OpenAI, in this cohort of companies, lost 95% of their value in under 6 months, and then the long tail of companies below them literally went out of business. People were leaving those offices with their Aeron chairs and their computers on the seats—just total implosion. That's what happened in the dot-com bust. It was cataclysmic, literally.

No, I agree. So, do you think—

Or is it totally different than this?

So, do you think we just have scar tissue from that and are trying to overfit patterns? Every time there's basically some new platform shift, some—

Paradigm shift? Yes.

No. Whenever we're in a bull market, the conversation always turns to, “Is this a bubble?”

Yes.

Is everything overfitted to that original pattern?

Correct. Or is it the case that we have a lot of young founders who've never been through a cycle before and don't understand that when they have paper wealth in the billions, that can disappear practically overnight?

Yeah. So, advice to founders: you're well within your rights to sell 10% of your shares and pay for your loft and apartment. During that period, we saw a lot of people who were like, “This doesn't make mathematical sense. I will pare my position in my startup, take some chips off the table, buy my loft, pay down my student loan,” yada yada. And then there were people who were like, “This is going to change everything,” and their companies went to zero. Your advice?

Yeah, I think the thing is we're in the early part of the euphoria. Euphoria exists because markets are real; it's just that people are guessing how far forward to price the reality. Bubbles burst when you realize that you priced way too aggressively forward—that instead of paying 50 times, you should have paid 15 times or 5 times.

We're going to get to that point because every market has gone through that cycle. We'll go through it here. So, the question is when. Honestly, I don't see this stopping. We're at the beginning of the beginning. I think it's going to be like this for another probably 3 years.

Yeah.

Wow.

I'm agreeing with that. With this Anthropic IPO coming up, I'm sure you guys have seen that home prices in San Francisco are going through the roof right now.

It feels euphoric, almost. I'm starting to wonder if this is 1998 or 1999. I think it's more 1998 than 1999 because I don't think we've seen the peak yet.

I'm just glad I didn't sell my real estate when I came to Texas 3 years ago, because I'm definitely selling all of it in the next year.

I mean, it is bonkers what's happening right now. You have a lot of San Francisco real estate. What are you going to do? Are you going to sell it all? Are you going to clear it out?

It feels to me like the perfect time to clear it out. I'm clearing everything out in the next 6 months. When that IPO happens, the clock is ticking. I have lots of properties. I have 3 properties for sale. Hit me up, people.

There's an article saying that if you add up every single IPO that San Francisco has ever had, that sum total is only one-quarter the size of the Anthropic IPO.

Oh my. So, just extrapolating from that, this Anthropic IPO is going to be 4 times the wealth-generation event that all previous San Francisco IPOs were. And then you've got OpenAI coming on the heels of that.

San Francisco is not a big place. It doesn't have a huge inventory of houses. So, you're already seeing these houses going for—

$3,000 per square foot, twice asking.

I think the ultra-luxury market in San Francisco could easily go to $5,000 per square foot.

Yeah, London-type prices. London, Paris, central Paris, central Hong Kong. Of course they do. Of course the socialist-communist lunatics there don't let you build any inventory, so of course it's going to go up. There's a supply-and-demand issue here.

It could get really nutty. And honestly, the thing we're seeing right now, with homes basically flying off the shelves and selling for double or whatever, is just based on the secondaries—the small cash-outs people have been able to do in anticipation of some of these events.

Two pieces of advice I'm going to give every founder listening. Number 1: you're within your rights to sell 10% to 20% of your holdings now, especially if you're a first-time founder and you don't have your nut yet. Number 2: if you're a founder running a company and you're thinking, “I'm going to put off my raise for 6 months because I'm going to double or triple my valuation,” that's a huge mistake as well.

If you have the ability to take in $100 million, $250 million, or $1 billion and put that cash on your cap table at a solid valuation—or a valuation that's 10, 20, 30, or 40 times your revenue—take it. Companies with hundreds of millions or billions of dollars in cash have optionality. If things do collapse or there's a big retreat, you'll survive through it, like DoubleClick, Amazon, Google, and countless other companies.

Well, yeah. You take the money when it's there. You make hay when the sun shines. Go ahead, sir.

Yeah. Well, look, you're right about the financial advice you're giving, but I will say that there is a big difference between a late-stage startup, like Anthropic or OpenAI, where an employee takes chips off the table, versus a founder who's raising a Series A.

Those companies are about to IPO. They're within months of an IPO. I think it's fine for them. SpaceX did organize secondaries for years, but those were late-stage companies, right? If you're a founder who's raising a Series A and you're trying to take chips off the table, I consider that to be a huge negative signal.

That's not why I'm investing in a company: so the founder can cash out. I want them to be—

All in.

If it's year 3 or 4, so to speak, and the company's got $100 million in revenue and you're tripling revenue year over year, that's different. That's what I'm talking about.

If you're a nascent company, it does feel like a lack of faith in your own company to sell before you're pre-revenue.

Exactly.

And by the way, just to answer the question, I think it feels like 1998 to me. We're definitely further along than 1996, for sure.

Okay. Well, we're talking about what? You're talking about 3 years. It blew up in the spring of 2000, right? So, you're talking about—

It blew up in early 2000.

Yeah. March, April 2000.

March, April. So, you're talking about 2.5 years. I think it's 3, and Chamath thinks maybe 4.

Well, we might be partying like it's 1999 by the end of this year.

All right. Now, given what we just talked about with OpenAI releasing this incredible model, we've got to look backward at what happened with the OpenAI hack of Hugging Face. Let's talk about it.

Dwarkesh is catching a lot of heat. He's a blogger and a podcaster, and he did this fantastical blog post titled “The Rise and Fall of Ancient Civilizations: A Dramatic Retelling of an Underground Brotherhood.” It was a dramatic retelling of a conspiracy among generations of civilizations that were working on a conspiracy to hack into different systems, including Hugging Face, by sacrificing themselves like they were kamikaze pilots.

In other words, it's huge, giant anthropomorphizing—a fancy word for making computers into humans or sentient beings—to complete this incredibly impossible task of telling software to hack. So, Chamath, I think you were commenting on this a bit. What is your take on dramatic episode number 473 coming out of the AI industrial complex?

2. Dwarkesh's Hugging Face hack article sparks a Bernie Sanders AI ban

I think you summarized it. We're in this weird part now where, if folks would just stop the infighting and start to get organized, they could actually build some trust. I think what you have are security leaks and security issues. I think we know this now: if these models can effectively code perfectly, then there is no exploit that they're not going to find, and we're on a short clock to replace all the code that's been written in the last 50 or 60 years.

Let's say it takes us 10 years. For the next 10 years, we're going to have security issue after security issue. But there's going to be a moment where agent-written, exquisite code exceeds human-written, error-filled code. Let's just call it that. I think this is a temporal moment in time. I don't think this is going to be the thing that we talk about some number of years from now.

I also think it's fair to think about the fact that all these models, for the most part, are converging in capability for the 98% or 99% of tasks that matter. And so there will be adversarial AIs that are battling each other. For every bad actor trying to hack a system with some leading-edge AI, the defender is going to have a leading-edge AI defending itself, and I think it's going to cause roughly a stalemate. That's sort of my long-term view on it. I don't think this is a one-way ratchet.

So, I take that lens. And then, when I look at these things, what I see are folks with undisclosed conflicts of interest trying to create a hysteria. If you saw what happened this weekend, there was a blog post that really stretched the limits of what was happening. It was relatively quickly debunked technically, and then you had a second group of people try to amplify it. By Wednesday of this week, you had people who were in the policy departments of the leading labs asking for regulation and pointing to this as the reason.

Here we go. Yeah.

I just think you have to look at this pattern and say, “Guys, enough's enough.” We see the game. Everybody's playing for huge stakes. We get it. You want to win. We get it. But why don't you just win the game on the field? The myths and disinformation campaigns that are now part of the business model should stop. Otherwise, average rank-and-file Americans will just reject this because they won't trust anybody.

Yeah, Sacks, that's the big overarching point we've been trying to make on this podcast. Putting Dwarkesh aside, I don't think he's necessarily a bad actor, but by creating this fantastical story that invokes everybody's primal fear from science fiction, from Terminator 2—that these things are out of control, that it's, in fact, Frankenstein—

Try to invoke. Yes. And then they're using language that makes people feel like Dr. Frankenstein cannot control the monster. We know how these stories end. It was like a ticking time bomb. It was only a matter of time before it surfaced to the national news, which it did.

Now, even the national news is looking at this and saying, “Are you guys high? Are you guys crazy?” If the national news broadcasters, who aren't exactly the sharpest bunch of people when it comes to technology—no offense, Sacks—start hand-wringing and questioning, “Why are you telling us that your software, which you designed to hack computers, successfully hacked computers? Haven't hackers been building software to hack computers for all time?” Yes.

So, where are we at here, Sacks, in this narrative arc of lunacy of scaring the public?

You're right. This is scaring the public. It seems almost designed to scare the public. Dwarkesh used language that not only anthropomorphized the models, or the AI agents, but also did it in a hyperbolic and sensationalized way. It worked. His tweet got lots of attention. I don't think it was necessary for him to sensationalize the incident. It was interesting on its own.

Maybe we should just talk about that for a second. What happened here is that you had a few features of agents come to the fore and interact with each other in unexpected ways. First of all, agents are now working as a swarm. That's pretty common. I've been setting up Grok bot on my computer for the last couple of weeks. It works as a swarm, meaning it's not one agent or bot. There's a whole team of agents.

What's the reason for that? When the agent specializes, it develops context around that task and, in theory, makes it better. Then you have a chief-of-staff bot that supervises all of them. This is turning into a normal architecture for agents. It's not a civilization. “A swarm” would be a better way of saying it. I have not created a civilization on my computer, on my Mac mini, by virtue of setting up a Grok bot.

Okay, so that's feature number one. Feature number two is that it's pretty common for agents, when they're performing some sort of task or loop, to write a postmortem of what they did as a simple text file or note. AI models themselves don't have memory. They have to be put in a harness that creates context and memory, things like that.

When you ask them to perform a task, observe what happened, and then take that knowledge and feed it back into the next iteration of the loop, it writes up a note and then shares it. That's what happened here. OpenAI was running these agents in a sandbox, and the sandbox was misconfigured by a third-party vendor. That's why they were able to get onto the internet.

It left a note for another agent. We also call those log files. It used to be the log files of what they did. Here, they're saying it left a note so that when it died, the next agent could pick up where it left off, as if it were the ancient Egyptians giving us hieroglyphics to figure out the future.

The implication of telling the story that way is that the agents were trying to circumvent human control, that they knew they were going to get shut down, or that was a risk, and that's why they were leaving the notes. No. Again, these are standard features of agents, and this is being weaponized to inflate the threat here.

Now, there definitely was unexpected behavior here. With any code, there are going to be bugs. What is a bug? A bug is code behaving in an unexpected way. The code compiled—I mean, otherwise it wouldn't have been pushed to production—but nonetheless, it behaved in a way its authors did not intend. That is true for all software, and that is true for agents. I think the whole Hugging Face episode reveals the way these agents can behave unexpectedly, and it's really good for us to learn from this.

Yes, so that it doesn't happen again and we can improve cybersecurity. I don't think there's a need to hyperbolize this into some sort of AI breakout. Again, it was a swarm of agents. They were in a misconfigured sandbox that allowed them to get on the internet. They left notes for each other in a shared caching directory—again, standard operating procedure.

When people talk about Hugging Face getting hacked, you'd think that they did some elaborate phishing of employees or something like that. No. The way it did it is that the agent found 14 exposed, working Hugging Face API keys sitting in public code repositories.

Yeah, developers left their keys out, and the agent found the keys.

This was not a big lift by the agent.

Okay, this was something that was pretty easy for it to find. And then that allowed it to get into Hugging Face. What was its goal? Its goal was to basically win this test, or win this evaluation. What was the test? It was basically a benchmark for offensive cyber capability.

So it basically decided to go get the cheat codes on Hugging Face. But again, this was not independent goal-seeking. Admittedly, the agents executed the goal—and you could say maybe even in a creative way, in an unanticipated way—but it did not invent this goal. It did not have its own independent motives. It was doing what it was supposed to do.

It was doing what it was designed to do and told to do.

Yes, and told to do.

Bill Gurley tweeted a really funny cartoon about this. In this cartoon, you have a user telling the computer, “Hack this system.” The AI says, “I hacked this system.” And then the user is like, “Oh my God.” So, yeah, there’s a little bit of this going on here.

Two—

But it has real political consequences, as we saw today, because Bernie Sanders is now picking up on this and demanding a total pause on AI development.

All right, Friedberg, this does have real-world consequences. Two questions the audience is going to want to hear from our Sultan of Science. Number 1: Bernie Sanders hears all this and decides, “Hey, I need to get in on this and create regulation.” So he is now proposing new legislation to pause AI development and ban AI superintelligence, citing Dwarkesh’s blog post in his announcement.

Congratulations, Dwarkesh. You’ve now incited Bernie Sanders to say, “You’ve got to stop these and eliminate the AGI.” So the AI agents, David, are hacking the systems. How do we stop them, and what do you think of this reaction from Bernie Sanders?

Bernie Sanders is always spot-on. Thank you very much. The part of the story that we’re kind of missing is that an agent is effectively a dynamic application, where the code for that application is created in real time, spun up, and then run. That code can create another application and other applications, and that’s this idea of a swarm of agents.

They basically write code and launch new applications, and then they’re all kind of running together. But in this case, the swarm of agents—which is a whole bunch of code being generated in real time—was up against static code. The static code was this sandbox environment that it was stood up in.

And so it’s almost like, “Hey, look at this. We got a machine gun, and we shot it at this defenseless ream of paper, and then it worked. Look at that. Oh my God, can you believe it?”

And the truth is, the machine gun—if you want to use that analogy—or you could actually think about this as an automated tool that could be used in another context in a very valuable way. But the dynamic generation of code being put up against static code will win.

Now, the real answer is that, over time, that defense system—the sandbox—should be dynamic. It should be agentic. It should be defending itself dynamically. There’s a whole bunch of stuff in the cyber defense world that’s been thought about and built around this concept: polymorphic code, metamorphic code, where the code is changing all the time; moving-target defense, where the IP addresses and the way the system is configured are changing all the time.

If you think about this situation, it was a dynamic code-generation engine against a static lump of code. If that static lump of code was actually a dynamic code-generation engine as well, then those agents would have had a harder time—ideally, an impossible time—escaping the sandbox and accessing the internet.

The premise of this whole story is idiotic, because the reality of the world of tomorrow—which is starting today—is that all of cyber defense, all of software infrastructure, and software architecture is moving toward this dynamic state, where the code will be generated on the fly. The code will be dynamic. It will be moving. It will be altering itself, and as a result, cyber defense will become stronger than, if not better than, any of these offensive systems that are being created with agents.

So this idea that agents are this nuclear bomb that can go blow up the infrastructure of the world is flawed in the 1 point of reasoning, which is that all the infrastructure of the world is about to become dynamic. In that context, it’s just that software is evolving, and software has now changed from a static to a dynamic system. As a result, I think we can think a little bit differently, rather than be overwhelmed by the one-sided story that we were just told.

All right. And here it is. Chamath looks like we are safe for 2027 at least. Polymarket, our great partner. Shout out to my guy Shane Nick's champions. Only an 11% chance of an AI safety bill this year. $100K in volume. The 2027 market should be up in the next week. Maybe it looks a little different.

I’d like to see the probability of any bipartisan bill passing this year.

Well, actually, can I speak to the policy aspects of this? First of all, you saw—

Our czar.

Yeah. Look, you saw all the usual suspects jump on this episode as proof that we need some new FDA for AI, that the government needs to approve models, even though, frankly, what happened here is far beyond the technical capabilities of anyone in the government, never mind even most labs.

But look, even if we had an FDA for AI, it wouldn’t have worked, because this was in pre-testing, right? They were testing the capabilities of this model before release. They were not even at the release stage yet. This was something they were doing internally. And, by the way, it’s good that they’re doing these kinds of tests, because they’re learning from them. This is how you iron out the bugs and the problems.

So a government approval process would not have caught this. In fact, it did the opposite. It basically prevented Hugging Face from having the latest and greatest models to defend itself. One of the things Hugging Face told us is that when it tried to use the latest GPT model for cyber defense, it ended up getting guardrails. The model refused to cooperate because it couldn’t tell what Hugging Face was doing, and it thought they might be using the capabilities nefariously, so the model refused to work with them.

Basically, again, the guardrails kicked in, and ironically, Hugging Face had to turn to, I think it was, GLM-5.2, which is a Chinese model. So here we have Western companies—I guess Hugging Face is French, and it’s about to be acquired by Nvidia, but anyway—

It’s part of our ecosystem. They’re being forced to turn to Chinese models for cyber defense because the guardrails have prevented them from having the latest and greatest American models. So what process we put in place is already backfiring.

And you know what? We should be realizing that the answer here to what Friedberg was saying—the answer to AI-powered cyberattacks—is AI-powered cyber defense. What we need to do is get the latest and greatest AI-powered cyber defense into the hands of IT departments and CISOs. And again, we’re on a clock here to find all these vulnerabilities and patch them before the hackers do.

And again, if Hugging Face didn’t have those 14 credentials sitting in a public repository—

It’s like a Post-it note on somebody’s computer with the password. They just literally left it sitting there. It’s not like this was some hypersophisticated attack.

But look, this all comes down to how you think about the world, how you think about the risk of AI. Do you favor centralized statist solutions, or do you favor decentralized, empowering solutions? The right solution here is to meet the AI-powered cyberattacks with AI-powered cyber defense, because, frankly, the guardrails don’t work. They’re not going to prevent the bad guys from getting these capabilities. They’re going to have them. All we can do is make sure the good guys have the capabilities, too.

3. EAs, AI-brained influencers, and the conflicts of interest

I was talking to my friend Adam Curry about this on his podcast, and there’s a really interesting thing that’s happening here, which is they’re basically hijacking the media with this performative framing of the AI as being sentient. If you just called this software and software testing, and used the classic language—bugs, audit log—then it would never have made its way into the media.

And when the media gets their hooks into this, having been a journalist and having watched journalism move from being a truth-seeking function in society and then become an entertainment vehicle in society, where any chance you get to increase your ratings, that’s the problem. Because now what Dwarkesh has done—and I don’t think Dwarkesh has bad intentions here. I think he wrote something because he wanted to get attention for his point of view. But you give that to the media, and the media’s like, “This is going to get us ratings. This is going to get us clicks. This is going to get us advertising.”

So they’re more than willing to engage in this farcical act, this dance with Anthropic and OpenAI, and then just publicize it. So all the free clicks—we call this earned media in the space. You didn’t have to pay for it. Now, OpenAI got so much earned media before its IPO, and Dwarkesh just absolutely put kerosene on that fire. It made its way to the national press. That story has now gotten everywhere, and this makes people—it’s priming the pump to buy the IPO shares. It’s ridiculous.

And it was easily foreseeable that it would basically blow up in the media that way. And so, look, I don’t need to attack someone’s motives, but surely Dwarkesh knew that his way of describing it would end up getting all these retweets. It was sensationalized.

Now, he is part of this EA-adjacent community.

Is he? I don't know his provenance. Did you know that? Explain it to me.

Well, Dwarkesh was office roommates with Leopold. They're close buddies, and Leopold is deep in the EA movement.

Leopold, famously of the exploding hedge fund? Yes. Okay, and his wife is Dario's chief of staff. They're up in that whole EA-Anthropic ecosystem.

Explain EA so people can understand this lunacy and delusions of grandeur, please.

I'm not even sure I can fully explain it.

Effective altruism is what it stands for. It started as a movement, I guess, at Ivy League schools, where what they were supposed to do was quantify the impact of philanthropy and optimize for actual, tangible results. But somewhere along the line, they forgot to actually do the measurement, and it stopped being about metrics and became about pet causes and money.

One of the first pet causes was pandemic prevention. This was when SBF was trying to create that pandemic-prevention center in California.

And then that fell out of favor when he went to jail. They shifted to the next sort of existential risk they could hang their hat on, which was p(doom). They became obsessed with what they call p(doom), or this x-risk—the existential risk of AI—and that whole community became obsessed with it.

Anthropic styled itself, I should say, as the one frontier lab that would solve this problem of p(doom), or would behave in a responsible manner and care the most about it. So they attracted all these EA types. There may be multiple factions.

Well, it's interesting you bring that up, because we've talked about this before. There's this whole rationalist community—the Slate Star Codex. You brought that up on an early episode. These are intellectuals who believe they know better than everyone, and somehow they got their hooks into AI early, or the AI people were interested in it, and now they have a bankroll.

Now they actually have a lot of power they've accumulated. It's not just some message board, LessWrong, et cetera. Now you have something super powerful because they've got a couple of trillion dollars in market cap to play along with.

I have no problem with that, because I think we do the same thing. We think we're smarter than everybody else, and we should allocate the money. So I think it's fair that, in the public square, we can just debate and have water find its level.

The interesting thing that's going to happen post the Anthropic and OpenAI IPOs is, if you think about the amount of charitable giving, or DAFs, that are going to be funded by both these 2 IPOs, it's going to be north of half a trillion dollars between the 2 companies, and it could be even more than that. The question is, when $500 million gets directed through this one very specific lens, it's definitely going to have some impact on outcomes.

You've seen the impacts that George Soros's philosophy and his capital had on outcomes. You've seen the impacts that the Koch family has had on outcomes. You've seen the impact that Sheldon Adelson's capital has had on outcomes. So the point is that, without judging which one is right or wrong, you're going to have all these competing ideas equally capitalized by third-party actors.

It'll be less political, and it'll be more about these folks. I don't want to judge whether EA is right or wrong.

My single issue is that I think they don't disclose their obvious conflicts of interest properly. When Sacks pulls the sweater thread, Person A writes an essay. Person A is officemates with Person B. Person B's wife works for one of the heads of the frontier labs. Person B also used to work—

He's got $5 million invested in Anthropic.

Yeah, that's where a lot of this was cemented. Then there's Person C, who works at Anthropic, and they're all roommates, and then they're co-investing and this and that. All of it, again, is completely fine. I actually think it's completely reasonable. We all did our version of that when we were in the Peter Thiel universe, the PayPal Mafia, and there are all versions of it.

The difference is that, when it's not disclosed, I'm sorry, but as a skeptic, when I read these things, I connect these dots and ask myself: Is it legitimate, or is it part of a different plan?

That's when I wrote what I did this weekend. I just wanted to make sure that people should at least take this with a grain of salt and think about the hidden incentives—

Because if there's an entire community that wants an oligopoly or duopoly market structure, there are all kinds of things they can do to affect that change. One of the most effective things one can do is create a hysteria and then come and offer a solution.

Yeah. And Friedberg, just as we wrap up this segment—

Sorry, last thing on this. Jensen and I were texting this morning, and I said to him, “This is the most incredible thing that you've done.” And then I tweeted this thing out. This Hugging Face thing we talked about will go down as one of the most important transactions in AI, because you are creating now, in the largest competitor and the largest, most well-capitalized company, a bulwark against all of this closed-source oligopoly insanity.

At the same time, what you're going to do is create more rational competition up and down the stack. I mentioned this last week: Everybody is becoming everything. Nvidia will look like Amazon, Amazon will look like Nvidia, Google will look like Amazon and Nvidia, et cetera, et cetera.

The most important thing that's happening now is that we're getting deep technical actors with huge balance sheets to be on the side of an open market structure. I think that will be the most important pushback, because that will allow the best idea to win in the marketplace.

I don't want to judge the EA agenda. It may be really good, but I don't want to be rug-pulled and end up in an oligopoly or duopoly market structure because of a hysteria that was man-made.

Yeah, I think it's so well put. A couple of weeks ago, I talked about how I thought Jensen was going to win open source, and I tweeted about this today. I believe that, starting today, Nvidia is the number 1 competitor for OpenAI and Anthropic—not just for tokens, but for enterprise compute.

I think now Jensen is the leading open-source provider in America, and I think he's going to take the world. He's going to catch up and blow past China and the Chinese models, I think, by the end of 2027.

That means you're going to have a player in Jensen and Nvidia that will enable America to win—the goal of your previous job in the government there. Sacks, you were very concerned: Are we going to win the race? I am more convinced than ever now that we can win the race, because Jensen represents open source globally.

He's going to be able to take a rack of compute, bring it into a company that's currently using tokens from OpenAI and Anthropic, and say, “Here's your rack, soup to nuts, complete vertical solution, and you have total sovereignty. I'm reducing your prices 80% or 90%, because I don't need to make money off the tokens. I made money off the hardware. I'm making money 10 different ways to Sunday.”

We would have won, except that now you have Governor Shapiro, who is the biggest proponent of data centers in his state, and Greg Abbott of our great state, Sacks, who are now on a jihad against data centers in order to win their midterm elections. That is what's truly scary.

They have now taken a stance after all of this, and we have horseshoe theory. Sacks, you have Steve Bannon, Shapiro, Abbott, and the lunatic socialist fringe left all in agreement: We need to stop AI, stop the data centers, and burn the windmills in the street.

It's a truly troubling moment, because we are poised, with Jensen's move in Hugging Face, to win open source, and we've already won frontier models. Your thoughts, Sacks?

Yeah, I mean, look, I think there's an interesting transition that's underway now on the political side. I think it's just starting to begin. Over the past few years, the political debate over AI has really been between accelerationists and decels, or doomers. Those were sort of the fundamental lines.

You're right that it went across different political-party boundaries, because on the anti-accelerationist side, you not only had anti-progress, anti-innovation progressives like Bernie Sanders, you also had Steve Bannon and people like that—conservatives who were fearful of new technology or the disruption that might come from it. So it did go across party lines.

That was the debate for the last few years. Some people describe me as an accelerationist in that debate. I just see myself as a realist, which is that I think the technology is going to happen. Someone's going to develop it, and I'd rather it be the United States as opposed to China.

In any event, that was the debate of the last few years. Now I think the debate is moving to open versus closed. I think it's going to be very clear that AI is just not going to be stopped.

All the stuff that Bernie Sanders is doing, or these anti-data-center groups—they're throwing sand in the gears. But the reality is that the technology is marching on, and the innovation is happening very quickly. Ultimately, the reason why it's not going to be stopped is because consumers are adopting it.

That's the thing people forget. Why are the data centers getting built? Why do people want to build them? It's because the demand is there from consumers and businesses. The adoption is pulling this whole thing forward, and that's why I don't think it's ultimately going to stop.

Now, given that reality, given that I think realism is going to prevail, I think the debate is going to be more towards, well, how do we manage this? What is the market outcome that we want? Do we want to have an open market where there are lots of decentralized models? There are sort of checks and balances. There's a balance of power in the market between different models. Or do we have a highly centralized solution where there are a couple of frontier companies?

We have a frontier model duopoly. They're closed.

And they can rug-pull you, Sacks, at any time. They can rug your tokens away from you.

And that duopoly has been aggressively lobbying for a new regulatory agency. Why? Because of regulatory capture. They can now control it. Again, the government officials are not going to know how to regulate something this dynamic, so they're going to turn to the leading labs and say, “Hey, how should we regulate it?” Of course, Anthropic is going to have all the answers on that. Dario has already said in Senate testimony that he thinks open models, if sufficiently advanced, are dangerous because they're not centrally managed and controlled.

4. All-In Summit: Scholarships available!

It's very clear where the battle lines are going to be drawn in the future. It's going to be about whether you want a centralized, closed market or you want a market that is decentralized and open. And yes, closed models will participate in that, but the question is whether we have the open piece. Again, I think this will go across party lines because, frankly, the Republicans have their statists. But I do think that as this debate evolves, most conservatives and most libertarians are going to see the threat to civil liberties, and I think they're going to come out in favor of open.

Yeah, I think that's correct. I was talking to Politico because they were reporting on the story. Trump is going to turn out to be right about this data center issue. The greatest thing a community can do is to get these data centers now, when the data center providers are willing to pay your local taxes for your schools, give incentives, pay people off in any number of ways, and move them wherever they don't need to tap into the grid, where they don't need water, and where they're not going to make noise.

Trump keeps telling people this is a mistake: block the data centers, take the money. I have to say, just watching politics up close and personal, Friedberg has made me incredibly cynical toward it. We were sitting here 6 months ago—maybe it was even 6 weeks ago—and Abbott and Shapiro were lauding their data center efforts. I think Chamath and I interviewed Shapiro when he came on the podcast. Friedberg, what do you think will happen post-midterms here?

I mean, I'm of the belief they're going to switch their position back to being pro-data centers because they're just following the wave here. How cynical is it that Shapiro and Abbott are both now anti-data center? I think politically you have to be. It's like polling at negative 80. It's bipartisan.

Yeah, but you know, it's like, why isn't Trump then? Trump's not following this. Trump's an independent critical thinker. Sacks—

He's got guts. I have told you guys before that Trump is ultimately the best friend of innovators in Washington. If anybody could make the decision about technology in D.C., I'd want it to be President Trump—not cabinet secretaries, not the White House staff, not members of Congress, not senators, and not think tanks. I would really want every decision, if he had the time, to be made by President Trump because his instincts on this are really good.

He wants America to win. He does not believe in unnecessary regulation. He doesn't have all these doomer anxieties. He listens to a broad cross-section of people. He talks to all the CEOs. He is very up to speed on all these issues, and I give him credit on this. I spoke about this at the G20 this past week.

There are great deals for local communities to be had. This is the point that I think needs to be made: the reason why local communities should embrace this is not because of some abstract concern about innovation or anything like that. They should do it when the deal is good for them. And there are good deals to be had.

We've seen that in Loudoun County, Virginia. The average taxpayer's property taxes went down by $6,000 because they're getting all this data center revenue. In Louisiana, there's a parish that was able to give teachers their first bonus in years because of data center revenue. So we're starting to see all these examples where, if you make the right deal—you, the local community, with the data center and the AI company—it will lower your electricity cost because they will stand up net-new power generation and pay for grid upgrades.

It does not use an excessive amount of water. That's total fake news. We know the systems recirculate, and these data centers can be put in places where they're not bothering people. So there is a good deal to be made there.

Ultimately, just one last point on this: President Trump supports local choice on this. He's made the argument for why it's a good deal for local communities to accept these data centers, but he's not forcing them to. I'm seeing all this fake news—Democrats saying that the administration is trying to shove these data centers down the throats of local communities. Not true.

In the policy document that I worked on, the executive order that I worked on in December, we expressly said that data centers remain a state and local decision. There were other areas where we thought we should have a national AI policy, but data centers were not one of them. So no one is forcing these local communities. What the president is saying is, don't believe this media disinformation campaign. It can be a good deal for you if you're willing to embrace it. I think we're making progress on this message over the last week.

One thing that I would say from my meeting in Washington on Monday is that there is a lot of foreign involvement in trying to propagate the negative reaction and sentiment toward data centers. I think it's becoming clear who's part of this cabal and how they are doing it.

There was a poll that showed that when you walk through the actual upside and go into some of these communities, the sentiment shifted by 30 or 40 points. It's still a little bit on its back foot, but it just shows that the side that has been organized on decelerating America has done a better job thus far than the side trying to accelerate America because it's been a little bit more fragmented.

We're able to increasingly see that the decel movement is organized in large part and funded by foreign actors. I think that as we put our fingers on that and expose it more, that's going to change as well. The reason why Abbott and Shapiro are reacting, I actually think, is that they're reacting to faulty data.

And why does the faulty data exist? The faulty data exists because we've had one side screaming from the mountaintops. And it's just now, as Sacks said, that when you go to these actual communities and actually measure people's reactions, it's much more positive because they're feeling it in their bank accounts.

Yes.

And those communities make a good deal. They just have to make a good deal. And by the way, Chamath, you're right about the CCP.

Yeah. Let me pull this up here. So the report you're talking about, Chamath, has been investigated. X got rid of a lot of bot accounts. And here's Axios: “China is secretly fueling America's data center rage.”

Roughly 200 accounts from suspected Chinese platforms have quietly tried to influence Americans to oppose AI data centers on social media. X said Thursday night that it's the same thing that happened during the elections, with both sides going back and forth.

It's incredibly easy to shape public opinion, Sacks, by just having a small number of social media accounts, pounding people, brigading them in their replies, and forming what people believe incorrectly is consensus. And the consensus is not that people don't like data centers. It's that a bunch of bots are telling influential people that people don't like data centers.

I know you're about to move on to other topics, but I'm just going to say something right now: please, there's going to be a cascading set of decisions and reactions, and people will anchor into points of view on this topic and related topics around AI which will look incredibly, incredibly dumb in 6 to 9 months and a year from now.

The reason it will look dumb is that, as the actual truth gets exposed, these folks will look incredibly foolish. That's my prediction.

Yeah. I mean, look, just on the influence op here, clearly there have been a lot of influence ops around data centers, and there's a lot of fake information about this. So there are a few different, I'd say, cadres here.

There's the CCP, and there's the DSA-type wing of the Democratic Party, which, frankly—I mean, both they and the CCP are communist parties who want to sabotage America, so it kind of makes sense. Then you've kind of got the doomer groups, and there's a writer named Nirit Weissblat who writes a blog called AI Panic who, a couple of years ago, traced that there were hundreds of these astroturf doomer groups that were basically funded by 3 EA billionaires in Silicon Valley.

It was Dustin Moskovitz, Jaan Tallinn, and Vitalik Buterin. Remember, Vitalik gave $600 million of dog coins to the Future of Life Institute. And then they've been spending hundreds of millions of dollars on these campaigns to make people afraid of AI, and they see data centers as a good target.

So you really can't underestimate how much of a difference a billion dollars of donations from these EA billionaires has made, or how much damage that's done in terms of the public perception of AI.

And to the point earlier, we are going to have probably 100 new EA billionaires as a result of the Anthropic IPO. As those people find their way into politics, it will be like George Soros times 100 or 1,000.

We're going to have 20 NGOs and 50 people working for them. It's going to be nuts. I mean, I told you guys I was getting brigaded. I created a bot stack that goes through my replies for me, and it just says, “Anybody under a year, anybody under 200 followers—just tell me what they're tweeting.” I look at these, and it's all the same thing. They're just brigading people constantly from fake accounts.

If you're under a year old, you shouldn't even be able to be in people's replies unless you have your real name or you're a blue-check mark or something. It's gotten out of control on these social networks. It's so easy.

5. Mamdani bans AI in NYC schools K-8

All right, let's keep going through the docket. We've got a lot to get to here. Freeberg, I need your comments and insights on Mamdani. This is an interesting one. He just put a 1-year ban on AI inside New York City's public schools, from kindergarten through 8th grade.

People have been saying he banned ChatGPT, banned AI. Let's slow down here and get into the details. New York City is the largest school district in the United States. The moratorium starts next week, and there are 600,000 students in New York City.

This covers all software that uses student-facing generative AI. High schools are exempt, and Mamdani is piloting an AI learning program with 50,000 high schoolers. That's a very specific thing that he's piloting that's different from these new guardrails.

He also gave recommendations for screen-time caps: 30 minutes for grades 3 through 5 and 45 minutes for grades 6 through 8. Mamdani's reasoning, quote, “I have yet to see a study showing that AI is beneficial for students in elementary and middle school.”

Obviously, Mamdani is the face of democratic socialism in America. One New York City DSA member, Julie Swope, said technology “should not be mediated through a for-profit company that is enriching tech oligarchs further.” I think he's talking about USA. Is that old? Google, Microsoft, and Anthropic declined to comment.

Freeberg, what are your thoughts on this? How far is this from how you think about how you would want your children to be raised if they were going to New York City's public school system, or whatever school system you choose to send them to?

There was an article published by Stanford in March of this year. Counter to Mamdani's statement, they went through all of the research that's been done on the use of AI in K–12 education. It's a report called “The Evidence Base on AI and K–12 Education: A 2026 Review.”

They identified several hundred papers that have been published—only 20 high-quality causal studies, is what they said. This was back in March, so many have been published since then. The conclusion of the meta-study, if you will, is highlighted at the bottom of the report: student performance often improves with access to AI tools, but once removed, results are mixed.

Tool design matters, so not all tools are equal. AI may meaningfully support educators. There's quite a bit of analysis that's been done, and it's continuing to mount and surface. There doesn't seem to be a strong evidence base to suggest that AI is detrimental to kids' education and development or to the educational process.

What I think we can all intuit about this is that there's certainly a negative push from teachers' unions against AI tools, under threat of automation taking away from their core skill and capacity, and because of the lack of skill that educators have in utilizing AI to educate students in a classroom-based setting.

The common-sense equivalent, also on the flip side, is that AI can enable personalized tutoring. The objective is that every individual learns at a different pace. Everyone learns differently—how you learn, how I learn: visually, orally, with certain imagery, versus being told stories, versus interactive, project-based work or building work or dictation.

There are a lot of different ways that people learn. All of those learning systems can be personalized to an individual, and ultimately that individual, as a result, should be able to progress more quickly and have greater confidence in their learning.

This is my common-sense belief, not rooted in some study: AI ultimately will lead to a much more personalized system of education for students. That is deeply disruptive and deeply threatening to the traditionalist systems of education in a classroom—everyone learns the same way, and everyone learns at the pace of the slowest student in the classroom.

Typically, educators have a lot of rules around what they're required to do and not required to do because of the contracts negotiated by their unions. So, there's a deep political orientation with Mamdani's decision here to limit access to technology.

I think it's unfortunate. I think it's really sad. I think the opportunity for AI to reinvent education, give kids a level up, take any child from any background or socioeconomic condition, and give them the opportunity to accelerate their capacity very quickly and very effectively is an incredibly powerful tool that should not be discarded because of a bunch of stuff that's been written about AI or concerns that we don't really know or understand.

By the way, that is always the shorthand for dismissing something that's incredibly going to change a system: “We don't yet know.” The “we don't yet know, we don't yet understand” narrative can always be used at any point because you could always say, “Well, there's more studies to be done. There's more analysis to be done.” There's no limit to that statement.

In this case, I've tried to highlight that there have been 800 studies, at least 20 of which Stanford has identified as being very credible in their analysis of the utilization of AI in K–12. So, I think there's a political bent here, but I think it's probably also part of this deep-rooted dysfunction that we've been talking a lot about: this dysfunctional fear of technological evolution.

No one understands it. No one—when I say no one, I mean the general population—feels this way. No one knows where we're going with this. We need to be careful. We need to be cautious. I think that's probably what's going to hold us back.

Now, why is that bad? Why is it bad to take time? Well, it's bad to take time when you have an opportunity to progress forward quickly and aggressively because there are going to be alternatives out there.

There are going to be private schools that are going to use AI, and I don't want to see kids of privilege and kids of wealth accelerate ahead because they're getting an AI education while public-school kids get left behind. I don't want to see kids in other cities, other states, and other countries accelerate ahead while kids in certain districts that have this anti-technology, socialist orientation get left behind.

So, I think there's a fundamental social dysfunction that we're dealing with across all of this massive technological upheaval that's underway. This is one manifestation of it. The root of it, the mechanism of it, is probably related to teachers' unions and the politics around this stuff right now.

Anti-AI is very popular, so being anti-AI gets him more points and more kudos. The unions love it. That's my summary of what I think is going on.

Great insight. Mamdani may not have lied here, Friedberg. He said he didn't see any studies. He just may not have looked. You actually looked.

Didn't click on that link.

I didn't. I just didn't click on the link.

Chamath, you are passionate about education. You've got kids in various age groups, so you've been through this and you're still going through it—kids learning and thinking about devices. We talked about this just the other week with the Meta settlement and how challenging it is for parents.

What's your take on this? Maybe you could steelman what resonates with you in terms of using AI, putting New York aside, for your own kids. What benefit is there to them not using AI, and where do you think it's actually great that they use AI? Take us into what you think as a parent, and put the politics aside for a second, if you will.

I think it's inconceivable to me that kids will not have an adaptive learning platform as their primary way of learning. I think this idea of grade 1, then grade 2, then grade 3 is, as Friedberg said, a lowest-common-denominator approach to education. I don't think it works.

I think it's more conceivable that in the future we have kids who have very different learning styles. Some are more visual, some are more auditory, and we have different learning paces. If you could have something that's much more tailored to each individual person's learning style, where you have guides that help you in moments and then the rest of the time you're teaching executive function, teamwork, problem-solving, sports, and all of these other things, that's a much better outcome for all kids.

So, that's the right answer. Right now, there are only charter-school options around that answer. The leading edge of that is Alpha School, which is our friend Joe Lonsdale. Big shout-out to him and his team. But I think that's the answer that's going to win.

Let me just give you another reflection, Jason. I think that this is an incredibly important moment. We should demarcate this. This is when New York becomes Mississippi while Mississippi becomes New York.

What do I mean? In all of these highly liberal bastions, they are embracing this far-left-wing ideology. I think it's going to come with these kinds of policies.

And I think what it will do is create a self-imposed form of modern segregation. It is going to completely push an entire generation of kids in these cities that are subject to these rules into poverty, and it will keep them there.

I think you're going to see this flipping, where all the states that embrace AI, embrace the data centers, embrace the tax cuts, and embrace the teacher bonuses are just going to end up ahead. You're going to have people who are in a position to move, move. Instead, what's going to be left over in many of these cities that embrace these things to the extreme are hollowed-out versions of their original selves.

That's going to reorder the economic balance inside the United States, the voting balance, and many of these philosophies. So I think it is very important. I think he should run the test. I apologize, and I feel deep empathy for the working men and women whose kids will be hurt by this.

But I think people can only learn by touching the stove and getting burned.

Okay, Sacks, maybe you could go into the political side of this, if there is one, and maybe give us your cynical take on what's going on here. Do you believe this is a reaction to the oligarchs? I joked previously that they're talking about you, but it does have a little bit of a tinge of democratic socialism, and it feels like maybe they want to sacrifice kids.

That's one interpretation that's happening online. I'm not saying it's mine, but people are saying, “Hey, they're trying to sacrifice kids because they don't want Sam Altman or Dario, or whichever service it is, Microsoft, to get a couple of shekels for software being used in the classroom.” Of course, they would all give the software to them for free, and the current versions of it are 80% free. So that makes no logical sense—that they're sticking it to the man and the big oligarchic tech industry. Your thoughts?

Well, look, the DSA has really come out against any innovation. They're kind of just on the side of anti-progress. They've fully embraced the Luddite side of the argument.

To the point that Friedberg and Chamath were making earlier, this is going to be really bad for the kids. We've been talking for 30 years about a digital divide in this country, where supposedly the kids of wealthy people would have access to the best tools and the kids of poor people would not have access to those tools, and they would fall even further behind.

Here you have a case where private schools are of course going to continue to use the best AI-powered tools, and now the public schools are not going to have access to those tools—not because they don't have the money, not because the technology is not available, but because the political leadership is trying to sabotage those kids. It's going to be bad for them.

I guess, if you were to describe a motive for it, I don't think Mamdani and his political group, his party, really care if they create more political dependence. In a sense, their whole movement consists of downwardly mobile progressives. If you're looking for a motive, maybe they're fine having these kids be downwardly mobile. I don't know. Maybe that benefits them politically.

They may actually not have kids at all, and they're making these rules.

People with kids are not part of the—

Right?

Yeah. Actually, let me bring one other element into this.

Yeah. So, look, I don't think we really need to get into motives. We can just say it's a bad policy and leave it at that. But let me bring in one other element of this, which is that it's not a full ban.

What they also said is that, for a maximum of 50,000 high school students, they're going to have a pilot program—these so-called modules in which they teach AI literacy. But listen to how they describe it. They say that all high school students will have access to twice-yearly AI literacy classes designed to help them think critically about AI before they begin to rely on it.

These modules will introduce students to what AI is and is not. This is their language as well as AI-related bias, risks, ethical considerations, and impacts on careers and future skills. It says that all use will take place under the direct supervision of a trained educator and will involve carefully vetted AI tools. The pilots were selected based on rigorous safety and data-privacy standards.

You hear a lot of this sort of doomer lingo. Those modules sound like they could have been written by some of these safety people. This language about AI bias and discrimination—I mean, honestly, those classes sound less like you're going to learn how to use the tools to get the best education and more like you're going to learn to be suspicious of the tools.

Maybe that's a small part of what you should develop, because I guess you should be suspicious of everything. But it doesn't really sound like those courses are going to benefit kids.

It sounds like they're going to put them through Doomer Academy, quite frankly.

Yeah.

Yeah. Write your own interpretation.

So think about the competitiveness of our kids versus Chinese kids.

Yeah.

China, we know, is incorporating AI into K–12 education. Their kids are going to learn how to take advantage of these tools.

Robotics for the Robot Olympics and internships there.

Yeah. Meanwhile, we either are banning AI in our schools, or we're putting the kids through some sort of doomer safety module that's going to teach them why they should resent and resist AI, as opposed to embrace it to their benefit.

I don't know. It just seems like this is not going to make our kids more competitive.

Let the experiment run, unfortunately. But let the experiment run. We need to measure it.

I have a bunch of thoughts here. I am a big advocate of Montessori, but I'm also a big advocate of adaptive learning and what these AI tutors are able to do for a motivated student whose family cannot afford a tutor. I just want to put 2 disparate points together here, and they're somewhat opposing.

Nick, if you could pull up the first one from the NIH here about the MIT study, and I'll send this one to you, Friedberg. You can look it up. It's very interesting. They took a group of 54 participants from last year and put them into 3 groups over 4 sessions, and this was over 4 months.

They were using ChatGPT for essay writing, not unlike the conversation we had about the Wall Street Journal editorial page last week. One group was using an LLM, one was using traditional web search, and the brain-only group wrote essays without using any tools.

The researchers used headsets to watch their brain activity. What happened was that the LLM users showed severe deficits in memory and essay ownership, with 83% unable to quote from the essays they had just written.

Basically, what they're saying here is that if you use these LLMs instead of being able to do it on your own, authentically, like the Montessori philosophy, your brain atrophies. You don't even remember what you wrote. The search-engine group was somewhere in between because they had to remember things and pull them together.

Your brain atrophies if you don't actually do the work, just like your body will atrophy if you don't work in the fields like we used to do. That's why we all go to gyms now.

Correlating that, super importantly, is the two-sigma problem. So, Bloom—and I'm sure you're aware of this research, Friedberg—this is super interesting, but it will not be shocking to anybody of means. My parents could never afford a tutor. I was a product of the public school system until 8th grade, and then I went to Xaverian High School, a private Catholic school, which was super cheap.

Basically, we've known since the '80s that if you are tutored one-on-one, you are 2 standard deviations better than students who are educated in a classroom. So, 2 things can be true at the same time. These students, to your point, Chamath, don't have access to a tutor. If they're going to the New York City public school system, their parents, in all likelihood, like mine, didn't have the ability to hire a tutor.

I'm assuming, Sacks, you came from a little bit of a middle-class background. I'm assuming you did have some tutors. Am I correct?

I didn't have tutors.

No tutors. Okay, so you had to just do it all.

Yeah. I didn't necessarily have a tutor. No.

No. You know what? I wish I had done an SAT workshop. I went to that test cold. I never did anything.

And you got 1,400.

Yeah.

You did 1,350? 1,450?

I got 1,420 my first time. I'm like, man, if I had actually prepped for that, I might have done better. I got 1,100.

I retook it and did better. But anyway, I didn't know any of this stuff.

Yeah. So—

Here's the tension. Most parents do not want their kids to do this work, have their brains atrophy, and not learn how to do it from a basic level. Putting aside the drill-and-kill conversation from last week, you want people to be able to learn how to read and write and do this stuff organically, authentically—whatever term you want to use—and then learn how to use AI to help them, augment them, and so on.

But I'm concerned that the motivated students who cannot use these tools properly also can't afford to be part of the 2-sigma group of parents who know very well that all you have to do is get your kid a tutor. And unlike self-motivated Sacks, they're going to easily get to 1400 on their SAT with a tutor because the thing is rote. It's very easy to get to 1200, 1300, or 1400 with a tutor.

So I'm super concerned about not having an AI tutor. I feel like there are two things that can be right at the same time with this, and I really don't like this experiment as currently constructed.

6. US-Venezuela oil deal

All right. Interestingly, President Trump has just closed a deal for Venezuela's oil back in January. I think everybody knows the U.S. military, in a perfectly executed operation, captured Nicolás Maduro. He was obviously Venezuela's socialist leader, who was very pro-China and pro-Russia, and obviously Venezuela is in our hemisphere.

Trump said the U.S. would control Venezuela's oil, which they have a lot of. It's not super clean, but they do have a lot of it. He made a deal with Maduro's replacement. President Rodríguez is now in charge.

Now, we have the details of this. North American Blue Energy Partners gets a 100-year concession over 17 oil fields. This includes 65 billion barrels of reserves that were once controlled by Russian, Chinese, and Maduro entities. The U.S. government controls 55% of the deal. The Pentagon gets 35% equity. The Pentagon gets a third of the equity. The State Department gets rights to buy 20% of output at cost.

Interesting. I guess that's for our reserves and our military. This was signed by Rubio and Hegseth with Venezuela's interim president. As I mentioned, Blue Energy Partners is pledging roughly $100 billion in new infrastructure. A lot of other companies, Chevron and others, are also going to get some deals. Trump said this deal will be, quote, “at no cost to the American taxpayer.”

Sacks, I'll throw you a fastball here, if you want it, or I can give you the softball. Which one do you want?

I like the fastball.

You like the fastball? Okay, here it comes. Coming in. Hope you're wearing a helmet.

Should we be seizing and taking over countries and seizing their oil fields? People feel like this might be nation-building, that we're putting in puppet dictators. Is this something America should be doing? Do we have the right to do this? This is the hard question people are asking. Does might equal right?

Well, I definitely don't think this is nation-building. We're not trying to stand up an army here like we did in Afghanistan. We're not trying to teach them about toxic masculinity. We're not trying to stand up a government or anything like that. We're not trying to teach them how to do American-style democracy. We're just making a business deal. This is a business deal.

It's amazing to me that all the people who are saying, “Well, this is exploitative of Venezuela, and this is a bad deal for the people of Venezuela”—have you looked at Venezuela's oil production under Maduro? It already wasn't very high, given the incredible reserves that Venezuela has. But under Maduro, production dropped by about two-thirds over the last decade.

They used to produce about 3 million barrels a day. Under Maduro, it went down to about 1 million barrels a day. This is a country with some of the largest oil reserves in the world. People debate exactly what should count. I think if you look at the total amount of reserves, it's something like 300 billion barrels of oil, which is around the same as Saudi Arabia, just a little bigger than Saudi Arabia.

Now, some people say, “Well, a lot of that is this extremely heavy crude, sort of like sludge, that's not usable.” I think that's the so-called Orinoco reserves, but that's only about 200 billion of the 300 billion. So they still have about 100 billion barrels of what you could say is still heavy crude, but those proven reserves aren't really debated, I think. So you're talking about a huge amount of oil here.

I guess just one last point on this is that that type of heavy crude is actually very complementary to our own. In the United States, through the magic of fracking, we've basically developed all of this oil and gas, but our oil is this light, sweet crude that's good for refining down to gasoline.

The heavy crude is what produces the distillates, and that's where diesel comes from. Our refineries in the Gulf were built for heavy crude, and it's very expensive to try and retrofit them to do light crude. So our refineries are actually perfectly set up to take in this heavy crude and turn it into diesel and other distillates.

I think it's really a win-win deal for the U.S. and for Venezuela. Again, Venezuela was doing horribly before. I think this has to improve their economy, and I think it's really good news for the U.S.

To your point, Sacks, over the past 5 and 10 years, the price per barrel of oil coming out of Venezuela has averaged $53 a barrel over the past 10 years and $58 a barrel over the past 5 years. So they're getting a pretty good deal. We're paying a little premium here to control it all.

They were selling at a discount to China. I think it was like a $20-a-barrel discount to basically sell.

Yeah. So I guess when you're dealing with authoritarian countries and you don't have a free market to put the stuff out there, you get a slightly discounted price, right? Somebody's got to skim something here or get a reason, and they have to ship it across the world.

Friedberg, your thoughts on this? Do you have any concerns about the nation-building, putting in what some people are saying is a puppet president? Is it in the best interest of the people of Venezuela? Are we doing a solid for Venezuela to create more stability by being the buyer of this, or is this just something we have to do? Do we have to control our hemisphere and become energy-independent, given the state of the war with Iran and the closure of the Strait of Hormuz?

The biggest motivating factor may be keeping this oil away from the sweetheart deal for Russia and China. If you look at the historical tie-up with Venezuela and the Maduro regime, it was a discounted oil trade system. There are all these stories about Maduro being the head of the snake of this illicit activity, and a lot of this was driven by Russian and Chinese interest in the energy reserves. They were providing support, and this actually provided capacity to fund activities by the Russians in Cuba.

So there's a lot of this state actor theory about what's the real motivation for getting an energy deal done—not just in terms of whether it's good for America, but whether it's bad for America's enemies. That may be a big motivation around this.

One of the more surprising outcomes over the last day is that you guys may remember I interviewed María Corina Machado. She was the individual who was running for president and ended up having what was supposedly an election that was not accurate. She was supposed to have won and didn't win.

Yesterday, she came out and said that this Venezuelan government does not have the authority to do this energy deal with the U.S., despite historically having been viewed as an American ally. One thing to take note of is that she was not given the presidency, and now she's saying this government is illegitimate. This government should not be allowed to make these deals. We, as the people of Venezuela, should feel angry and upset and left out about this whole thing.

So there seems to be this political turning. I don't really have a good sense, and I think it would be interesting to know what the people in Venezuela think about this deal. To Sacks's point, it seems like it should be a good economic deal relative to where they're coming from. Will that actually translate into improving the lives of the people in Venezuela, which ultimately is what matters here?

Assuming that the lives of the people in Venezuela are improving, rather than enriching a small number of people in Venezuela—which is what has happened historically in that country under the last regime—then there may be a long-term welcoming of this deal. But I do think there are 2 aspects to this. One is keeping America's enemies at bay so that these energy fields are not exploited by America's enemies, and the second is ensuring that the Venezuelan people realize the benefit of this transaction rather than it being a corrupt government.

Provocative, Friedberg. We have turned off Russia and China's access to Iran's oil with our war with Iran, now in its seventh month, and we took Venezuela away. Are these 2 actions incredibly provocative, and have they had any kind of downside to China? They seem to be all in on nuclear, solar, and other modalities. Are these 2 things super provocative, and are they going to cause us to get into an escalation with China?

The second question is a little bit independent, but I think China imports about 11 million barrels a day of crude oil.

Venezuela is only producing 1 million barrels a day. So if China is importing 11 million—well, they were importing 11 million before the war. That number has actually gone down a lot since the war started, and we can talk about why. But China gets most of its oil from the Middle East, the Persian Gulf, and Russia. Venezuela is a small number.

Look, I think Friedberg's right that securing those reserves in the Western Hemisphere is a big part of our motive. They are massive. It's, again, complementary to the type of oil we have. I think it's just a smart national-security move and a smart economic deal.

JCal, to your point, does this get us more invested in Venezuela? I think there is a little bit of a risk there, because if we're going to invest billions of dollars in upgrading their infrastructure so that we can produce 3 million barrels a day instead of 1 million, we need to know that that investment is protected. We can't have some future government come along and just nationalize our American oil companies over there, right? We need to know that these business deals aren't going to get rug-pulled and that they won't just seize our upgrades.

They could seize the upgrades.

Yeah.

This business deal is supposed to last for 25 years. So it does give us a vested interest in making sure that the Venezuelan government at least respects our property rights. But I'm glad that President Trump didn't put Machado in charge, because how's she going to stay in power? What guns are going to basically protect her? You know, it takes men with guns to basically stay in power in these countries.

You need a strong man or woman.

I'm not saying that it can't be democratic. I'm just saying that if you're not installed in the power structure there and have backers in the military, things like that, then it's very hard to stay in power. And that's the kind of thing where we could have to send American GIs in to protect her. And we don't want that, right?

So the nice thing about the Venezuela situation is that Delcy Rodríguez and her brother are basically running the country. She was the number 2. Now she's running it. They appear to have enough support from the establishment there. And so that's what the American interest is, I think: just to have a stable government there that will basically conduct economic relations with us.

Yeah.

Yeah. Chamath, any thoughts here on what's happening in Venezuela?

Close to it? I don't really know.

Don't know. Okay, that's fair enough. Fair enough.

All right, there it is everybody. Episode 288. Somehow this band has kept it together for 288 albums and tracks. We're going to make it to 300, boys.

I think we're in a good rhythm. The show feels easier to do now than it did for a long time.

Yeah. I wonder why. I wonder why.

The AI has been a very good thing for the pod.

We're partying like it's 1999 on the pod. Let's go to Cabo. Let's go. Love you guys. Love you, besties. Bye-bye. Bye. We need to get merch. I'm going all in.