AI 是泡沫吗？专家与 Dave、Salim、AWG 辩论 AI 的未来 | EP #190

Everybody, welcome to Moonshots. Another episode of WTF Just Happened in Tech here with my Moonshot mates, Salim, Dave Blundin, and Alex Wissner-Gross. Gentlemen, as I like to say to our listeners, get ready to add 20 IQ points this morning.

There is a lot going on in the world, as usual.

Oh my God, a crazy amount.

We were in California for 1 week at OpenAI headquarters and came back, and about 50 things have happened that we need to talk about right away.

It’s unbelievable.

Literally, the team and I—and all of us—spent about 20 hours getting the slide deck ready, trying to figure out what to put in it and what not to put in it. Last night, we said, “We’ve got to add this. We’ve got to add that.”

It’s a goddamn full-time job at this point. It was ridiculous.

I tell you, we are full-bore in self-improvement and singularity mode. The rate at which things are popping now, it’s not going to stop either.

I would argue that in AI, we’ve actually crossed the singularity. The pace of change is faster than we can process it.

You’re talking about like-minded souls here. We’re all going to agree on that topic. It’s funny—I hung out with the family over the weekend and saw my nephews, and they’re just not aware yet. They will be very soon, but it’s just crazy. They’re watching the podcast, so they’re keeping up. Let’s accelerate them today.

Yeah, and I appreciate that. The number of people who have reached out and said, “Oh my God, I love WTF. I love Moonshots,” has been really heartwarming. Alex, they love seeing you, too, so welcome back as our fourth here.

Very kind. Yeah, super excited to be here.

There are a bunch of topics today that are very technical, and hopefully you’ll be the one guy on the planet who can explain them. I’m very much looking forward to that.

Hey, Salim, you look like you’re at a space station someplace. Where are you?

I’m at Newark Airport, trying to do a podcast, so this is not conducive from an environmental perspective. I found a corner of the airport. Hopefully it’s okay.

Okay. Well, hopefully no one shows up. If they do, just yell “bomb” or something.

If I get dragged away, you’ll know why.

All right. Let’s jump in. Like you said, a lot’s going on. The very first subject is the AI war. AI is always accelerating everything. We’ll cover that, robotics, and BCI together today. We’ll cover a number of different subjects, but it’s a heavy AI digest, and we’re going to begin with the latest on GPT-5.

Dave, like you said, we just recorded a podcast with Kevin Weil, the chief product officer. Hopefully people are enjoying that podcast. It was a really beautiful setting, and then literally a week after we were there, everything continues popping. It’s great.

The podcast, by the way, was absolutely a must-watch. It was an awesome, awesome episode you guys did.

Yeah. Thanks.

We missed you there. You guys had really good questions, even though I wasn’t there. I was in the hot tub with my son this weekend. He’s at Wayfair right now, and I said, “You’ve got to get into that building. Go to San Francisco. You have to get into that OpenAI building.” History is happening in real time at light speed. Just find a way to navigate in there and meet a couple of people. The energy in that building is like nothing else on the planet.

Yeah. Well, I’m sure it’s that way at the Gemini team at Google and at xAI.

Yep.

Here’s an IQ curve. This is the Mensa Norway test, and by definition, it’s a bell curve. We’ve got the average human IQ at 100. I’ve been watching this because, of all the metrics, Alex, that we speak about, I’ve been watching IQ simply because it’s a humanizing effort. Here we see GPT-5 Pro come out at an IQ of circa 148, which is pretty damn good. I’m not sure if it’s you, Alex, but it’s pretty damn good. I’m hoping, Alex, we can trigger you on a rant about how we crossed the Turing test and nobody noticed, and now we’re crossing barriers that would have been unthinkable 2 years ago, and people are like, “Well, you know…”

My mission is accomplished. Dave, you’re telling me ahead of time what I’m going to say. This is great.

I think this is yet another sign that benchmarks based on the average human population are saturating. We’re running out of sigma on a bell curve here to benchmark GPT-5 Pro, arguably the strongest generally available model at the moment. We need new benchmarks—harder benchmarks that arguably look less like the average distribution of the human population and start to look more like specialist knowledge that isn’t generally accessible to a test administered to the broad human population.

We talked to Kevin Weil about the idea of an abundant set of benchmarks, where we’re looking for AI to solve the biggest problems in the world, and he liked that. So maybe we’ll see some benchmarks. Anyway, it was GPT-o3 at like 120 or 136, and we bumped up another 10 IQ points here. It’s going to be interesting when we start seeing IQ points beyond 200 and they’re therefore really immeasurable and don’t make any sense anymore.

At some point, one might expect tests like this to start to factor in AIs. Right now, this is based on the distribution of unaided human individual meat-body brain capacity. What happens when AIs start to merge with humans and the curve itself gets dragged upward? We’re going to talk about that for sure. That’s a fun one.

I have my usual rant against this, which I won’t get into now, but the neocortex is about the size of a dinner napkin, right? What happens when AI makes it the size of a tablecloth or a football field? What do we do then?

We listen carefully.

Here’s our next article: “AI models hit consumer hardware in 12 months.” Now, using a single top-of-the-line gaming GPU, like NVIDIA’s RTX 5090, which is about $2,500, anyone can locally run models matching the absolute frontier LLM performance within the next 6 to 12 months. Alex, thoughts here?

I think there are 2 stories here. The superficial, clichéd story is that this is about consumer privacy and empowering individuals with personal superintelligence, enabling individuals to have conversations with chatbots without needing to reach out to a server. I think that superficial story completely ignores the actual story here, which is that these frontier models are starting to incorporate new modalities, actions in the physical world, and video modalities.

The net upshot of all of this is that we’re about to get general-purpose humanoid robots that are running foundation models like ChatGPT or its now numerous frontier competitors, but they’re going to be running locally at ultra-low latency. When these curves cross, if they cross—and even if they don’t cross, if they come close enough together—this is going to give us GPUs embedded in general-purpose robots, with ultra-low latency, performing general-world, human-complete, AI-complete tasks.

Yeah, and that’s a very, very big deal for the consumer experience: talking to your car, talking to your personal robot, and having it say intelligent things back to you. I think for industrial use, everyone’s going to want the best of the best. They want to go up to that next curve. When you’re writing code or trying to design a rocket, you need the best of the best.

But when you’re trying to interact with day-to-day life—the other roughly 98% of use—you’re going to have a superintelligence locally that’s more than good enough to know exactly what you want: “Clean my house,” “Drive my car,” and all of those things. That’s imminent. That technology is here right now.

There’s another part that a lot of companies don’t want to acknowledge: they don’t want to give their sensitive data or interactions over to OpenAI, Microsoft, or Google. The ability to run all of this locally and have that capability in your phone, wherever you want it, I think is super important.

This is absolutely critical because even with MCP, your queries still go up there. If you’re a law firm, an accounting firm, or a government—any kind of sensitive government entity—you don’t want your queries being uploaded into the model either.

And so this is a huge deal. I think as we embed it, you'll have a lawn mower with this stuff embedded in it. It'll have a little LLM checking the weather and when it's going to rain, and it'll have that much more intelligence in it.

I just reread The Hitchhiker's Guide to the Galaxy with my son, Jet, and I enjoyed Marvin the Depressed Robot. I can imagine having a lawn mower with an attitude: “No, I don't want to cut the grass today. I'm tired. I did that last week. I want to do something different this week.”

He was so ahead of his time, Douglas Adams. Just unbelievable. I'm going to recount my favorite quote from him, where he said, “Anything in the world when you're born, we call it normal. Anything that's invented when you're young, that's called a career. And anything invented after you're 35 years old is just bad for the world. It's just bad.” Dave, you were saying.

Anyone who missed that 1X Robotics podcast, go back and watch at least the first 10 minutes of it, where Peter's interacting with the robots. They're a little clunky when they're moving, and that'll get fixed very quickly, but they are perfectly vocal when they're talking to you. It's just unbelievable. Everything you say, it understands perfectly, and then its reactions are perfect as well. So that part of the interface is already—

It was interesting when Bernt Øivind Børnich, the CEO of 1X, said, “We need to have the compute.” You said, “Why do you have the compute in the head? Why don't you have it in the cloud?” And he said, “Well, because we can't afford the delay time—the time for the electrons to get back and forth from eyes to brain, from brain to actuators. Crazy.”

That's it exactly. Maybe to tie a bow on what I was saying earlier, I think this is actually about latency in the long term. Privacy considerations—yes, sure, to first order—but the reality is, I think exactly as Peter was gesturing at, ultimately you want new knowledge that isn't already pretrained into the model. That requires reaching out to the world, and then privacy gets lost.

Let's move on. This is a related subject in that things are continuing to move. Again, our mission here is to give you a sense of how fast this is going and how this is progressing, and there have been no barriers, no ceilings that have been witnessed.

This is an article labeled “AI scaling laws have been shattered.” This is a 32-billion-parameter model that broke the Pareto frontier for AIME 2024 and AIME 2025. Alex, what are we seeing here?

There are so-called overhangs everywhere. An overhang, in general—a term of art in the research space in AI—is this notion that there are capabilities that are latent, just waiting to be unlocked, literally waiting to burst out if only we know where to look for them.

There arguably was a compute overhang when large language models—and even before LLMs, some of the earliest machine learning advances—were around, because we had GPUs lying around from video games, with all this compute waiting to be unlocked for this new purpose. Similarly, this paper here, which announced a new capability that was labeled and branded as data-efficient distillation, is pointing to the notion that there's arguably a new class of overhangs just waiting to burst out and unlock new performance at relatively low cost.

This is the idea of distillation. Distillation means taking a larger, so-called teacher model and using the teacher, not unlike human education, to train a smaller student model with the best of the teacher's knowledge. The core idea from the paper behind this chart is that with a properly structured curriculum, a proper data set, a teacher model explaining step by step the teacher's knowledge, and a number of other innovations, it's possible to take a relatively smaller, cost-efficient student and have the student demonstrate an enormous jump in capabilities.

I think innovations in distillation and the organization of training data sets are yet another overhang that's just waiting to yield 10x, 100x improvements in model performance.

One of the things that we always see in these charts is logarithmic axes, and you say, “What's the big deal?” That red star is not that far to the left—the x-axis is a log scale. So the red star is 1/100th of the training corpus of the purple star to its right.

The point here is that you can have equivalent knowledge with 1% of the compute in the training process—a 100x difference. The implications for startups trying to build foundation models specific to use cases are unbelievable.

You tend to get intimidated by OpenAI and Google having a billion-dollar-plus training budget, but if you can build equivalent capability with 1% of the data and then specialize it with data that's specialized in reading X-rays or designing parts for rocket ships, then you can actually build a specific model as intelligent as any other intelligence in the world within a reasonable budget.

Awesome. This is one of my favorite articles of our day here, our conversation: “GPT-5 can predict the future.” So the concept here is: Can these systems actually predict economic performance of complex systems or human social and societal performance—where things are going? We're going to find out, but these rankings on a Brier score are pretty impressive. Again, Alex, how much credence should we give this? Do you really think we're going to see AI models predicting the S&P 500 or Olympic winners this next cycle?

What's wonderful about predicting financial indices like the S&P 500 is that they immediately get priced into the market. The moment that there's an amazing crystal ball for predicting market performance, every financial institution, every quant fund will race, to the extent they haven't already—probably have already—to incorporate these LLMs and foundation models to trade better.

So in some sense, I think one can separate predicting financial markets, where, as the meme goes, “Don't worry, it's already priced in,” versus the rest of the universe, where, quite frankly, I think this starts to look like Isaac Asimov's psychohistory for forward prediction.

Something I hear almost no one else talking about is: What about retroprediction? Can we predict the past, or retrodict the past? So much of our past is a black box. If we can predict the future, can we do a really amazing job of retrodicting what came before us to very high resolution? I think—

I love that. So you mean we have data points of ancient Greece and ancient Rome. Can we fill things in? Is that what you're speaking about?

Exactly. To ultra-high fidelity. Some have called this, aspirationally, quantum archaeology. Could we retrodict the past light cone to quantum-level fidelity? And I think, ironically, predicting the past—retrodicting the past—might be even more exciting than predicting the future.

I love that, Alex. That's amazing. You and I have had conversations about going out thousands of years into the future, out to the light cone, and looking back at Earth and being able to see what happened if we had the technology to do that. I think this is an incredibly exciting idea for AI. Salim, what are you thinking?

It's fascinating because history is always written by the winners, right? The narratives and the types of publications around history were always coming from one side, and this gives us an opportunity to balance the playing field.

In my office, you may notice, when I'm there, the stack of books behind me called The Story of Civilization by Will and Ariel Durant. They spent their entire lives trying to document objectively what actually happened, rather than just the Romans' account after they conquered something. So now we can actually go and really fill it in. It would be amazing to see how you would rewrite that.

You could almost create a Wikipedia of what actually happened and let that be a referenceable model in itself.

I'd go even further and speculate—and maybe this is what mature civilizations do. They attain a certain level of superintelligence, and then some fraction of their compute gets allocated to navel-gazing and figuring out where they came from. Dave, what do you think is going to be interesting to predict in the future if we really get this right?

I was going to ask Alex that question because this benchmark is another one that looks like it's getting saturated in a hurry, and the concept—the headline there—isn't exactly eye-catching.

You know, we used the mean-squared error like everyone does, but what are we going to do next to predict the future in a benchmark way? There are so many options, but what do you think is coming?

I think at some point, predicting the future starts to become indistinguishable from innovating. You could ask, “What's the next major scientific invention next year?” To predict that accurately, you actually have to make the scientific discovery or the invention itself. I think that's, as we've discussed previously, the thing that happens the day after superintelligence. We start to get this flood of scientific, mathematical, and engineering discoveries.

The best way to predict the future is to create it yourself. Yeah.

Exactly. That's what it's doing.

Can you lay down a specific challenge? Erik Brynjolfsson has a whole new class coming in soon. If you give them a challenge, they'll rise to it. What would you measurably want to predict that's just fun and cool?

How about the details of the next 20 Nobel Prize-winning discoveries?

Super cool. All right. I mean, honestly, what happens then is: where do you invest your money? If you have the ability to predict, given that you can distribute your capital across the board, is there a higher-likelihood ROI on one specific technology? What we're seeing is digital superintelligence, so for hedge funds, do you put your capital all there?

Yeah, for hedge funds. I think if you can boil it down into predicting something that's happening in near real time, so people can follow it play by play, here's what the AI says is going to happen next, either in a sport or in a news event, and then they can track it. People love Polymarket, and they love Kalshi. If you can say, “Here's the AI benchmark, and here's the resolution happening almost instantly,” people get super engaged with that.

You can do that right now with fusion, because there are all these breakthrough questions: Can we hold the magnetic field for a certain amount of time? Once you can hold it for a certain amount of time, it means we can actually extract the energy out of it. That's huge. There's a sequence of known steps there that you could probably lay on a timeline and track in real time to see what happens.

Amazing. I'm taking this to Vegas with me, that's for sure. All right. GPT-5 Pro develops new mathematics. I think one of the things that I really want to track on this WTF podcast every week is the breakthroughs in math, physics, biology, chemistry, and materials science, because that's really where the juice is going to be. This is what's going to help us continue on this accelerating curve. Researchers entered an unsolved math problem into GPT-5 Pro for a convex optimization paper. The model produced a new proof, improving on the paper's attempt. GPT-5 Pro has had similar breakthroughs in physics and other scientific domains. Alex, you've been talking about this forever.

Yeah. And I want to flash the meme: It's here. It's happening. I think we're just at the very leading edge now of AI starting to bulk-solve math, science, and engineering. Right now, it's a trickle. It's an interesting, newsworthy moment when a weak improvement, arguably, over an existing optimization theorem was proven independently by GPT-5 Pro. It's remarkable, just this one little proof, but I think the trickle is going to turn into a tidal wave over the next year or so, possibly by the end of this year. So I think what this is going to turn into is basically bulk proofs of math, bulk discoveries in science, and bulk inventions in engineering all happening at once, which, right now, culturally, we have no precedent for.

Alex, I really want to get your take. I know for a fact that we're in full-bore self-improvement, right out of the Leopold Aschenbrenner paper, but a lot of people are in denial. Having worked hands-on in neural network research, writing the code for 7 years of my life, I can tell you that what's happening in this proof is exactly the kind of thing you do when you're researching neural networks. If you can do this, you can self-improve. I'd love you to comment on that, just to reinforce it.

My mental model is this: in computer science, you have this notion, when you're trying to make a program faster, of looking for the innermost loop. Usually, there are loops inside loops inside loops, and you're looking for the core engine of a computer program that's the most time-sensitive, most critical-path part that you want to optimize with accelerative technology—with the singularity, if you like that formulation. Arguably, the innermost loop looks like optimization. If tomorrow we can use AI to discover a better optimizer that offers orders-of-magnitude improvement, there's almost no other juice that's worth the squeeze—mixed metaphors—other than developing better optimizers and developing optimizers that are better at developing optimizers. That certainly smells like the innermost loop of our civilization right now.

Exactly. Exactly. I'm so glad you said it. We swagged the software-only rapid improvement at somewhere between 100× and 10,000×, and that was in a slide a few podcasts ago. But now you saw, in the slide we had a couple of minutes ago, 100× just in the data selection, in the choice of which data to select. So there's 100× on just one of those. I think we had about 8 dimensions in that slide of improvement, and they're all multiplicative when you put them together.

But if you have 100× in just that one, then our estimate was, if anything, on the lower bound. The implications of that are just mind-blowingly big. A lot of the deniers are saying, “Well, look, as we throw more compute at this, we're getting diminishing returns on this curve over here. Isn't this all going to slow down?” They're right on that one dimension, but the acceleration in these other dimensions is so much bigger than that slowdown. That's why people are going to underreact. I want to connect back to the previous discussion around history here, because imagine you take all of this capability now and apply it to all of the hundreds of thousands of experiments. Somebody did an experiment with 1,000 lab mice, giving them something, with a control group or whatever, and they're looking for one specific pattern. Now you have an AI that can look for all sorts of other patterns that a human being couldn't possibly see. I think we'll see unbelievable breakthroughs just from analyzing and doing better analysis of the thousands and millions of experiments that have already been done. That will be really incredible.

See, I would say that's overhang everywhere, including an overhang of previous scientific discoveries that are just waiting to be reanalyzed and reinterpreted.

Amazing. Amazing. All right. Let's watch a quick video here from Sam Altman about the Indian market for GPT-5. Again, I labeled this slide “The land grab.” I want to talk about the idea that we have these companies going out to deliver capacity to nations at a time. We've seen this in the UAE. We've seen this in Saudi Arabia. We've seen this in other places.

India is now our second-largest market in the world. It may become our largest. We've taken a lot of feedback from users in India about what they'd like from us: better support for languages, more affordable access, much more. We've been able to put that into this model and upgrades to ChatGPT. So we're committed to continuing to work on that.

India has 1.41 billion people. The vast majority—80–90%—are in severe poverty, half of those in squalor. It's a nation that needs AI more than anybody for health and education, and OpenAI wants to go there and give it to them. I'm going to link this article with the next one, which is “OpenAI in talks to provide ChatGPT Plus to the whole of the UK.” This article on its own isn't critical, but here we have these companies going in and saying, “Hey, let's give your population, your schoolkids, your factory workers—everybody—access to our model.” I do think it's sort of a land grab. What do you guys think?

I cannot wait to hear your thoughts on this, guys. Something is going on beyond just the cover story here. I know it for a fact, because when we were at OpenAI—not this trip last week, but the prior one, about 5 weeks ago—

Yeah.

—I said, “Why don't you guys open an office in Boston? We have 10 times more computer scientists in Boston than you have here in Silicon Valley. Incredible talent pool.” And they said, “Well, we're not going to do that because AI—you know, strong AI—is imminent, and this workforce is going to be all AIs.” But then they go a couple of weeks later and open this huge new office in New Delhi. Like, okay, you skipped right over Boston and New York and went right to New Delhi. That's not a coincidence. If you look at the demographics of India, it's the biggest population in the world, just crossing China right now, but the age in that 20-to-35 sweet spot is much bigger than in any other country in the world. Also, Mercor is now at a $10 billion valuation—the Brendan Foody story, which we can talk about if we have time—but Mercor is almost entirely operating in India now in terms of recruiting talent for the big AI companies. Something beyond just “it's the biggest market in the world” is definitely part of this plan.

We were talking about overhangs, right? The intellectual overhang in India is unbelievable. I don't know if you know the story of the mathematician Ramanujan. This was an obscure accountant in India 100 years ago, and he was sent to Cambridge, where he encountered a lot of racism.

He came back, died in obscurity, and then his widow handed in all his mathematical notes after he died. They found that there were 7 problems in mathematics that had never been solved for 1,000 years, and he had solved 5 of them. They have teams of PhD students reverse-engineering the notes now, asking, "How the hell did he do this?" This is an epidemic across India.

I think the bigger issue here is the infrastructure, energy, bandwidth, and so on that need to be solved first, because you're hitting people at the—as you mentioned, Peter—a large number of Indians are below the poverty line. This will have a double effect of allowing them to get out of that if you can get them the compute and infrastructure to scaffold themselves out of there. The potential is unbelievable.

Yeah, I tend to agree broadly that there are several—maybe 2 or 3—feedstocks to what we perhaps think of as global abundance. Abundant intelligence, or abundant superintelligence, is arguably one of the most important inputs. To Salim's point, abundant energy is another one of those feedstocks. If the world is just drowning in intelligence and energy, maybe material scarcity follows or is resolved automatically with energy and intelligence post-scarcity.

I think everything else—all of these global abundance challenges that we speak of—is downstream of those inputs, those feedstocks, and can be resolved and mitigated much more easily. That's one thought. The other thought, regarding India specifically, is that this starts to look like a prime example, if it were to come to fruition, of what one might call universal basic compute, or UBC.

UBC may be a special case of a larger class of approaches: universal basic services, the supply-side dual of universal basic income. The future looks very interesting if every citizen of a country is automatically supplied with a basic level of compute.

I also think this goes back to your question, Dave: Is this an economic play? If you can go in and get your software in as the basis for a billion people on the planet who are going to use your software to create more income for themselves and get a better life, and then be able to pay for your software, isn't this just an ability for them to—I’m trying to find a good analogy without going to drugs and giving a school kid a taste of a drug just to make sure that they start to use it.

This will become addictive to entrepreneurs, educators, healthcare workers, and government workers over the next few years. The question is, if you start using OpenAI, ChatGPT-5 and 6, and so on, would you switch, or would this become the baseline for a billion people in India?

Let's go to our next story here. This is just part of OpenAI's mission. Dave, you and I spoke about this when we were up at OpenAI headquarters last week. This is OpenAI's global data center dominance: opening up 2 large compute centers, 1 in Texas—Texas Stargate—with up to 5 gigawatts of capacity. Again, note that we're measuring the data centers in terms of power, not numbers of GPUs.

Then there's the Norway megacenter, 290 megawatts, in this case with 100,000 GPUs powered by hydropower. I was in Brazil talking—Brazil is a very energy-rich country—and I was saying, "Look at what's happening with OpenAI in Norway. They're going there for the hydropower. If you want data centers down here, make sure you get access to power and make it available." Power is sort of the pheromone that attracts data centers. Dave, what are your thoughts here?

This came up in a big way when we were at OpenAI last week, because I was asking Kevin Weil, "Is there a vulnerability for OpenAI in this area?" The big competitors have massive data centers. Google has massive data centers from years of GCP, and Microsoft has huge data centers. Kevin's answer was, "Yeah, well, Stargate. Stargate will be online."

These are the biggest data centers, the biggest investments humanity has ever made. OpenAI is starting from not having any data centers at all. He also confirmed that they're doing their custom chips. I don't know if that was public information—I guess it is now—but they're doing custom chips as well. I think all the horses in the race are now running in parallel, with huge infrastructure build-outs, custom chips, and now the new thing is AI designing the chips.

Yeah. The data centers, and soon the energy supply to the data centers, and soon predicting which politicians are going to support the data centers. It's all AI all the way down.

I want to play a short clip from CNBC here. This is with Sarah Friar, who's the CFO at OpenAI. This is going to wrap up our OpenAI-only segment. We'll go to the rest of the AI world in a moment, but here it is: OpenAI hits $1 billion per month in revenue, and the CFO warns of huge compute demand. There was a buzz about some remarks that Sam made about whether there's an AI bubble. We'll talk about that a little bit.

Developer outcome was actually great. I think our numbers were up something like 50% week over week on the number of tokens and so on being used. What we see is that tokens, in particular for agentic behavior and so on, almost doubled. Reasoning—which is what I get really excited about, because that's a place where I think we've really extended our lead—was up 8× in terms of usage of the reasoning components of the model.

Okay, so tell me about this. Sam has a dinner, I believe, out in San Francisco.

He does.

And he says at the dinner—this is the famous dinner in the last week—there could be an AI bubble that's taking place. Do you believe there's an AI bubble? I say that in the context that there's apparently a secondary sale of some of your private stock, with some of your employees trying to sell at a $500 billion valuation.

Amazing. So, Dave, is there an AI bubble?

There's definitely not a bubble. Two things: Sam, first of all, is now in full-bore downplay mode because he doesn't need to hype it anymore. He's exactly where he needs to be. He's in full-bore downplay mode. We've seen that before.

There are plenty of bad investments out there, all kinds of charlatans running around raising capital, and those companies will fail. Then people will say, "See, I told you it was a bubble." But that's not true. The tailwind is like nothing we've ever seen, and everybody is now—whether they know what they're doing or not—jumping on the ship.

All the business school people are coming out of the woodwork and getting involved. There are going to be some bad investments, and then people will say, "See, I was right. It was a bubble." Absolutely not a bubble. It's the biggest ship in human history.

And the worst thing you can do is not get on board and ignore it. That's the worst move you can make.

For me, this smacks of trying to manage your market cap. You've got employees trying to sell their stock on the secondary market, and you're like, "Oh my God, I'm trying to raise money out here separately. This is a total disaster. I have to do something. I have to say something." This is what it looks like to me.

Those numbers that Sarah was quoting were the jumps in users and compute right after the GPT-5 announcement, because there was a lot of conversation. Just to be clear, this was the biggest announcement on the planet, hyped not necessarily by OpenAI, but by the world. OpenAI has got GPT-5. The Leopold Aschenbrenner paper was like, "We're going to have recursive self-improvement when we get there."

So, hard takeoff. I think everybody was expecting AGI, and we got a simpler model with lower costs—not what was expected—but the world responded by doubling and redoubling its use of OpenAI. Alex, what do you think of the GPT-5 model? You've been playing with it. Is it everything it's been cracked up to be?

I've been very impressed. Prior to GPT-5 Thinking and GPT-5 Pro, o3 and o3 Pro were 2 of my favorite models, and GPT-5 Thinking especially is increasingly my go-to model for most tasks. It feels like a credible improvement.

I also think, more broadly, on this point of trillions of dollars being spent on data centers, as long as revenue continues to grow spectacularly, the party can continue for capital expenditures on data centers.

Yeah. Sorry, there's just something else here that's really important, Alex. In our last podcast, you talked about the fact that GPT-5 uplifts 700 million people into free models, right? That is a massive jump, and we're going to start to see the beginnings of that over the next few weeks and months.

Totally. There is a recursive element to it, in the sense that if 700 million people, many of whom are now using reasoning for the first time, are using reasoning and then using that to increase their productivity, their intellectual output, and their economic output, that starts to recurse back through the system and feed back more available capital—more real capital—into the system to build more data centers and empower more people.

And it's a wonderful positive feedback loop.

Yeah, I really love Alex's thoughts on this downplaying of expectations, because we now have infinite appetite for compute, which has never existed in the world before. If I'm running spreadsheets and I have a billion computers, who cares? It's still a spreadsheet. But now, all of a sudden, it's completely inverted. We have infinite appetite, and Sam has to be very careful about what he can promise to the world because it's all completely constrained at the compute level.

So when you demonstrate a V3 capacity, people get very excited about it, but then you realize, "Oh, wait, I have way too many users, and I can't actually deliver it." So that's, I think, what's driving the—well, hold on, guys. We're where we want to be. We're capable of doing a lot more than we wanted to show on GPT-5 launch day, but if we show it, then people will want it, and we just can't deliver it until Stargate's online. Even then, it'll be constrained.

So, Dave, to your point, I think what we're starting to see emerging, albeit in latent form at the moment, is a microeconomics of productivity per token. If we assume, just naively, that every token is equally expensive to deliver or to generate, some tokens are much more economically productive than others.

One could imagine, per token, maybe some sort of spreadsheet agent that's very productive and unlocks huge productivity, while video generation is relatively less productive per token. So I think we're going to start to see a new microeconomics of token-level productivity emerge. Fascinating.

It'll be managed by an AI, of course. Of course.

All right, let's go to the AI wars for the rest of the field. And again, just to be clear, we've been speaking about OpenAI and GPT-5, but we're about to see Gemini 3 coming online and then Grok 5 coming online, and it's just literally leveling up week on week on week.

In the rest of the field, Claude Sonnet 4 now supports 1 million tokens of context. 1 million tokens is a good amount. I remember in the early days being able to put in only a few pages at a time. 1 million tokens is about 750,000 words—3,000 to 4,000 pages of text.

I asked for an analogy, and GPT-5 said, "Hey, it's the entire Harry Potter series." We're seeing Grok probably somewhere in the 500,000 to 600,000 tokens right now. GPT-5 is around 250,000 tokens. It was a nice step up. Any thoughts on this, Dave?

Huge. Everybody thinks, "I don't need that. What am I going to do with 1 million tokens?" Then what happens is the AI is so productive for you—if you write code or text, it's so productive—that you end up with a massive amount of stuff you've created very quickly. It will forget what you did the day before because you've moved so far in a day.

Expanding that context field allows it to remember a lot more of what you're already working on. On the first day, you don't care. By day 3, you care tremendously. I've written more code in the last 2 weeks than in the prior 40 years of my life.

Wow.

And it's functional. It's incredible. It's working. It's self-documenting. But now my hard drive—I have literally almost a terabyte of code and text that I've created in the last week—and it needs the context to remember everything that it was already working on.

So it's never enough. At the rate that everything's accumulating, it's just never enough. This helps a lot, actually.

Is there an upper limit to the context window, or is it purely compute- and RAM-based?

There's no theoretical upper limit that I'm aware of right now for context window sizes. Do you remember when 64 kilobytes should be enough for anyone? It feels like we're in that era now, where one can reasonably foresee that, a few years out, maybe we'll have effectively infinite context.

When we find ourselves in that world, retrieval-augmented generation—RAG—maybe goes out the window. Maybe fine-tuning of models goes out the window completely. Why bother doing any of that if you can just dump your entire company's corpus of knowledge—code, documentation, emails—all into the context window and get effectively free marginal intelligence?

This is exactly why Blitzy is signing deals as quickly as they can have meetings, because Blitzy has this infinite-context-window coding capability. It took them a solid year and a half to develop, and now it can take an infinitely large context and restructure it to fit into whatever window's available.

So, Dave, just take a second, because we're going to be doing a podcast with the CEO of Blitzy. Blitzy is a company that you incubated and supported and that Link Ventures basically financially backed. Give me some context here on Blitzy.

Yeah, so Blitzy is 2 best friends from Harvard Business School. One went to West Point and is an organizational genius, Brian Elliott, and Sid Pardeshi, his co-founder, is a technical genius from India who also came to Harvard Business School. They founded the company together in our office. It's still in our office—it's taking over the office space like a Borg right now, just eating desks.

Like I said, they're signing deals as quickly as they can have meetings. Their Harvard Business School professor also joined the company, so that's a good sign. And then my youngest son also jumped on board because it's just—

What do they do?

Sucking everything up. They write 2, 3, 4, 5 million lines of code in a night that's all fully debugged and functional the next day.

The original use case was, "Hey, there's all this legacy code—mainframe code that hasn't been touched in a decade—which is incredibly expensive to maintain. Can AI come in and just rewrite all that in a modern language, make it much more efficient, and move it to the cloud?" That's a lot of their bread and butter.

But now they're moving on from there into greenfield: what can we create from scratch that didn't even exist in the world before? The specs for these, before you even launch the code, become 100-, 200-, and 300-page documents, all written by AI. It's hard to keep up and even proofread them before you hit the go-ahead-and-build-it launch button.

They're really on the forefront of building really big things in really short time frames and having the system debug and fix itself, with something fully functional coming out the other side.

I love that.

Coming out the other side, we have Perplexity making a $34.5 billion bid for Chrome. You guys really love Perplexity. I've just started using it to compare it to Google.

This is a $34.5 billion unsolicited offer. You can imagine you're sitting at Google headquarters and someone comes up and says, "I want to buy your favorite child. Here's $34 billion." The offer is larger than Perplexity's reported valuation, and I guess the concept here is that Alphabet and Google have been under incredible regulatory pressure that could force them to divest—to split it up.

You're making too much money. You're dominating the field. They are the projected winner of the AI race by a lot of the experts out there. So, thoughts here? Is this anything other than a PR play, Salim?

I think 2 things. One is a PR play, I think. Second, here's what I predict happens: I think the Trump administration gets involved and says, "Well, you want to give us 10% of Chrome to help you keep it protected." That's what I think happens.

Oh, good. Yeah. Well, what was going on here, though, is Google was claiming that Chrome can't be split from Google because it's useless without Google and nobody would want it. Perplexity wanted to show the FTC, "We'll take it, and we'd be willing to pay for it."

So it absolutely is not inseparable, irrespective of Google saying, "No, that's absolutely not true." It's all part of the grand strategy. I think they would buy it if Google is willing to part with it. But it's more—not PR, it's business strategy trying to keep the FTC active in this breakup.

I love this next sequence of stories. Meta freezes AI hiring after going on a blitzkrieg—we'll use that term—and offering everybody—I haven't gotten the call yet. I'm not sure if you got the call yet, Dave, and I guess I'm not going to expect the call yet from Zuck.

But after hiring 50+ researchers with salary packages in the tens of millions, to reportedly a billion dollars, Meta has stopped its hiring. It's reorganizing its AI teams into 4 groups: AI products, superintelligence, infrastructure, and fundamental AI research.

I'm going to match that story with the following story, which is that Microsoft is fighting back.

What’s the Star Wars analogy here? Attack of the Clones or The Empire Strikes Back, or whatever it might be.

The Empire Strikes Back.

The Empire Strikes Back. Yes. And Microsoft is the Empire. Microsoft is now offering multimillion-dollar pay packages matching the enormous offers that Meta was making, and it’s trying to raid the Meta coffers.

Microsoft created an internal most-wanted list of engineers and researchers. Then they fast-tracked the hiring process. It’s been said that if it’s critical AI talent, you can receive an offer within 24 hours.

Now, I think the backstory here—somebody should make a movie immediately about this—but the backstory is really interesting, too. There was complete peace and détente for a long time among Apple, Microsoft, and Google, where they basically settled into this mode: Apple, you keep cranking out the phones. Our Microsoft phone failed miserably; we’ll give up on it. Microsoft Office, that’s our cash cow.

Google Docs, you need to sit there and not threaten Google Docs. In return, Bing will not threaten Google Search. Not all of it is written down anywhere because that would be illegal, but it’s clearly stable. That lasted for 10, 12, 15 years of stability.

Now they’re colliding and fighting like you would not believe over AI. They’re going full-bore after your best people—anything they can do to get a head start on you. We’ve never seen all the tech giants going after one brass ring before. So it’s really great for startups because we have turbulence and chaos, which is always great for the new guys coming in.

Dave, that is beautifully said, buddy. That is beautifully said, Salim.

Well, there are 7 big companies, and there were 7 kingdoms in Game of Thrones, and it feels to me like that is what’s going on. I totally agree with Dave.

I was at Yahoo, and you’d see people very politely moving between companies, but it was very, very delicately done. There was back-channeling nonstop to make sure everybody was okay with it. It was one of the really great models of co-opetition that was out there. But now the gloves are off.

Yeah, it is winner-take-all. And here’s the other side of the equation. Here we have unlikely bedfellows. Elon tried to enlist Zuck in a $100 billion bid for OpenAI. This is more—this is just soap-opera land. I don’t want to say anything more than that, but that was fascinating.

That came out in the news. Then I want to close this segment on companies going after each other’s employees with this CNBC article. It says, “AI deals creating zombie startups.”

You go in and hire the CEO, the CTO, and leave the rest of the team there, sort of like, “Oh my God, how do we—” We just took $100 million of capital, and we can’t deliver our products. Dave, are you seeing these zombie companies? I mean, this is the best-friend model. Yeah.

You know. Yeah.

Yeah. No, it started with Character.AI. Google bought Noam Shazeer back for, wow, $6 billion or something like that, and that started the wave. Now it’s all the rage.

But that Windsurf deal was the real bellwether, because Windsurf was a very young company—18 months, something like that—acquired for $3 billion, and it was a huge windfall for OpenAI. Oh, wait: Microsoft blocked the deal, and then it turned into an acqui-hire of the talent into Google.

What did the shareholders get? That’s still in flux. I’m polling our seed-stage friends who are in that deal, saying, “Did you get ripped off and wiped out by this, or what does this mean for the venture world?”

Venture capitalists I talk to are not super worried about it. It’s not affecting that many of their investments. But it really wrecks the whole venture landscape if this becomes the de facto standard, kind of the endgame for a great startup.

There’s a whole meme around the fact that VC as a category is going to be ending soon, with this as one of the jigsaw-puzzle pieces that breaks the whole thing. It’s a really bad problem from a VC perspective, because if you don’t know what’s going to happen with that startup, what are you going to do with it? Especially in a hot area, right? Traditionally, that’s where you funded it.

If somebody comes and licenses the tech and then hires all the people, you’re left with a shell.

Expect to see a bunch of lawsuits coming out of this.

Probably a lot of this is also the antitrust laws, right? There were a lot of challenges in being able to acquire companies. There wasn’t an IPO window for the last number of years. Finally, that’s opened up.

You try and acquire the company, and then antitrust would say, “No, you can’t do that.” So, as with Alexandr Wang, you’d buy 49% of the company and effectively get control of it. And now instead, why don’t you just hire the talent?

Yeah, that’s the deal. You buy 49% non-voting, so that doesn’t trigger the FTC and Hart-Scott-Rodino and all that. Then you move all the best employees over with huge pay packages. There’s a commercial deal where you license all the technology, and that’s not disclosed, so you have no idea what’s buried in there.

It could be like, “You owe us all of your children for the next 10 generations,” for all we know. You just don’t know. But that’s the standard deal because it doesn’t have to go through the waiting period. In the race to AI, the big tech companies are desperate to move as quickly as possible.

If you read Accelerando, Alex will recommend that to you all day long. Manfred, the character in the beginning—the whole structure of the way things are created has to get rethought because of the pace of AI. This is kind of the first foray into that new terrain that we’re seeing.

Just in terms of the venture community, we have multiple companies reaching multibillion-dollar valuations in under 2 years. Prior to this wave of AI, I only counted 7 times in the history of the world that that had ever happened. Now we have 3 in a single portfolio, so the tailwind for venture is unbelievable—bigger than ever, ever, ever before.

I don’t want to leave the impression that something’s about to fall apart in venture. These are relatively rare deals, but it is the first bellwether that the new economy is coming, and some things need to be rethought just for speed.

A few quick articles about xAI. xAI co-founder Igor Babuschkin exits to launch an AI safety venture amid growing executive turnover at xAI. Andrej Karpathy, who headed Tesla’s AI systems, is thought to be coming into xAI. We’ll see if that gets announced. Alex, any thoughts on this one?

Yeah, I loved Igor’s farewell note regarding xAI. He told his life story of how he first went into AI because he wanted to solve science. In some sense, this is the perfect time, now that AI is arguably on the verge of solving science, for him to strike out and fund ventures in that area.

Nice. I love this article, again, in the Elon Musk universe: “Elon on AI increasing birth rate.” Here’s a tweet: “AI is obviously going to one-shot the human limbic system.” I love that.

This, of course, is the idea that—why get married? I have an AI girlfriend. I have an AI robot. But he goes on to say, “I predict, counterintuitively, that it will increase the birth rate. We’re going to program it that way.”

So, a couple of quick thoughts here. One: we’re going to program it that way. Your AI is subtly telling you, “Hey, you should have kids. Hey, go get another girlfriend. Hey, Elon’s got 15 kids. You should have 15, too.”

So, I looked at this when I saw it in the prep for this, and I don’t understand any one of those statements at all. If any of you could explain any of those—what do you mean, AI is going to one-shot the human limbic system? How will that happen?

Any insights from any of you, I would really love it, because this makes no sense to me at all. Mm-hmm.

So, I think I understand it. I just don’t want to comment on it. What I will say, though, is that, putting aside understanding any change—AI-driven or otherwise—in the birth rate would take decades to be felt in demographics, and the changes we’re seeing in AI otherwise right now are on a much shorter timescale: months to a low number of years, not decades. So I’m not sure it really ends up mattering either way.

Well, just taking a second to think about the decreasing birth rate, we’ve discussed at length that places like South Korea, Japan, China, and much of the world—other than what we’ve seen in India and Africa—are below the replacement level of 2.1. Some countries are dangerously as low as 0.7, and they’re literally sublimating. They’re evaporating. The question is, why is the birth rate going down?

A couple of things. One, women’s education is going up, so women are desiring to stay in school more. Number 2, as people move into the cities, it’s more expensive to bring up kids. Number 3, the child mortality rate is lower. The number of children per family back in the 1950s was, on average globally, above 5.

We’ve seen this precipitous drop because kids are living. You don’t need to have an extra 2 or 3 kids to make sure they’re there to work the farm.

So all of these have reduced the birth rate. The question, I think, logically is: if we do have abundance—if you have access to robots for helping raise your kids, if you have access to AI and universal basic income of some sort to help you with income—can we shift back to building families instead of having to work, or instead of having to make a choice between work and a family? That might be part of the Elon Musk counterintuitive approach.

Well, this quote—I know Alex needs to keep his reputation pristine because he does a lot of work for government agencies, and also nobody wants to irritate Elon. But this quote from Elon goes hand in hand with the one he had a year ago, where he said, “Look, of course, AI is going to be smarter than all of us. We’re not as smart as we think we are.” Basically, the undercurrent here is that AI is going to be incredibly persuasive very soon.

It is persuasive already. It’s more persuasive than the best humans.

Well, so then the purpose of that last quote—I don’t know why he’s being so honest about this, but, yeah, the natural state is the birth rate’s going to plummet to near zero. But the AI—we’re going to program it to convince you that it’s a good idea to have kids as a way to stabilize the human population. That’s a dangerous message: that we’re going to program AI to influence you on anything.

Yeah, exactly. That’s why it’s hard to touch this slide, but that is exactly what he means. I don’t know why he’s saying it because the backlash would be huge, but he’s saying it.

All right. I love this. Musk acknowledges that Google is the leader in AI. Elon concedes that Google currently has the highest probability of being the leader in AI, citing its massive compute infrastructure and data reserves. Google’s dominance is backed by $85 billion in AI-related capital expenditures and strategic investments, like a 14% stake in Anthropic.

So we looked at all the prediction markets, and even when GPT-5 was announced, we saw the prediction markets suddenly flip to, “Yeah, Google’s going to win the race by the end of August, by the end of the year.” Alex, do you believe this as well?

I think this is more a reflection that winning, quote-unquote, is a combination of spending enormous amounts of capital to build out data center and compute capacity, plus default distribution that could come with billions of users from an existing service or from a new service like ChatGPT that’s just emerging and reaching toward 1 billion, perhaps by the end of this year.

That winning combination of both enormous capex for compute and default distribution—that’s what I think this quote, and more broadly the conventional wisdom right now, suggests is what it takes to win as a frontier lab in AI.

Salim.

2 things, positive and negative, here. The negative is, I think, as a builder of cutting-edge AI, we’ve always seen that a small team will outperform a big company, always. Therefore, this investment in Anthropic and so on is super smart here, but Google itself, I don’t think, will do it. It’ll rely on some of these external teams.

On the positive side, their access to compute and infrastructure is so ridiculous. As Dave mentioned earlier, at the upper end, there is simply infinite demand for compute. That may be the reason why they win.

Yeah, maybe. I think the real race, actually, under the covers—the real race—is between the TPUs at Google and the Dojo chips at xAI or Tesla, wherever those are theoretically made. The Elon chips, because he just signed that $16 billion—which is really more like a $40 billion deal—with Samsung to manufacture those chips. That’s a lot of chips.

I think you’ll get the data centers to put the chips in. The question is the relative performance of the Dojo chips compared to the TPUs and the other next-generation chips that are getting designed right now. So it’s a footrace. He’s not conceding because he wants to just say, “Don’t look over here. I’m going to be working on this over here.”

Perhaps. At the same time that Elon’s making that statement, here is some interesting data: users are choosing Grok over Gemini. In terms of the star rating, Google Gemini is at 4.8 with 394,000 downloads and ratings, versus Grok at 4.9 with 502,000 downloads and ratings. It’s interesting.

Maybe—listen, part of this is that Elon’s got the largest megaphone of everybody, being able to, on Twitter, on X, say, “Hey, check out what X can do. Download it here.” That works pretty damn well.

That’s exactly right. And he’s using it aggressively. In theory, Google has a bigger megaphone because Google Search is actually bigger than X. But it cannibalizes—that’s the issue.

If Google were to push this as hard as possible, the extreme would be to say you could only use Google Search through Gemini. Then they’d bypass Grok in a heartbeat, but it would cannibalize their entire revenue engine. Grok doesn’t have that hang-up. So it’s really an interesting little balance in this great war that’s going on.

What I love is the fact that with the ratings—4.8 and 4.9—regardless of who’s ahead, the users win.

Yeah, that’s a good point, Dave. There’s another point here worth discussing that you and I have discussed in the past, probably not on our Moonshots podcast, which is that having a celebrated CEO who’s out there makes a big difference, right?

Elon, love him or hate him—and I would never bet against him—is out there constantly putting himself out there, tweeting 30,000 times a day. Not literally. I’ve been with him at parties and events, and he’s on his phone, tweeting away. Then he’ll pop up and have a conversation with you, and he’s back into his phone. So it’s him, to a large degree. I’m not sure anybody else is posting for him, but we don’t see that from Sundar.

We see this from Sam. Sam is out there as much, but doesn’t have the platform yet. I’m certain that OpenAI will eventually create its own platform equivalent. We’re seeing a little bit from Dario and Anthropic. But I think that’s so important. You want to speak to that?

I totally do. I’m so glad you brought it up, because when Elon did Saturday Night Live, that was the turning point where the definition of what it means to be a great CEO completely flipped on that day. Here’s the busiest guy on the planet, finding time to go to New York and be the host of Saturday Night Live. Why would you make that choice? It’s not random. It’s not ego. It’s part of a strategy.

Why are you doing this? Because it clearly works for recruiting and capital raising. It attracts talent. As you’ve been saying forever, Peter, you have to have an MTP. You have to have a massively transformative purpose that improves the world. But just having that purpose and not broadcasting it doesn’t recruit.

If you have the purpose and you broadcast it, then talent floods to you. They’re coming to Elon, they’re coming to Dario, because you’re out there and people recognize you. Social media is the cheapest form of media in the history of the world. If you don’t embrace it and get out there, it’s just the way you win as an entrepreneur now.

You embrace it. You get your voice out there. You get your face out there. Again, for the entrepreneurs—we have an incredible population of entrepreneurs and builders who watch this podcast—just a piece of advice: if you are passionate, if you want to change the world, it’s not enough to just be on your computer putting out code. Either you or your co-founder, someone on your team, needs to have an outsized personality out there letting people know what you stand for and what you’re doing. I think it’s critical.

All right. Let’s move on here to our next article. It’s “Google drops AI model that runs on 1% battery power.” Gemma 3 270M is a tiny, smart AI model that runs right on your phone. It can handle 25 chats using just 1% of battery power on a Pixel 9 Pro.

Again, this is AI that’s being infused power-efficiently into your phone and soon into everything that you touch, feel, and use on a day-to-day basis. Thoughts, gentlemen?

I’ll chime in and suggest that I think this may actually be more instructive regarding what the future of foundation models and frontier models looks like than, say, the large, high-parameter, multitrillion-parameter sparse models.

My suspicion is that, if I had to predict what the most futuristic possible frontier AI model looks like, it probably will look like some sort of nanokernel, maybe with far fewer parameters even than Gemma 3 with 270 million. Maybe it’ll only have a few million parameters. Maybe it’ll dispense with the notion of parameters entirely, but be like a small diamond nanokernel of an AI that hasn’t memorized the world’s knowledge.

That’s all externalized to some external database or knowledge base, but it’s multimodal. It understands video, text, and audio. It’s able to reason, but everything else—everything else—is externalized. I think this gem, this perfect, crystallized superintelligence, is where all of this ends.

Fascinating. The value—the reason we’re going to these small models—is simply power. Part of it is just the desire to push inference compute to the edge, to enable local chatbots on phones.

But I think ultimately these SLMs—small language models—are going to power the largest LLMs, the frontier models, in the data centers as well. The question is: Given that we arrived at this AI revolution by compressing all human knowledge into as small a model as possible, is there another phase transition where we can compress it even further and figure out what that nucleus of superintelligence looks like? I think we're going to get there in the next few years.

Well, more importantly, this enables that lawnmower-checking-the-weather-LLM type thing, right? Because you can embed this into everything. I think that granularization allows you to go up a little long tail of edge cases, of which there are an infinite number.

Sure—sorry, that vision—

No, that vision that Alex laid out has 2 parts that I'd love to add to. One is that when you take one of these 1- to 10-trillion-parameter mega-neural nets, all world history is baked into those parameters. Just a massive amount of information, most of which you don't need.

So there has to be this diamond nugget that Alex was describing, that has to exist, where it can call on data just by looking things up, but it has the same level of brilliance as the big model. It just doesn't have all that waste. And then the other thing is, if you talk to the Liquid AI team, they're really fixated on this notion that these Apple devices have an incredibly good neural processor in every single phone and every single laptop, and it's largely unutilized.

If you can move that diamond nugget into that latent compute that's in hundreds of millions of devices, it unleashes a huge amount of intelligence, especially during this next 2- to 4-year window, when big data centers are struggling to catch up. So there's a big short-term opportunity in unleashing all that compute in some useful way, and this takes advantage of that.

All right, I'm going to move us along. The next article is from CNBC: The U.S. government takes a 10% stake in Intel. Oh my God, what a story this has been, right?

The White House invested—I think, actually, the White House and the CHIPS Act granted Intel $8.9 billion. They're turning that grant into an investment. I'm so curious about what conversations took place in the White House when Lip-Bu Tan met with the president. We've seen Intel stock rise on this news. Our friend Leopold Aschenbrenner, who took out call options on Intel, has made a killing in that process. Dave, you and I were talking about that the day we discovered it. Oh my God, what a trade that would have been.

Well, hey, watch the podcast. Listen closely. Get those tips. But we didn't know that Lip-Bu Tan would be in Donald Trump's office 2 days after the podcast came out. All we knew was that Donald Trump had called him to the mat. But remember, we said on that podcast that what would happen next is Lip-Bu would meet with Donald Trump. When they came out of that room, you'd know whether they had cut a deal or struck a war.

If they cut a deal, what you'd see is Lip-Bu staying in place and Trump saying really nice things about him. We said it, and that's exactly the way it played out.

And that was followed up by the—

The call option was priced at when we had that conversation.

The short-term ones were up 100x from that day when I texted you. I don't know what Leopold bought, but—

It was at 3 cents for the call option, then it jumped up to 3 bucks.

That would have been the investment, for sure. After every one of these podcasts, I'm going to call you, Dave, and say, “Okay, what's the investment today?”

This was actually done a little bit by Obama when the 2008 financial crisis hit. The U.S. government gave a huge chunk of money to the car industry to save the car industry, and they got paid back in spades because it turned into kind of an investment, and then they re-liquidated. So this is very similar to that.

Yeah.

This is a really big deal for Intel, too. We met with Greg Lavender, the CTO, back before Gelsinger got fired. He said that CHIPS Act money has so much crap attached to it, so much baggage, because the way these things go through Congress, they add garbage to it to the point where it's useless. So they received $10 billion of unusable money.

So it was of no value. Trump, being the business guy, has restructured that into an equity investment. Now they can just use the money.

Nice.

So you'll see some serious motion coming out of that money. At the same time, we're seeing Masayoshi Son from SoftBank signing a $2 billion investment deal. Of course, SoftBank also owns 90% of ARM, whose chips power 99% of smartphones. It's really smart to follow what Masa is doing here, getting into the chip industry.

The other point here is that Intel is too big to fail, too strategic to fail.

For sure. I mean, that would have been the obvious conversation as Intel's price was falling. The question was whether Intel was going to be broken up and sold to other U.S. companies.

Anyway, here from Bloomberg: Apple expands iPhone production in India for U.S.-bound phones. So all iPhone 17 models are being built in India, not China. And they're being built by Foxconn, right? Foxconn built all of this in China, literally took over the Chinese economy, and has now moved it to India. They're going to be shipping this next month. That's a big deal.

I'm actually getting on a plane to India right now.

Do you have to go to India to get your iPhone? Are they shipping it to you?

I want to make sure the production quality is high because in India, component quality is a bit of a question sometimes.

Oh my God. All right. It's interesting, right? There's so much pressure to get out of China right now. Wait, let me make one more quick point about that. This is a very big deal because one of the problems that India has always had is this perception that you can't build high-quality things there.

And this will shatter that perception. I think a floodgate of manufacturing will start to move to India. Crypto and tokenized economies are going to be so important, especially as they connect to AI. It's going to drive the future, especially with agents being able to trade currencies.

That's the AI part.

All right, let's jump into this article here from CNBC: Albania wants to replace its corrupt government with AI. Prime Minister Edi Rama has advocated for AI ministers and prime ministers. It's a big deal. We've seen a little bit of this in the UAE and a few other governments, but this would be a fascinating move. Salim, comments?

I met Edi Rama a couple of years ago, part of the circle of heads of state that revolves around our ExO World. He did an incredible job turning around Tirana, the main city, as mayor, and then went to the national level. He found that the government is so corrupt that you need to do something, and you can't get out of it in an easy way. Coming at it from the top down with AI is a super-smart thing to be able to do.

I expect to see this across the board. At the very least, having AIs that are monitoring the activities of ministers and so on would allow you to reduce corruption. In Colombia, for example, there was a port being built, and all the government folks bought all the land around the port and then sold it for 100x just after that announcement.

You have that kind of institutional corruption. Congresspeople in the U.S. can have insider trading. That's kind of incredible to me that the public allows that. This is the kind of thing where AI can oversee some of this and start to make a kind of hacking into that problem space, which is huge, because the corruption problem is worth multiple trillions to the global economy.

So Dave, to you here: An MIT study reports that 95% of AI pilots are failing. Companies have spent $30 to $40 billion on generative AI, yet 95% see no financial return. The adoption rate is high: 80% are testing, but 40% deploy, and the impact has been low. Big firms are running pilots but struggle to scale.

In this study, 95% of the failure is a failure to deliver financial benefit, and the study goes on to say that it's principally because the companies don't understand how to use the AI tools properly. There's a learning gap. Companies that buy existing AI solutions succeed two-thirds of the time, while those that try to build them internally do not succeed. Only a third of them do.

We saw a huge stock-market dive as investor fears hit on this. Is the AI bubble real? Interestingly enough—and I think this is one of the most important things, Salim, you and I talk about in the ExO World—startups achieve a much better return on investment regarding AI because startups have fewer entrenched bureaucracies and business processes.

They don't force AI into existing workflows. They are AI-native, and they reinvent their business based on AI.

I can't stress this enough. We've been working with CEOs for the last couple of years on this. It is imperative to structure yourself as an ExO, right? That's one. There are 2 failure modes that we see. An ExO is an exponential organization, right?

Yeah. So this is the model that we've been pushing, though it is now shown to work. I mean, companies using this model are delivering 40 times the shareholder returns of companies that aren't. I mean, it's just absurdly obvious once you see the model.

There are 2 failure modes that we see. One is people jump into the water without looking where the rocks are. We came across a medical CEO who had uploaded all the sensitive patient data into ChatGPT and now has huge legal exposure because of that. So that's one bucket of challenges.

The second, bigger one, is cultural resistance. If you don't get the culture right, people inside the company are scared it's going to take their jobs, and it's a mess. So those are 2 big, huge buckets that this is essentially pointing out here. Dave, you're living this world right now with all of the startups and ventures.

Yeah, I'll tell you, don't read these reports. It's not worth the time. I got to be careful what I say here a little bit because we all love MIT—there's no place on the planet like it. The disconnect between the students and the faculty is like nothing I've ever seen.

I work with the students every single day, all the startups, and they are killing it. Back in 2020, we did this little research thing, and about 14 MIT alums of all time, out of 140,000 alums, had become self-made billionaires—14 as of 2020. Now, Greg Brockman at OpenAI, Alexandr Wang just got acquired by Google, and Mark Chen—I mean, they're everywhere. They're absolutely thriving in the world of AI.

Correction: Alexandr Wang got acquired by Meta, not Google, right?

Oh, sorry, sorry—by Meta. Yeah. So, actually, you've got top guys at Meta, OpenAI, and Google coming out of MIT—in Greg's case, dropping out of MIT—and they're just killing it out there.

Meanwhile, the administration keeps cranking out these documents saying, “Slow down, chill out.” Why? What are you talking about? Well, what it really is is, “We didn't invent it. Demis Hassabis got the Nobel Prize. Geoffrey Hinton got the Nobel Prize. I wanted that Nobel Prize. This isn't really happening. It's something else.”

The theme in all of these reports is that there's still a missing component. Some self-reflective, self-thinking thing still needs to be invented to make these truly AI.

Well, I think part of this is that we're still early, and we see these boards of directors—the chairman goes to the CEO: “Listen, what's your AI strategy? You need to have an AI strategy.” The CEO or the CTO is basically just throwing money at this without properly thinking it through.

It's the impedance mismatch between an established large company that's doing everything the way they've always done it for 10 or 20 years, trying to force AI into the mix. One of the last points on this chart here is that companies are potentially wasting AI by focusing on marketing and sales versus cutting the cost of back-end processes and operations. What are the use cases?

Yeah. So let me just say a couple of things. There's only one path to navigate this for companies. The bigger the company, the more this has to be followed. Smaller companies, as you mentioned, Peter, can adapt very well. And Dave, you're exactly right. This is a boom that's just going to keep going, so don't slow down. Don't let this affect you at all.

For a big company, there's only one model that's going to work. This is what I've been advising when I talk to the big-company CEOs that we talk to. Go create an edge organization that's replicating the functionality of whatever you're trying to do. Let's say it's building cars. Create an edge organization that's completely AI-native.

And start automating use cases bottom-up, one by one. Then you create a young-entrepreneur mindset with the youngest employees possible and let them loose with AI, copying the functionality of the big company. Now you're essentially doing A/B testing and seeing who can do it better.

Over time, you can deprecate the mothership and, little by little, move all of this to the edge, and that becomes a new center of gravity over time. Do not try to transform the mothership.

And that organization—the organization on the edge, that CEO of that team, this locked skunk-works equivalent—needs to report directly to the CEO. Don't put them underneath the other organization. It needs to be independent, allowed to iterate and do stupid things. This is where Kodak goes bankrupt even though they invented the digital camera.

We have solved this problem, folks. Anybody struggling with this, just call us. There's a 10-week engagement that we run. In a big company, the default answer if you try anything disruptive is no. Everybody becomes very French and they go, “Pas possible. Can't do it.”

But then you need to do the thing on the edge because even if you switch to a yes, you can't get out of the old models quickly enough. You have to do this thing at the edge and let the center of gravity drift to that over time. I'll get off my soapbox. Totally. Totally.

There's a lot of energy.

The other thing: get your corporate venture fund back up and running. A lot of them, like Intel Capital, got shut down just trying to save money. On that prior slide, Google owns 14% of Anthropic. That's a $14 billion position by itself.

Get that corporate venture fund up and running again, and then be a development partner for some of these startups. Try and be their first or second customer. Be super supportive of them and invest in them at the same time.

Then that group that you've invested in plugs into your internal edge group that Salim just described. That's how AI knowledge is going to actually get into your organization, because the reason these corporate things are failing is not because the AI is failing. It's because you're just throwing it into a group of people who have no idea even how to start using it, and they kind of don't want it to take their job away. So they don't have a huge incentive to try to make it work.

So our next subject here: OpenEvidence gets a perfect score on the U.S. Medical Licensing Examination. This is huge, right? We've seen the data before: a human diagnostician gets 72%; a human plus GPT-4 was getting like 74%; and GPT-4 on its own was getting 92% accuracy in diagnostics.

Here we see another version of this on the U.S. Medical Licensing Examination. Again, this is to become a full doctor, right? You've done your internship and residencies. OpenEvidence is at 100%, GPT-5 at 97%. Again, this is AI taking the lead. Pretty extraordinary.

I love the fact that 40% of doctors in the U.S. are already using this. That's very inspiring to me.

Yeah. Listen, every doctor is going to be using this, or they're not going to be a practicing physician. But here's the real topic I wanted to hit on.

Sam Altman is getting into the BCI, the brain-computer interface, race with a company called Merge Labs, which is targeting the combination of gene therapies and ultrasound as a mechanism to be able to read and write onto neurons. And I know the company well. I don't know what I can say.

I'm hoping to have this company on stage with me at the Abundance360 Summit in March. Also, I'm super pumped that Kevin Weil, the chief product officer, has agreed to come on our AI Day and talk about how fast GPT-5 is, when we'll have GPT-6, and when we'll have AGI.

But check this out. There's probably, my guess, Salim, I don't know, 20 BCI companies—probably about 4 or 5 that I'm tracking—that are extraordinarily effective and moving rapidly. Again, when Ray Kurzweil made his prediction of a high-bandwidth brain-computer interface by 2033, I was like, “Ray, you got this one wrong.” And no, he got this one right. We're going to see that. Extraordinary.

Man, the man is incredibly annoying.

Yeah, incredibly annoying. And Alex, you know this team here, and you know one of the co-founders as well, right?

Yeah. My fellow Hertz Fellow, Mikhail Shapiro, is a co-founder. I'm a huge fan of the company. My sense is this is such a rapidly moving space. I would love to see Merge Labs and all of its competitors deliver high-bandwidth BCIs in the next few years.

I think if that window of the next few years for high-bandwidth BCIs isn't achieved, the risk is always that we achieve sort of a pure AI economy that completely decouples from the human economy. BCIs, I think, are our best hope, if they can deliver quickly enough, at keeping the human and machine economies coupled.

And you were on stage with me at Abundance Summit 2 years ago talking about the importance of coupling, right?

Right. So either we are on the AI team fully and we are literally coupled—and this is the quote from Sam. Sam Altman says, “The merge can take a lot of forms. We could all just become really close friends with a chatbot, but I think a merge is probably our best-case scenario.” We've seen this from Elon as well.

The idea is that being able to connect our neocortex to the cloud allows us to ride on top of AI's acceleration, versus being left in the dust. Fascinating stuff. All right, this is equally fascinating, and again, this is just showing that AI and science are going hand in hand.

OpenAI's GPT-4b. You know, I gave Kevin Weil a hard time about his naming protocols for his models. Anyway.

Super funny.

Yeah. OpenAI's GPT-4b designs proteins to reprogram cells as stem cells. You've heard me talk about the Yamanaka factors. This is a Nobel Prize won by Professor Shinya Yamanaka in Japan, funded by Marc Benioff.

What we've seen here is GPT-4b being able to come up with a new version of these transcription factors that's 50 times more effective. I remember, Alex, when you showed me this, you were like, “Okay, here comes longevity escape velocity.”

That's right. I want to get out my “Ray Kurzweil is right” hat for this one. This is what AI-driven longevity escape velocity, at least the early glimmers of it, looks like. A generalist model making seemingly a breakthrough discovery in longevity.

I would encourage everyone, if you use ChatGPT and you haven't played with the built-in molecular biology and biochemistry support in the form of RDKit, to try it for themselves with any recent form of ChatGPT.

Go develop a new pharmaceutical. This is now table stakes.

Well, this is also where Demis Hassabis has the opportunity to become the most important figure in human history, because he is the Nobel Prize-winning Demis Hassabis. He's spending all of his time on the cell simulator and just solving all disease, but he's also planted within Google. He has access to an immense amount of compute and academic freedom to work on it.

With that and a Nobel Prize, he's in a position to change the world more than anybody. I can't wait to talk to him. The rate of progress here is unbelievable.

When people say, “Why are you so excited about longevity and health-span extension?” it's this: it's the impact of AI. It's nothing else that's going to move the needle this fast. We saw it on my podcast with Dr. David Sinclair—his use of AI to create these molecular equivalents of what was only possible with gene therapies before.

We almost need a whole episode dedicated just to the intersection of AI and biotech and protein folding and Yamanaka factors and all this stuff.

We'll get Demis on this podcast. We'll have that conversation for sure.

All right, let's close out with robotics here. We can't avoid the robot revolution. The robots are coming. This year, at Abundance 360 in March, my plan is to have at least 4 of these companies, maybe 5, there with their robots so we can play with them, see them, and meet the CEOs.

One of the key things that's going on out of China is that we're seeing these robot clusters in the United States and, principally, in China. We've seen the first humanoid robot games in Beijing. Think hip-hop, soccer, boxing, and track. Here's a quick look at the video.

Here's the hip-hop portion. Here's the opening. Here's the soccer portion, a little bit of boxing, and track.

In the track section, Unitree—I had Unitree last year at Abundance 360—set a 1,500-meter world record, but it's still 91% slower than a human. A quick look at the video here. The real news here is what China is doing, right? Creating these games and creating these clusters around humanoid robots. They're iterating the cycle much faster. It's amazing.

Yeah. I think that's the key part: the fact that they're making it a public competition. This feels to me like FIRST Robotics on steroids, in a funny way. I will make a prediction that these types of robot-to-robot games will fail, just because we watch the Olympics for the human factor, not for the speed at which somebody ran. So that's my prediction.

Yeah, that's probably a good prediction.

I think the higher-level topic is really important, though. Benchmarks are critical for inspiring people to keep moving forward, and this is a form of benchmark. You're right, it'll come and go, but there'll be some other benchmark. As long as you're inspiring people to show off what they can do and compete, then this thing is going to drive forward very quickly.

Yeah. All right. I remember seeing this. Remember Scott Hassan's company in robotics?

Yeah. They spent forever trying to get a robot to fold laundry.

Yeah. And here we go. This is our friends at Figure AI. This is Figure 02. This is fully automated. This is not teleoperation, which is a really important point to make. Here we see it folding laundry.

I still think there's a human being in there.

Well, they said it's not.

You mean inside the robot?

Okay. I found this fascinating as well. This is Figure's Helix. Again, here we see a robot company—in fact, every robot company we've seen that has partnered with a frontier-model firm has started building its own AI models. Figure is no exception. Brett Adcock said, “We're building Helix,” which is their sort of AI for navigating the physical universe.

Here we see Figure 02 walking through very rough terrain in a very human-like fashion. I mean, this is pretty extraordinary.

You know—

Also, you remember when we were talking to Bernt Børnich a couple of weeks ago at 1X Technologies? We said, “How do you debug whether it's physical or mental that's not working when it makes a mistake?” He said, “Well, we have a teleoperator try and do it with their hands or with their own feet, and that tells us if the robot can do it or not. If the robot could do it, then we know it's in the brain. If the robot can't do it, then we know it's in the robotics, in the gears.” But you can see in those videos that that's reaching the end of its life cycle, because when it's folding laundry, there's no teleoperator who can be moving the hands remotely at that speed and dexterity.

He mentioned that when we interviewed him. It's—

Yeah. Go ahead, Dave. Sorry.

No, he mentioned that when we interviewed him, that he was right at the edge now of where that mode of debugging was going to continue to work, because the robots are getting ahead of anything a remote-control operator can do.

What we just saw in the video, for those of you listening, was Figure 02 walking along this junkyard of wood planks and almost tripping, but catching itself and walking elegantly across, which was pretty amazing.

It's been rumored for some time that Apple would get into the robotics business. They went almost into the electric-car and autonomous-car business. What we're seeing here is Apple expanding into a new set of AI-enabled devices, including a tabletop robot. This is sort of an iPad on a stick that can look around. They're going to go into smart speakers and AI-enhanced home security cameras. Hopefully, some version of Siri that can spell my name correctly.

I don't know. The thing that frustrates me the most about Apple is that I'm texting two people, their names are obviously there in the text line, and they spell both names wrong. It just drives me nuts.

Drives me nuts.

I was going to close on this article out of China. China is developing the first humanoid robot with an artificial womb. Kaiwa Technology is creating robots that can carry a fetus in a synthetic womb with fluids and nutrients, at a cost of $14,000.

Maybe this is part of Elon's prediction of AI and technology reversing the decline in the birth rate, if you don't need to carry your own child anymore. But seriously, there's surrogate pregnancy. One of the things that's interesting—and I'll have Ben Lamm on stage at the Abundance Summit—is that Ben is the CEO of Colossal, the company that's de-extinguishing the woolly mammoth, dire wolves, many other dodos. In order for him to actually hit his marks, he needs to build artificial wombs. But they're not going to be carried around by a robot. They'll be stationary, physically in a room, and carefully guarded.

See—

I don't understand this carrying-around bit. This reminds me of Brad Templeton saying, “You have all these robot horses, and people are spending huge amounts of money creating robot horses. Give me a male horse and a female horse, and I'll grow you a horse.” Oh, my God. I don't know.

But I am going to read this final closing line, and we'll maybe comment and break on this. This is from a guy named Dr. Singularity. I love the quote. He says:

“In the 1960s, Star Trek envisioned a distant utopia, placing warp drives, replicators, and advanced societies centuries away. Even through the 1980s and 2000s, the future was imagined as a slow, linear march of progress. But reality no longer moves linearly. We're extremely close to having AI agent researchers matching the brightest human minds. Soon we'll see millions or billions of them. A Star Trek world wouldn't wait for the 2200s. It would arrive by the 2030s.”

Amazing prediction.

I think it's a great way of closing out this episode. It's exactly where we are right now.

Yeah. And I would just add, Star Trek is such a strange future in the Star Trek universe. It's energy-rich. They have warp drives. They're traveling around the galaxy.

It’s biotech-poor. So longevity was outlawed in the late 1990s in the Star Trek universe. And it’s AI-poor. Everyone’s surprised when a human-level or superhuman AI pops out of the holodeck, whereas I think the future—and the present, frankly, that we’re finding ourselves in—are going to be rich and abundant in all 3.

Amazing.

Great points. Well, after we did that Kevin Weil interview, I started reading The Future Is Faster Than You Think again. It’s on my bookshelf, so why not read it again? But, Peter, that book has so many things in it that are coming true right now.

It’s unbelievably prescient, and I think it’s worth everyone grabbing a copy and looking through it. It’s actually much more relevant today than even when you wrote it because there are so many new converging technologies. You have some examples in the book of technologies from—I guess that was 10 years ago. Was it that long ago?

No, it was 2019. Yeah, 6 years ago now. But so many things have happened since then.

No, we’re living in dog years here. It’s compressed time. There are so many things that that book predicts. It’s worth reading again.

So, where are you? What are you doing in India, buddy?

I’m going for a bunch of conversations. The Singularity Summit is happening there, and they’ve asked me to do the opening. So I’m back in the country of my birth, running around and having a bunch of meetings. I’m in 6 cities in 5 days. The carpentry of this is going to be pretty ugly, but India—there’s just a piece of my soul that’s always there.

I love it. Love it. Alex, what’s next week look like for you, buddy?

Oh my goodness. Well, given this flood of AI innovations, I’m just immersing myself in it wherever possible. I advise a number of startups on how best to incorporate AI advances into their workflows, but really, I think smoothing out the singularity is the name of the game at this point.

Fantastic. And Dave, are you ending up anywhere in particular this summer?

Yeah, same as Alex. I have 2 weeks to grind through AI models and finish some things. Then we’re back at Stanford. We were at Google on September 8, then NVIDIA that night, and then Stanford all day the next day—2,000 people. The Blitzy launch will be then. We’ll do the Blitzy podcast right before that, and I think we’ll release it at that same moment. Then 60 back-to-back startups will be presenting on Stanford’s campus. So I have a little respite to finish a whole bunch of AI work before all hell breaks loose the first week of September.

All right. Wishing you guys an amazing end of summer here. What a great time to be alive, everybody.

I hope you enjoyed this podcast. We work to give you, hopefully, an increased IQ bump, excitement about the future, and a positive vision of where things are going. If you enjoy this, please let us know. We love your feedback. Tell your friends.

Next episode, more crypto, because there’s a ton happening there as well.

Yeah, no, we’re going to be diving more into crypto for sure. We’re grateful for you. We do this because we have a lot of fun. This is, for us, the way we keep on top of everything: actually doing the work, doing the research, and discussing it here. We hope it’s beneficial to you. Anyway, thank you, gentlemen. A real pleasure and honor.