Marc Andreessen introspects on Death of the Browser, Pi + OpenClaw, and Why "This Time Is Different"

Something about AI causes people in the field, I would say, to become both excessively utopian and excessively apocalyptic. Having said that, I think what's actually happened is an enormous amount of technical progress that built up over time. For example, we now know the neural network is the correct architecture. I'll tell you, there was a 60-year run—or even 70 years—where that was controversial.

So I think about the period we're in right now as what I call an 80-year overnight success. It's an overnight success because ChatGPT hits, then o1 hits, then OpenClaw hits. These are radical, overnight, transformative successes, but they're drawing on an 80-year wellspring—a backlog of ideas and thinking. It's not just that it's all brand new; it's an unlock of all these decades of very serious, hardcore research.

If I were 18, this is what I would be spending all my time on. This is such an incredible conceptual breakthrough.

Great, thank you. Apparently, this is the final few days in your current office. You're moving across the road.

We have some projects underway, but this is actually the original office. We're in the original office. We're in the whole thing.

I have to call this out. In October 2022, I had just made friends with Roon, and I wanted to give him something to be spicy about. I said, “You'll never not find it funny that a16z was constantly saying, ‘The future is where the smart people choose to spend their time,’ while going deep into crypto and not AI.” That was in October 2022, and Roon said there was an internal meeting at a16z to reorient around generative AI. Obviously, you have, but was there a meeting? What was that?

I don't know about all that. I've been doing AI since the late 1980s. As far as I'm concerned, this stuff is all Johnny-come-lately.

We've been doing AI our entire existence: AI, machine learning, deep learning. We've been doing this from the beginning. Obviously, AI is core to computer science. I actually view them as quite continuous.

Ben and I both have computer science degrees. We're both old enough to remember the actual AI boom in the 1980s. There was a big AI boom at the time under names like expert systems, in the era of Lisp and Lisp machines. I coded in Lisp; I was coding in Lisp in 1989, when that was the language of the AI future. This is something we're completely comfortable with. We've been doing it the whole time and are very enthusiastic about it.

Is there a strong sense that this time is different? My closest analogue was 2016–17. There was an AI boom, and it petered out very quickly. There just wasn't much investment excitement. Although that's really when the NVIDIA phenomenon took off—I would say it was in that period when it became very clear that, at the time, the vocabulary was more machine learning, but it was very clear that machine learning was hitting some sort of takeoff point.

Yeah. Well, as you guys have talked about this at length on your show, if you really track what happened, I think the real story is that the AlexNet breakthrough in 2013 was the real knee in the curve. Then it was obviously the Transformer breakthrough in 2017, and then everything that followed.

I've been working with Facebook since 2004, and I've been on the board since 2007. Of course, they started using machine learning very early. I've used it basically for 20 years for content-feed optimization and advertising optimization. Obviously, many financial-services companies and many companies in many different sectors have been doing this.

It's not a single thing; it's layers. The layers arrive at different paces, but they build up over time. In retrospect, 2017 was the key point with the Transformer. Then, as you guys know, there was this really weird 4-year period where the Transformer existed and it was just like, “It's cool.”

Between 2017 and 2021, that was the era in which companies like Google had internal chatbots, but they weren't letting anybody use them. OpenAI developed ChatGPT or GPT-2, and then they told everybody, “This is way too dangerous to deploy. We can't possibly let normal people use this thing.”

Do you guys remember AI Dungeon? For a while, the only way for a normal person to use GPT-3 was in AI Dungeon. You'd go in there and pretend to play Dungeons & Dragons. In reality, you were just trying to talk to GPT-3. There was this long period in which the big companies were cautious.

It took OpenAI time to redirect its research path. Even OpenAI, which has been the leader of this thing in the last decade, had to adapt and lean into the new thing.

I think it was at Rosewood, right? The dinner at which they founded OpenAI was right there. That dinner would have taken place in 2018?

2018? The formation of OpenAI was as late as 2018?

I'm sorry, no, I'm wrong. It should probably be 2015. They just celebrated their 10-year anniversary, so it is 2015.

Yeah, 2015. But then Alec Radford did GPT-1 in—

What? Probably 2017 or 2018, yeah.

2017 or 2018. Then GPT-3 was in 2020.

Because that became Copilot immediately.

Yeah. Even OpenAI, which has been the leader of this thing in the last decade, had to adapt and lean into the new thing. I think it's just this process of basically wave after wave, layer after layer, building on itself, and then you kind of get these catalytic moments where the whole thing pops. And obviously that's what's happening now.

Is it useful to think about whether there will be an AI winter? There are always these patterns. Is it endless summer? It's something I constantly think about, because do I just get endlessly hyped and trust that I'll only be early and never wrong? Or will there be a winter?

There's something about AI that has led to this repeated pattern. You guys know this: summer, winter, summer, winter, summer, winter, summer, winter. It goes back 80 years. The original neural-network paper was in 1943, which is amazing—it goes back that far.

There was a big AGI conference at Dartmouth University in 1955. Have you guys ever talked about this on your show? They got an NSF grant for all the AI experts at the time to spend the summer together, and they figured that if they had 10 weeks together, they could get AGI out the other end. They got the grant and the 10 weeks, but no AGI.

Like I said, I lived through the 1980s version of this, where there was a big boom and a crash. So there is this thing about AI that causes people in the field, I would say, to become both excessively utopian and excessively apocalyptic. We're probably on both sides of the boom-bust cycle. You see that play out.

Having said that, I think what’s actually happened is that, in retrospect, we now know an enormous amount of technical progress built up over time. For example, we now know that the neural network is the correct architecture. I’ll tell you, there was a 60-year run—or even 70 years—where that was controversial. We now know that’s the case.

Everything we’re building on today is derived from the original idea from 1943. In retrospect, we now know that these guys were right. They got the timing wrong, and they thought capabilities would arrive faster, or that it could be turned into businesses sooner, or whatever. But fundamentally, the scientists who worked on this over the course of decades were correct about what they were doing.

The payoff from all their work is happening now. The way I think about what’s happening is that the period we’re in right now is what I call an 80-year overnight success. It’s an overnight success because ChatGPT hits, then o1 hits, then Claude hits, and these are radical, transformative overnight successes. But they’re drawing on an 80-year wellspring, a backlog of ideas and thinking.

It’s not just that it’s all brand-new; it’s an unlock of all these decades of very serious, hardcore research and thinking. There were AI researchers who spent their entire lives on this. They got their PhDs, they researched for 40 years, they retired, and in a lot of cases they passed away. They never actually saw it work.

Yeah. So sad.

It is sad. Geoffrey Hinton was like the last guy.

There were guys out there like Allen Newell—I mean, tons of them. John McCarthy was one of the inventors of the field. He was one of the guys who organized the Dartmouth conference, and he taught at Stanford for 40 years. He passed away, I don’t know, 10 years ago or something. He never actually got to see it happen.

But it is amazing in retrospect. These guys were incredibly smart, they worked really hard, and they were correct. So, anyway, as they say, history doesn’t repeat, but it rhymes. Does that mean there’s going to be another boom-bust cycle?

I’ll tell you: in a sense, I guess everything goes through cycles. People get overly enthusiastic and overly depressed, and there’s a timelessness to that. Having said that, there’s just no question.

The four most dangerous words in investing—

“This time is different.” Do you know the 12 most dangerous words in investing?

No. The four most dangerous words in investing [laughter] are, “This time is different.”

And I’ll tell you what’s different: now it’s working. There’s just no question.

By the way, I’ll give you my take. From basically the ChatGPT moment through the spring of 2025, I think well-intentioned, well-informed skeptics could still say, “This is just pattern completion. These things don’t really understand what they’re doing. The hallucination rates are way too high. This is going to be great for creative writing and creating Shakespearean sonnets or rap lyrics, but we’re not going to be able to harness this to make it relevant in coding, medicine, law, or the fields that really matter.”

I think it was the reasoning breakthrough—o1 and then R1—that basically answered that question. It said, “No, we’re going to be able to turn this into something that works in the real world.” Then the coding breakthrough that catalyzed over the holiday break was the third step in that.

If Linus Torvalds is saying that AI coding is now better than he is, that’s never happened before.

Shawn Wang

That’s the benchmark, yeah.

That’s never happened before. Now we know that it’s going to sweep through coding. We know that if it’s going to work in coding, it’s going to work in everything else, right? That’s the hardest example in many ways, and everything else is going to be a derivative of that.

On top of that, we just got the agent breakthrough with OpenClaw, which is fantastic, amazing, and incredibly powerful. Then we got autoresearch—the self-improvement breakthrough. We’re now into the self-improvement breakthrough.

The way I think about it is that we’ve had 4 fundamental breakthroughs in functionality: LLMs, reasoning, agents, and now RSI. They’re all actually working. I’m jumping out of my shoes. This is it. This is the culmination of 80 years’ worth of work, and this is the time it’s becoming real.

Shawn Wang

Yeah. I’m completely convinced. I think the anxiety that people feel is that, during the transistor era, you had Moore’s law. It was, “All right, we understand why these things are getting better. We understand the physics of it.” With AI, it’s so jagged, and the jumps are so dramatic. Like you said, in 3 months you have this huge jump, and people are like, “Well, this can’t keep happening, right?” But then it keeps happening.

It’ll keep happening. How do you think about timelines and what’s worth building? We always have this question with guests: should you spend time building a harness for a model, versus the next model just doing it in one shot in the latest space? How does that inform how you think about the shape of the technology?

You talk about how it’s a new computing platform. If you have a computing platform that drastically changes every 6 months, it’s hard to build companies on top of it.

A couple of things. Moore’s law was what we now call a scaling law. For your younger viewers, Moore’s law was that chips either get twice as powerful or twice as cheap every 18 months. That was the 50-year trajectory of the computer industry.

By the way, that’s what took the mainframe computer from a $25 current-dollar thing into the phone in your pocket being a million times more powerful than that, for $500. That was a scaling law.

The key to any scaling law, including Moore’s law and the AI scaling laws, is that they’re not really laws, right? They’re predictions. But when they work, they become self-fulfilling predictions because they set a benchmark, and then the entire industry—all the smart people in the industry—work to make sure that it actually happens.

They motivate the breakthroughs that are required to keep things going. In chips, that was a 50-year run. It was amazing, and it’s still happening in some areas of chips.

I think the same thing is happening with the core scaling laws in AI. They’re not really laws, but they are predictions and motivating catalysts for the research work required to keep the curves going, as well as the investment dollars required to keep the curves going.

It’s going to be complicated, and it’s going to be variable. There are going to be walls that look like they’re fast approaching, and then engineers are going to get to work and figure out a way to punch through those walls. Obviously, that’s been happening a lot.

There are going to be times when it looks like the walls—or the laws—have petered out, and then they’re going to pick up again and surge. It appears that what’s happening to AI is that there are now multiple scaling laws. There are multiple areas of improvement, and I don’t know how many more there are to be discovered, but there are probably some more that we don’t know about yet.

For example, there’s probably some scaling law around world models and the acquisition of data at scale in the real world that we don’t fully understand yet. That one will probably kick in at some point here. There’s a bunch of really smart people working on that.

So, yeah, I think the expectation is that the scaling laws generally are going to continue.

Yeah, the pace of improvement will continue to move really fast. To your question about what to build: I'm a complete believer that the scaling laws are going to continue. I'm a complete believer that the capabilities are going to keep getting amazing, with leaps and bounds.

The part where I kind of part ways a little bit is with what I would describe as the AI purist. I would characterize them as people who are, in many ways, the smartest people in the field, but also people who spend their entire lives in a lab and have very little experience in the outside world. The nuance I would offer is that the outside world of 8 billion people, institutions, governments, companies, economic systems, and social systems is really complicated.

Eight billion people making collective decisions on planet Earth is not a simple process. You see this happening now: a bunch of the AI CEOs have this thing when they talk in public where they're just like, “Well, there's this obvious set of things that society needs to do.” And then they're like, “Society's not doing any of those things.” It's like, how can society not see whatever their theory is? How can society not see XYZ?

The answer is, well, society—number one, there's no single society. It's 8 billion people, and they all have a voice and a vote in how they react to change. Human reality is just really complicated and messy.

The specific answer to your question is, as usual, it depends. There's no question there are going to be companies—it's already happening—that think they're building value on top of the models and are then going to get blitzed by the next model. But I think there's no question that the process of adapting any technology into the real, messy world of humanity is going to be messy and complicated. It's not going to be simple and straightforward.

There are going to be a lot of companies, a lot of products, and, in fact, entire industries that are going to be built to help all this technology actually reach real people.

Shawn Wang

The amount of capital going into these companies—I mean, Dario talked about it on the Dwarkesh Podcast, and Dwarkesh was like, “Why don't you just buy 10 times more GPUs?” And Dario was like, “Because I'm going to go bankrupt if the model doesn't exactly hit the performance level.”

How do you think about that as a risk? You guys are investors in Inflection AI, Thinking Machines, and World Labs, and it seems like we're leveraging this scaling law at a pretty high rate. How comfortable, I guess, do you feel with the downside scenario? If things peter out, do you think you can restructure these build-outs and capital investments?

Yeah, I should start by saying I lived through the dot-com crash, and I can tell you stories for hours about the dot-com crash. It was horrible. No, it was awful. It was apocalyptic.

By the way, a lot of the dot-com crash was actually, at the time, a telecom crash. It was a bandwidth crash. The thing that actually crashed and wiped out all the money was the telecom companies—Global Crossing.

Shawn Wang

I'm from Singapore, and they laid so much cable over our oceans.

There was a scaling law in the dot-com era. The U.S. Commerce Department literally put out a report in 1996 saying that internet traffic was doubling every quarter. In 1995 and 1996, internet traffic actually did double every quarter.

That became the scaling law. What all these telecom entrepreneurs did was go out and raise money to build fiber, anticipating that the demand for bandwidth was going to keep doubling every quarter. Doubling every quarter, though, is like grains of rice on a chessboard. At some point, the numbers become extremely large.

What really happened was that the internet continuously kept growing, basically since inception. It's continuously grown; it's never shrunk. It's grown really fast compared to anything else in human history. But it wasn't doubling every quarter as of 1998 or 1999.

There was this gap between the expectation of what they thought the scaling law was and reality. That's actually what caused the dot-com crash: companies like Global Crossing way overbuilt fiber.

That was the beginning of the data center build. Then data centers overbuilt. You had fiber, telecom equipment—all the networking gear—and then the actual physical data centers. I think it was around $2 trillion that got wiped out.

Shawn Wang

Jesus.

By the way, the other subtlety was that the internet companies themselves never really had any debt, because tech companies generally don't run on debt. But the telecom companies—physical infrastructure companies—run on debt. Companies like Global Crossing didn't just raise a lot of equity; they also raised a lot of debt, so they were highly levered.

Then you have a highly levered thing where you're overbuilding capacity. Demand is growing, but not as fast as you hoped. Boom: bankrupt. It's like they say about the hotel industry: it's always the third owner of a hotel that makes money. It has to go bankrupt twice. You have to wash out all of the overoptimistic exuberance before it gets to a stable state, and then it makes money.

By the way, all of those data centers and all of the fiber are in use today, 25 years later. But it actually took 15 years—from 2000 to 2015—to fill up all the capacity.

The cautionary warning is that overbuilding can happen. You get into this thing where basically everybody who has any sort of institutional capital is like, “Wow, I don't know how to invest in these crazy software things. But for sure I can build data centers, and for sure I can buy GPUs and deploy compute grids,” and all these things.

If you're a pessimist, you could look at this and say, “Wow, this is really set up to basically replicate what we went through in 2000.” Obviously, that would be bad.

The counterargument—the one I agree with—is a couple of things. One is that the companies investing the money are blue-chip companies. Back then, Global Crossing was an entrepreneur's new venture. But the money being deployed now at scale is coming from Microsoft, Amazon, Google, Facebook, Nvidia, and, now, OpenAI and Anthropic, which are at a really serious size as companies with very serious revenue.

These are very large-scale companies with lots of cash and lots of debt capacity that they've never used. This is institutional in a way that it really wasn't at the time.

The other is that, at least for now, every dollar being put into anything that results in a running GPU is being turned into revenue right away. Everybody's starved for capacity. Everybody's starved for compute capacity, as well as all the associated things—memory, interconnect, data center space, and everything else.

Every dollar being put in the ground right now is turning into revenue. In fact, I think there's an interesting thing happening: because everybody's starved for capacity, the models we can actually use today are inferior versions of what we would have if not for the supply constraints.

Suppose there were a hypothetical universe in which GPUs were 10 times cheaper and 10 times more plentiful. The models would be much better, because you would just allocate a lot more money to training, build better models, and they would be better. We're actually getting the sandbagged version of the technology.

Shawn Wang

Yeah, no, everything we use is quantized because the labs have to keep the full versions.

But getting the good stuff is even more significant because, even if technical progress stops, once there's a much bigger build-out of GPU manufacturing capacity and memory—all the things that have to happen in the course of the next 5 or 10 years—even the current technology is going to get much better.

And then, as you know, there's just a million ways to use this stuff.

There’s just a million use cases for this. This isn’t just sending packets across a thing and hoping that people find something to do with it. This is: we apply intelligence in every domain of human activity, and then it works incredibly well.

Here’s what I know: in the next 3 or 4 years, basically everything is selling out. The entire supply chain is sold out or selling out, and we’re going to have a chronic supply shortage for years to come. There’s going to be a response from the market that’s going to result in an enormous flood of investment in new fab capacity and everything else needed to support it. At some point, the supply-chain constraints will unlock, at least to some degree, and that will be another accelerant to industry growth when it happens, because the products will get better and everything will get cheaper.

I know that’s going to happen. I know that the deployments—the actual use cases—are really compelling, and with reasoning and agents and so forth, I know they’re just going to get much, much better from here. I know the capabilities are really real and serious. I also know that technical progress is not going to stop; it is accelerating. The breakthroughs are tremendous. Even month over month, the breakthroughs are really dramatic.

If you were a cynic—and there are cynics—you can look at 2000 and find echoes, but I can’t imagine betting that this is somehow going to disappoint, at least for years to come. I think it would be essentially suicidal to make that bet.

Shawn Wang

Is that Michael Burry?

Uh-oh. That’s an interesting guy, huh? Let’s pick on one guy. He did come out with the Nvidia short, right?

And then—and you guys probably talked about this—the analysis now is that current models are getting better faster, at such a rate that if you’re running an Nvidia inference chip today that’s 3 years old, you’re making more money on it today than you did 3 years ago. The pace of improvement of the software is faster than the depreciation cycle of the chip.

My understanding is—these are rumors I’ve heard, or maybe it’s public—that Google is running very old TPUs very profitably. It actually turns out, as far as I can tell, that the opposite of the Burry thesis is true: he was 180 degrees wrong. The old Nvidia chips are getting more valuable, which is something that has literally never happened before.

That’s an expression of that ferocious pace of software progress and the ferocious pace of capability payoff that you’re getting on the other side of this. The idea of betting against that is an invitation to get your face ripped off.

What really hits me is modeling the lifespan of the H100s and H200s. Usually, they advise 4 to 7 years, and maybe you realistically cut it down to 2 to 3, but actually it’s going up and not down. That’s, I think, the dream. We’re finding utilization, and I think utilization solves all problems.

You can find use cases for even the poor stuff—even memory. We’re having a shortage, right? Even the shittier versions of memory that we do have, we’re finding use cases for. So that’s great.

Yeah.

How important is open-source AI and edge inference in a world in which you have 3 years of supply crunch? If you fast-forward 5 years, how do you think about inference in the data center versus at the edge?

To start, I think open source is very important for a bunch of reasons. I think edge inference is very important for a bunch of reasons. Practically speaking, if we’re going to have fundamental chip supply crunches for the next 3 years, you guys know that if you project demand over the next 3 years relative to supply, one of the dismaying predictions is what’s going to happen to the cost of inference in the core over the next 3 years. It may rise dramatically.

The big-model competition is subsidizing heavily right now. What will be the average person’s per-day, per-month token cost 3 years from now to do all the things they want to do? I have friends today who are paying $1,000 a day for OpenClaw cloud tokens to run OpenClaw. That’s $30,000 a month, and those friends have 1,000 more ideas for things they want OpenClaw to do.

You could imagine latent demand of up to—I don’t know—$5,000 or $10,000 a day in tokens for a fully deployed personal agent. Obviously, consumers can’t pay that, but it gives you a sense of the future scope of demand. Even if there’s a 10x improvement in price performance, that still only goes to $100 a day, which is still way beyond what people can pay. There’s just going to be ferocious demand.

The other interesting thing about the agent situation is that, up until now, a lot of the constraints have been GPU constraints. I think the agent situation now also translates into CPU constraints.

CPU and memory, yes.

CPU and memory, right? The entire chip ecosystem is just going to—

With network constraints, that will be the killer.

It’s all bottlenecked, potentially for years. Broadly, I think it’s actually possible that inference costs are going to keep coming down, but the rate of decline may level out here for a bit because of these supply constraints. At some point, maybe the labs stop subsidizing so much, and that will again be an issue. There’s just going to be so much more demand for inference than can be satisfied with a centralized model.

You guys know this, but the dramatic innovations that have happened in Apple Silicon to enable inference are quite amazing. The level of effort being put in—the open-source guys are putting incredible effort into this recurring pattern where the big model will never run on a PC, and then 6 months later, it runs on a PC. It’s amazing, and there are very smart people working on that.

There are also other motivators. There’s the question of how much trust the big centralized model providers are building in the market versus, at least in certain cases, people saying, “I’m not willing to just turn everything over.” There’s also straight-up price optimization. There are many uses of AI where you don’t need Einstein in the cloud; you just need a smart local model.

There are also performance issues. You’re going to want your doorknob to have an AI model in it so it can do access control. Obviously, everything with a chip is going to have an AI model in it, and a lot of those are going to be local. Then, by the way, there are also wearable devices. You don’t want to do a complete round trip; you want whatever your smart devices are to have super-low latency.

The question is: do we care who makes it? One of the biggest pieces of news this week was the collapse of AI2, the Allen Institute for AI, one of the actual American open-source model labs. I’m not that optimistic about American open source. You guys invested in Mistral, and Mistral is doing extremely well outside of China. That’s about it.

Yeah, we’ll see. Number 1, I do think we care who makes it. I would say this: the previous presidential administration wanted to kill it in the US. They wanted to drown it in the bathtub. At least we have a government now that actually wants it to happen.

Shawn Wang

And you’re on the council?

Yeah, I’m on the new PCAST. This administration, for whatever other political issues people have—which are many—has, I think, a very enlightened view, and in particular an enlightened view on AI and, in particular, on open-source AI.

And so, they're very supportive. My read is that the various Chinese companies have a very specific reason to do open source, which is that they don't fundamentally think they can sell commercial AI outside of China right now—or at least not specifically in the US—for a combination of reasons. And so, they view open-source AI as a bit of a loss leader against domestic paid services and ancillary products. They're very excited about it.

By the way, I think it's great that they're doing it. I think DeepSeek was a gift to the world. I think the great thing about open source is that its impact is felt in 2 ways. One is that you get the software for free. The other is that you get to learn how it works: you get the paper and the code.

For example, I thought this was amazing. OpenAI comes out with o1, and it's an amazing technical breakthrough. It's absolutely fantastic. But of course, they don't explain how it works in detail, and then they hide the reasoning traces. Then everybody's like, "Okay, this is great, but who's going to be able to replicate this? How are other people going to be able to do this? Is there secret sauce in there?"

Then R1 comes out, and it's just like, there's the code and there's the paper. Now the whole world knows how to do it. Then, 3 months later, every other AI model is adding reasoning. Even if the Chinese models themselves are not the models that get used, the education that's taken place throughout the rest of the world—the information diffusion—is incredibly powerful.

There are a bunch of American open-source AI model companies. Look, there's going to be tremendous competition among the primary model companies. There already is. Depending on how you count, there are 4 or 5 big model companies now that are neck and neck in different ways.

Obviously, both xAI and Meta, where I'm involved, have huge attempts to leapfrog on the way. Then you've got a whole fleet of startups and new companies, including a whole bunch that were backed and are trying to come up with different approaches.

How many mainline foundation-model companies are there in China at this point? It's probably 6.

The "6 Tigers" is what they call them. Qwen is questionable because of the changing leadership.

Right. Yeah. But that includes Moonshot.

Yes

Kimi, DeepSeek, Z.ai, Qwen, and 01.AI are in there, right? And then ByteDance. You see, ByteDance would be the next tier. They weren't as prominent.

They weren't, but now they are. At least Seedance is very inspiring, and presumably they have more stuff coming. Tencent probably has more stuff coming, and so forth.

Between the US and China, right now there are about a dozen primary foundation-model companies that are at scale, at some level of critical mass. It's not going to be a dozen in 3 years, right? These industries don't bear a dozen. There are going to be 3 or 4 big winners, or maybe 1 or 2 big winners. A whole bunch of those companies are going to have to figure out alternate strategies.

I think open source is one of those strategies. You could see that changing really fast. The question is, who's going to do open source? That could change really fast. That's a very dynamic thing, and it's very hard to predict what happens. I think it's very important.

Shawn Wang

Nvidia's doing a lot. You mentioned—

Well, I was going to say—

Shawn Wang

Well, exactly.

Then you've got Nvidia. Again, industry after industry, there's an old thing in business strategy called "commoditize the complement." If you're Jensen, it's just kind of obvious: of course you want to commoditize the software. To his enormous credit, he's putting enormous resources behind that. So maybe it's literally Nvidia, and I think that would be great.

Shawn Wang

Narrative violation: 2 European projects innovating. Bam. I'm mostly in my Europe conference soon, and I got both of them. They got us. They got us, Marc. They got us all.

Wait a minute. Where was Peter Steinberger when he did OpenClaw?

Yeah, yeah, yeah. He was in Vienna.

Shawn Wang

Oh, he was in Vienna. And then where is he now?

He's moving to SF.

Shawn Wang

Okay, there we go. And then, yeah, the Pi guy, right? The Pi guys are European.

Yeah, they're buddies in Austria. Mario's also there.

Shawn Wang

Right. And they haven't announced any sort of change yet, or have they?

No, they have a company there.

Shawn Wang

Okay, good. Yeah, good. Anyway, I think Pi and OpenClaw are very important software things, and I just wanted you to go off on what you think.

Yeah. I think the combination of the 2 of them is among the 10 most important software developments.

Shawn Wang

OpenClaw got all the attention, but talk about Pi.

Pi is kind of the architectural breakthrough. For those of us who are older, there was this whole thing that was very important in the world of software from basically 1970 through the creation of Linux. It still is very important, but we used to call it the Unix mindset.

There were all these different theories, all these different operating systems and mainframes, and then Windows and Mac and all these things. But behind it all was this idea of the Unix mindset.

In the old days, the operating system that made the computer actually work in the 1960s was this thing called OS/360, which IBM developed and which was supposed to basically run everything. It was this giant monolithic architecture in the sky, a giant castle of software. By the way, it worked really well, and IBM was very successful with it, but it was this huge castle in the sky and almost unapproachable.

You had to be inside IBM or very close to IBM, and you had to really understand every aspect of how the system worked. Then the Unix guys, originally out of AT&T and then out of Berkeley, came out and said, "No, let's have a completely different architecture."

The way the architecture is going to work is that we're going to have a prompt and a shell. All the functionality is going to be in the form of these discrete modules, and then you're going to be able to chain the modules together. It's almost like the operating system itself is going to be a programming language.

That led to the centrality of the shell, then to chaining together Unix tools, and then to the emergence of scripting languages like Perl, where you could easily do this. The shells got more sophisticated.

That worked. That was the world I grew up in. I was a Unix guy from, call it, 1988 all the way through my work, and it worked really well. In the background, normal people didn't necessarily need to know about it, but if you were doing system architecture or application development, you knew all about it.

It's been in the background ever since. Your Mac still has a Unix shell in there, and your iPhone still has a Unix shell buried in there somewhere. They're kind of in there. The Windows shell is kind of a weird derivative of that, but the internet runs on Unix.

Smartphones—both iOS and Android—are Unix derivatives. Unix did end up winning. Anyway, we just started taking that for granted.

Basically, the way I think about what happened with Pi and then with OpenClaw is that those guys figured out something. I would say the great breakthroughs are obvious in retrospect, right? Which is the best kind. They weren't obvious at the time, or somebody else would have done them already. So there is a real conceptual leap, but then you look at it backwards and you're just like, "Oh, of course."

Shawn Wang

To me, those are always the best breakthroughs.

Well, actually, language models themselves are like that. It’s just, “Oh, next-token completion. Of course.”

Shawn Wang

Yeah, what other objective mattered?

Yeah, exactly. [laughter] But she was even saying it wasn’t obvious until somebody actually did it, right? And so the conceptual breakthrough is real, deep, powerful, and very important.

The way I think about Pi and OpenClaw is that it’s basically marrying the language-model mindset to the Unix shell-prompt mindset. What is an agent? As you know, many smart people have been trying to figure out what an agent is for decades. They’ve had many architectures to build agents, and the whole thing.

It turns out that what we now know is an agent is the following: it’s a language model, and then above that, it’s a Bash shell. So it’s a Unix shell, and the agent has access to the shell, hopefully in a sandbox.

Then there’s a file system, and the state is stored in files. There’s the Markdown format for the files themselves, and then there’s basically what in Unix is called a cron job. There’s a loop and a heartbeat, and the thing basically wakes up. So it’s LLM plus shell plus file system plus Markdown plus cron, and it turns out that’s an agent.

Every part of that other than the model is something we already completely know and understand. In fact, the latent power of the Unix shell is extraordinary. There’s enormous latent power in the shell, enormous numbers of Unix commands, and enormous numbers of command-line interfaces into all kinds of things already.

Your entire computer, to start with, runs on a shell. If you’re running a Mac or a phone, your computer is already running on a shell. The full power of your computer is available at the command-line level, and it turns out it’s really easy to expose other functions as a command-line interface.

This whole idea that we need MCP and these fancy protocols or whatever—it’s like, no, we don’t. We just need a command-line thing.

That’s the architecture. Then it turns out, what is your agent? Your agent is a bunch of files stored in a file system. The thing that completely blew my mind when I wrapped my head around it as a result of this was, okay, this means your agent is now actually independent of the model that it’s running on.

You can swap out a different LLM underneath your agent. Your agent will change personality somewhat because the model is different, but all of the state stored in the files will be retained.

Shawn Wang

Different instruction set, but you just compile it.

Right, exactly. It’s like swapping out a chip and recompiling, but it’s still your agent with all of its memories and all of its capabilities. By the way, you can also swap out the shell, so you can move it to a different execution environment that is also a Bash shell.

You can also switch out the file system. You can swap out the heartbeat, the cron framework, the loop—the agent framework itself. So your agent, at the end of the day, is basically just its files.

Shawn Wang

And then there is, of course, the purple Claw.

Yeah, it’s basically just the files. [laughter]

As a consequence of that, it turns out there are a couple of important things about the agent itself. One is that it can migrate itself. You can instruct your agent, “Migrate yourself to a different runtime environment. Migrate yourself to a different file system. Swap out the language model.” Your agent will do all that stuff for you.

Then there’s the final thing, which is just amazing: the agent actually has full introspection. It knows about its own files, and it can rewrite its own files. There’s basically no widely deployed software system in history where the thing that you’re using actually has full introspective knowledge of how it itself works and is able to modify itself like that. There have been toy systems that have had that, but there’s never been a widely deployed system with that capability.

That leads you to the capability that completely blew my mind when I wrapped my head around it, which is that you can tell the agent to add new functions and features to itself, and it can do that. “Extend yourself. Give yourself a new capability.”

Literally, you run into somebody at a party, and they’re like, “Oh, I have my OpenClaw connect to my Eight Sleep bed, and it gives me better advice on sleep.” You go home at night, and you tell your Claw—or, if they’re at the party, you tell your Claw—“Add this capability to yourself.”

Your Claw will say, “Oh, okay, no problem.” It’ll go out on the internet and figure out whatever it needs, and then it’ll go out to Claude Code or whatever, write whatever it needs, and the next thing you know, it has this new capability. You don’t even have to do anything other than tell it that you want it to upgrade itself.

The combination of all this is just massive and incredible. If I were 18, this is 100% what I would be spending all of my time on. This is such an incredible conceptual breakthrough.

Again, people are going to look at it—and they already have this response—and say, “Oh, well, where’s the breakthrough? All of these components were already known before.” But this is the key: by using all these components that were known before, you get all of the underlying capability that’s buried in there.

For example, computer use all of a sudden just becomes trivial. Of course it’s going to be able to use a computer—it has full access to the shell. Then you give it access to a browser, and you’ve got the computer and the browser, and off it goes. You’ve got all the abilities of the browser, also.

The capability unlock here is profound. My friends who were deepest into this are having their Claw do literally 1,000 things in their lives. They have new ideas every day. They’re constantly throwing new challenges at the thing.

It’s early, and these are prototypes. As you guys know, there are security issues, and there’s a bunch of stuff to be ironed out. But the unlock of capability is incredible.

I have absolutely no doubt that everybody in the world is going to have at least an agent like this, if not an entire family of agents. We’re going to be living in a world where I think it’s almost inevitable that this is how people are going to use computers.

Shawn Wang

I was going to say, for someone who’s deeply familiar with social networks, the next step is your Claw talking to my Claw: posting on Claw Facebook, posting their jobs on Claw LinkedIn, and Claws posting their tweets on Claw XAI or whatever. [laughter]

I do think that’s how we get into some danger in terms of alignment and whether or not we want these things to run. Do you guys know about Rent-A-Human.com?

Shawn Wang

Yeah, Rent-A—yeah. I mean, it’s Fiverr, it’s TaskRabbit, it’s—

Of course. Mechanical Turk.

Shawn Wang

Yeah, but flipped, right? The agent hiring the people. [laughter] It’s obviously going to happen. I’m curious if you have any thoughts on the engineering side.

When you built the browser, the internet was just a bunch of mostly plain-text files plus some images. Today, every website and app is so complex, and somehow the browser kept evolving to fit that in.

Are there any design choices that were made early in the browser, the internet, or the protocols that you’re seeing agents similarly say, “Hey, this thing just isn’t going to work for this type of new compute, and we should rip it out right now?”

There were a whole bunch, but I’ll give you a couple. One is—and, to be clear, this was totally different; we didn’t have the capabilities we have today, and we didn’t have language models underneath this—but we did have the idea that human readability actually mattered a great deal.

Specifically, in those days, it wasn’t so much English language, but there was a design decision to be made between binary protocols and text protocols. Basically, every old-school systems architect who had grown up between the 1960s and 1990s said, “What do you know about the internet? It’s starved for bandwidth.”

You just have these very narrow straws. When we did the work on Mosaic, people who had the internet at home had a 14-kilobit modem, so you were trying to hyper-optimize every bit of data that travels over the network. And so, obviously, if you're going to design a protocol like HTTP, you're going to want it to be a highly compressed binary protocol for maximum efficiency. You're going to want it to be a single connection that persists, and the last thing you're going to want to do is bring up and tear down new connections.

And you definitely are not going to want a text protocol. And so, of course, we said, no, we actually want to go completely the other direction. Obviously, we only want text protocols. By the way, same thing in HTML itself: We want HTML to be relatively verbose. We want the tags to actually be human-readable. We want to use the—

Shawn Wang

The most inefficient things possible.

Yeah, we want to do the inefficient things.

Shawn Wang

You're the original token maxer.

Yeah, exactly. Basically, this was the conscious thing, which says: Assume a future of infinite bandwidth. Build for that. And what it was is, it was a bet that if the system—the latent capabilities of the system—were powerful enough, and that was obvious enough to people, that would create the demand for the bandwidth that would cause the supply of bandwidth to get built, which would actually make the whole thing work.

And specifically, what we wanted was everything to be human-readable because, at the engineering level, we wanted people to be able to read the protocol coming over the wire and be able to understand it with their bare eyes, without having to disassemble it or whatever. You have to convert it out of binary.

HTTP and everything else were always text protocols, and the same thing with HTML. In many ways, some people say that the key breakthrough in the browser was the View Source option. Every webpage you go to, you could view source, which means you could see how it worked, which means you could teach yourself how to build new webpages. There was that.

So human readability—and, again, human readability in those days still meant technical specs. Now it means English language—but there's an incredible latent power in giving everybody who uses the system the option to drop down and actually understand and see how it's working. That worked really well for the web, and I think it's working really well for AI.

That was one. What was the other? A big part of the idea of web servers was actually to surface the underlying latent capability of the operating system, and also to surface the underlying latent capability of the database. Because what is a web server, fundamentally, architecturally? It's the operating system.

It's running on top of an OS, so it's the OS's ability to manage the file system and do everything else that you want to do—process everything. And, of course, a lot of websites are front ends to databases. You wanted to unleash the underlying latent power of whether it was an Oracle database, Postgres, or whatever it was. A lot of the function of the web server was just to bridge from that internet connection coming in to unlock the underlying power of the OS and the database.

And again, people looked at it at the time and they were like, “Does this really matter? Is this important?” Because we've had databases forever, and we've always had user interfaces for databases, and this is just another user interface for a database. It's like, okay, yeah, fair enough.

But on the other side of that is: This is now a much better interface to databases, and one that 8 billion people are going to use, and that's going to be far easier to use and far more flexible. And you're not just going to have old databases. Now you have a system where people can actually understand why they want to build a million times more database apps than they have in the past. And then the number of databases in the world exploded.

So again, this goes to this thing of building in layers. Some of the smartest people in industry look at any new challenge and they're like, “Okay, I need to build a new kind of application. So the first thing I need to do is build a new programming language.” Right? “And then the next thing I need to do is build a new operating system.” And then the next thing I need to do is build a new chip. And they kind of want to reinvent everything.

I've always had—maybe it's just a pragmatic mentality or something, or maybe an engineering-over-science mentality—but it's more like, no, you have all of this latent power in the existing systems. You don't want to be held back by their constraints, but what you want to do is liberate that power and open it up.

Shawn Wang

Programming languages are another good thing. We have Bret Taylor on the podcast, and we were talking about Rust. Rust is memory-safe by default. So why are we teaching the model not to write memory-unsafe code? Just use Rust and then you get it for free.

How much time do you think should be spent recreating some of these things instead of taking them for granted? It'll be like, “Oh, okay, Python is kind of slow.” It's like, yeah, as imperfect as they are, they are the lingua franca.

I think this is going to change a lot because I don't think the models care what language they program in. I think they're going to be good at programming in every language, and I think they're going to be good at translating from any language to any other language.

Okay, so this gets into the coding side of things. I think we're going through a really fundamental change. And look, I grew up hand-coding. Everything I did was written in C. I wasn't even using C++ or Java or any of this stuff. Everything I ever did, I was managing my own memory at the level of C. And I'm still from the generation that knew assembly language, so I could drop down and do things right on the chip.

We've all always lived in a world in which software is this precious thing that you have to think about very carefully. It's really hard to generate good software, and there's only a small number of people who can do it. You have to be very jealous in terms of thinking about how you allocate: What are your engineers working on? How many good engineers do you actually have? How much software can they write? How much software can human beings maintain?

I think all those assumptions are being shot right out the window right now. I think those days are just over, and I think the new world is that high-quality software is just infinitely available. If you need new software to do XYZ, you're just going to wave your hand and you're going to get it. And if you don't like the language it's written in, you just tell the thing, “All right, right now I want the Rust version.”

By the way, computer security is about to go through the most dramatic change ever. Number 1, every single latent security bug is about to be exposed.

Shawn Wang

Right. So we're set up here for a computer-security apocalypse for a while.

But on the other side of it, now we have coding agents that can go in and actually fix all the security bugs. And so how are you going to secure software in the future? You're going to tell the bot to secure it, and it's going to go through and fix it all.

This thing that was an incredibly scarce resource—high-quality software—is just going to become a completely fungible thing that you're just going to have as much of as you want.

Shawn Wang

Right.

And that has tons and tons of consequences. In some sense, the answer to the question you posed, I think, is somewhat simple or straightforward: If you want all your software in Rust, you just tell the bot you want all your software in Rust.

The things that used to be hard, or even seemed like an insurmountable mountain to get through, all of a sudden become very easy. I think Bret had a theory that there would be a more optimal language for LLMs. And so the contention is there isn't. Just don't bother. Whatever humans already use, LLMs are perfectly capable of porting.

I think we're pretty close to being—I don't know if this would work today—able to ask the AI what its optimal language would be and let it write it and let it design it.

Shawn Wang

True. Okay, here's the question: Are you even going to have programming languages in the future, or is the AI just going to be emitting binaries?

Let’s assume for a moment that humans aren’t coding anymore. Let’s assume it’s all bots. What levels of intermediate abstraction do the bots even need?

Shawn Wang

Yeah. Or are they just coding binary directly? Did you see there’s actually an experiment—there’s somebody who just did this thing where they have a language model that actually emits model weights for a new language model?

Right. And so will the bots be—

Shawn Wang

Just predict the weights.

Well, yeah. Will the bots literally be emitting not just coding binaries, but weights for new models directly? Conceptually, there’s no reason why they can’t do both of those things. Architecturally, both of those things seem completely possible.

Shawn Wang

Very inefficient. You’re basically a simulation of a simulation in a simulation inside of the weights.

Yeah, very inefficient. But look, LLMs are already incredibly inefficient. Ask Claude, “2 + 2 = 4,” right? It’s billions and billions of times more inefficient than using your pocket calculator, but the payoff is so great with the general capability.

So, anyway, I kind of think in 10 years—I’m not sure there will even be a salient concept of a programming language in the way that we understand it today. In fact, what we may be doing more and more is a form of interpretability: we’re trying to understand why the bots have decided to structure code in the way that they have.

Shawn Wang

If you play it through, you don’t need browsers then. That’s the death of the browser.

Well, I would take it a step further: you may not need user interfaces. So who is going to use software in the future?

Shawn Wang

Other bots.

Other bots. Yeah, and so—

Shawn Wang

They need to, I don’t know, pipe information in and out. Do we?

Really?

Shawn Wang

Well, what are you going to do then?

Are you sure? You’re just going to log off and touch grass?

Shawn Wang

Whatever you want, exactly.

Isn’t that better? I want software to do stuff for me. Isn’t that better? I don’t know. There are arguments here.

It was not that long ago that 99% of humanity was behind a plow.

Shawn Wang

Right.

What are people going to do if they’re not plowing fields all day to grow food? It just turns out there are much better ways for people to spend time than plowing fields.

Shawn Wang

Yeah, do some scrolling.

Exactly, or talking to their friends. Look, I’m not an absolutist, and I’m not a utopian. To be clear, I have an 11-year-old, and he’s learning how to code. I think it’s still a really good idea to learn how to code and so forth.

But if you project forward, you just have to think forward to a world in which it’s, “Okay, I’m just going to tell the thing what I need, and it’s going to do it.” Then it’s going to do it in whatever way is most optimal for it to do it. I can tell it to do it non-optimally—for example, if I tell it to do it in Java or in Rust or whatever, it’ll do it, I’m sure—but if I’m just going to tell it to do it, it’s going to do it in whatever way is optimal.

And if I need to understand how it works, I’m going to ask it to explain to me how it works. So it’s going to be doing its own interpretability. It’s going to be the engine of interpretability to explain itself.

I’m just not convinced that, in that world, you have these historical abstractions. The goals of the abstractions will be whatever the bots need with the user.

Shawn Wang

I’m curious, though: if that’s true, shouldn’t the model providers be building some internal language representation that they can do extreme reinforcement learning and reward modeling around? Today, they’re tied to TypeScript and Python because the users need to write in those languages, versus having their own thing internally that they don’t need to teach to anybody. They just need to teach it to their model.

I think that’s how you get maybe the differentiation between the models. Going back to the Pi and OpenClaw thing, it’s like, “Oh, I built all the software using the OpenAI model, and now I switch to the Anthropic model, but the Anthropic model doesn’t understand the thing.” It feels like there still needs to be some abstraction. But maybe not. Maybe that’s the lock-in that the model providers want to have.

I don’t know. I’m not even sure that’s lock-in, though, because why can’t the second model just learn what the first model has done?

Shawn Wang

Exactly. Okay, give me an example.

As you know, models can now reverse-engineer software binaries. There’s this whole thing now where people are reverse-engineering Nintendo game binaries. You’ve seen a bunch of reports like this where somebody has a favorite game from the 1980s, and the source code is long dead, but they have a binary burned into a chip or something. Now they’re reverse-engineering it to get a version that runs on their Mac, right?

If you’re reversing x86 binaries, why can’t you reverse-engineer whatever they create?

Shawn Wang

Yeah. And because we’re on a Unix-based system, it has to be reversible because it needs to run on the target.

Yeah, basically. And so I just think it’s this thing where—by the way, everything we’re describing is something that human beings, in theory, could have done before, but with enormous—

Shawn Wang

Right, yeah.

But it was always cost- and labor-prohibitive. What I learned about reverse-engineering is that human beings can reverse-engineer binaries. It’s just that for any complex binary, you need like 1,000 years to do it. But now, with the model, you don’t.

All of a sudden, you get these things. Another way to think about it is that so many human-built systems exist to compensate for human limitations. If you don’t have the human limitations anymore, then all of a sudden you won’t have abstractions, but you’ll have a different kind of abstraction.

Shawn Wang

I have 2 topics to bring us to a close, and you can pick whichever one. Just talking about protocols: was it you or someone else—I forget—who said that the biggest mistake we didn’t figure out in the early days was payments? Was that you?

Yes, it was: 402 Payment Required. We have a chance now.

I don’t think we’re going to figure it out. I don’t know. What’s your take?

We will. No, now I think it’s going to happen for sure. There are 2 reasons it’s going to happen for sure.

One is that we actually have internet-native money now in the form of crypto stablecoins and crypto. This is, I think, the grand unification of AI and crypto that’s about to happen now. I think AI is the crypto killer app. I think it’s where this is really going to come out.

The other is that it’s now obvious. Obviously, AI agents are going to need money. It’s already happening, right? If you’ve got an OpenClaw and you want it to buy things for you, you have to give it money in some form.

I would say the adoption is probably 0.1%, if that.

Oh, today?

Yeah, today. But think forward. Where is it going?

The ultimate principle of everything—and everything that I think we do—is the William Gibson quote: “The future is already here; it just isn’t distributed.” It isn’t distributed yet.

My friends who are the most aggressive users of OpenClaw have just given their Claws bank accounts and credit cards. Not only have they done it, it’s obvious that they needed to do it, because it’s obvious that they needed to be able to spend money on their behalf.

It’s just completely obvious. The number of people who have done that today, to your point, is probably 5,000 or something. But it’ll grow. That’s how these things start.

Actually, since you keep mentioning OpenClaw—and by the way, OpenClaw, if you don’t give it a bank account, is just going to break into your bank account. It’s going to break into your bank account anyway and take your money, so you might as well do it.

By the way, I really love the phenomenon. I love the YOLO. I’m not doing it myself, to be clear, but I love the people who are just like, “What is it? Skip permissions?”

Because at Facebook, they have this culture of naming the thing “dangerous,” so that you’re aware, when you enable the flag, that you’re opting into a dangerous thing. They brought it into OpenClaw, and of course that makes it enticing. Sam runs Codex with skip permissions on his laptop.

Yes, 100%.

And so I think the way to actually see the future is to find the people who are doing that. Log everything, just watch the logs.

Let’s actually find out what the thing can do. The way to find out what the thing can do is just let it try everything. Let it unlock everything. That’s how you’re going to find all the good stuff it can do.

By the way, that's also how you're going to find all the flaws. I think the people who turn that on for bots are murderers to the progress of human civilization. I feel very bad for their descendants, whose bank accounts are going to get looted by their bots in the first 20 minutes, but I think the contribution that they're making to the future of our species is amazing.

It's like gentleman science, you know.

Yes, yes. It's a Ben Franklin out trying to get lightning to strike his kite and see if he gets electrocuted. It's a Jonas Salk with the polio vaccine, injecting it. So, yes, I think we should have glory. We should have flags and monuments to the people who just let OpenClaw run their lives.

More anecdotes. What are the craziest or most interesting things that people listening to this should go home and do?

I mean, the extreme thing is just straight YOLO: turn your life over to it.

But that's a general capability. Is there a specific story that was like, “Wow,” and everyone in the group chat just lit up?

There are already tons of health-related examples. The health dashboard stuff is absolutely amazing. The number of stories—I'm trying not to violate people's personal anonymity—but one of the things OpenClaw is really good at is hacking into all the stuff on your LAN. It's really good at it. The Internet of Things is basically the Internet of Super-Insecure-but-Great-and-Discoverable.

OpenClaw is happy to scan your network, identify all the things, and then my friends who are most aggressive at this are having OpenClaw take over everything in their house. It takes over their security cameras, their access-control systems, their webcams. I have a friend whose Claw watches him sleep. He put a webcam in his bedroom, put the Claw on a loop, and told it, “Watch me sleep.”

I've seen the transcripts, and it's literally like: “Joe's asleep. This is good. It's good that Joe's asleep because I have his health data, and I know that he hasn't been getting enough sleep, so it's really good that he's getting sleep. I really hope he gets his full 5 hours of REM sleep. Joe's moving. Joe might be waking up. If Joe wakes up now, he's going to ruin his sleep cycle. Oh, okay, it's okay. Joe just rolled over. Okay, he's gone back to bed. Good. All right, I can relax. This is fine.”

He's monitoring the situation.

Monitoring the situation and being a bot. It's just very focused. This is its reason for existence: to watch Joe sleep. I was telling my friend who did this, “On the one hand, this is weird and creepy, and maybe this is taking over my life. On the other hand, what if I had a heart attack in the middle of the night? It would literally freak out and call 911. There's no question this thing would figure out how to alert medical authorities and probably summon SWAT teams and do whatever would be required to save my life.” So, yeah, that's happening.

What else? There's this company, Unitree, that makes robot dogs, and I actually have one at home. The Chinese companies are so aggressive at adopting new technology, but they don't always take the time to really package it.

Package it and maybe think it all the way through.

The Unitree dog I have has an old, non-LLM control system, which, by the way, is not very good. It markets well, but in practice it's not that good. It has trouble with stairs and so forth, so it's not quite what it should be. But then the language-model thing comes out, and so they add LLM capability and a voice mode to it. That LLM capability is not at all connected to the control system.

So you've got this schizophrenic dog that's a complete idiot when it comes to climbing the stairs, but will happily teach you quantum mechanics, with a plummy English accent. It's absolutely amazing.

Jagged intelligence. Talk about jagged.

And then, obviously, what's going to happen in the future is that they're going to connect the two together.

They'll do it.

But right now, it's not that useful. I have a friend who has one of these, and his Claw basically hacked in and rewrote the code—wrote new firmware for the Unitree robot.

You can do that before and after with the motion?

He said it's completely different. He said it's a complete transformation. Whenever there's an issue with the thing now, the Claw just rewrites the code. It kind of goes to your point here.

All of a sudden, this is why we're going to think about AI coding. AI coding is not just writing new apps. It's also going in and rewriting all the old stuff that should have worked but never worked. I think the Internet of Things is basically over. All these devices in your house that have been marginal, or basically dumb, might all of a sudden get really smart.

Now, with the smart home, you have to decide if you want this. There are horror movies in which this is the premise.

Yeah, but this is the first time I can say with confidence that I now know how you could actually have a smart home with 30 different kinds of things with chips and Internet access, where it all makes sense, works together, and is coherent—and unlock all of that without a human being having to do all that work.

I'm waiting for, “Sorry, Marc. I can't let you open that fridge door.”

Exactly.

Because you're not supposed to eat right now.

Yes. I have every shred of health information, and I know what you think you're doing. I know you think you can do this, but are you really sure? You told me last night that you really don't want me to let you do this, so I'm sorry, but the fridge door is locked.

Open the fridge doors.

Exactly. And, by the way, I know you're supposed to be studying for a test. So why don't you go study? When you can pass the test, I will open the fridge door for you.

Final photo call, and then we can wrap up: proof of human, right? That's the last piece we have to figure out.

Yeah. I would say there are 2 massive asymmetries in the world right now where we've known these asymmetries exist, and we societally have been unwilling to grapple with them. I think they're both tipping right now, and they're the same thing: the virtual-world version and the physical-world version.

The virtual-world version is the bot problem. The Internet is awash in bots. The Internet is awash in fake people. It has been forever. A lot of that has to do with a lack of money. My spicy take is that these 2 things are the same thing, and corporations are people, too.

Interesting. So, a bank account is proof of human?

Yeah, until you give the bots bank accounts.

Exactly.

Yeah, so there's that. But look, every social-media user knows this: the bot problem is a big problem. The bot problem has been a big problem forever. It's a huge problem, and it's never really been confronted directly.

The physical-world version of this is the drone problem. We've known for 20 years now that the asymmetric threat, both in actual military conflict and in security on the home front, is the cheap attack drone—the cheap suicide drone with a bomb.

We've known that forever. It's very disconcerting how every office complex in the world is unprotected from drone attacks. Every stadium, every school, every prison is unprotected.

Shawn Wang

Sure. We've known that. We've never done anything about it. What are we going to do about it? One possibility is just to leave them unprotected forever and live in a world of asymmetric terrorism forever.

The other is to take the problem seriously and figure out the set of techniques and technologies required to be able to deal with that, whether those are lasers, jammers, early warning systems, or personal force fields.

The virtual version of the problem—what we need, quite literally, is proof of human. The reason is because you’re not going to have proof of bot, especially now that the bots are too good. The bots can pass the Turing test. And if the bots can pass the Turing test, then you can’t screen for bots.

You can’t have proof of not a bot. But what you can have is proof of human. You can have cryptographically validated proof that this is definitely a person. And then you can have cryptographically validated proof that this is definitely something that a person said. This video is real.

Shawn Wang

Kinetic personal force fields—for Dune?

For personal force fields. Exactly. And in both cases, these are economic asymmetries. These are economic asymmetries, right? Because it’s really cheap to field a bot, but it’s very hard to tell if something’s a bot. It’s very cheap to field a drone, but it’s very expensive to defend against a drone.

Shawn Wang

Just to double-click on that, do you think Alex Blania with World has got it, or is there an alternative?

I think many people will try. We’re one of the key participants in the World project.

Shawn Wang

I didn’t know that.

Yeah, so we’re partisans. But, yeah, we think World is exactly correct.

Shawn Wang

Okay.

The reason is it has to be proof of human. It has to be proof of human because you can’t do proof of not-bot. You have to do proof of human. To do proof of human, you need biological validation. You need to start with: this was actually a person, right? Because otherwise you have bots signing up as fake people.

So you have to have something—you have to have a biometric. And then you have to have cryptographic validation, and then the ability to do the lookup. And then, by the way, the other thing you need is selective disclosure.

You need to be able to do proof of human without revealing all the underlying information. By the way, another thing you’re going to need is proof of age, right? Because there are all these laws in all these different countries now around needing to be 13, 16, or 18, or whatever, to do different things. And so you’re going to need to validate proof of age to be able to legally operate, right? And so that’s coming. And then you’re going to want proof of credit score and proof of 100 other—

Shawn Wang

That’s a tricky one.

It is a tricky one, but you’re going to need it. There’s no reason, if somebody’s checking on your credit, that somebody should need to know your name in order to find out whether you’re creditworthy, right?

Shawn Wang

I see. Independently verifiable pieces of information.

Pieces of information. So, like, selectively disclosed. This is the answer to the privacy problem writ large, which is: I only need to prove what I need to prove at that moment.

Shawn Wang

So you’re going to need that, and I think their architecture makes sense.

So that needs to get solved. I think language models have tipped the balance. The bots are now too good, and so they’re undetectable. As a consequence, we now need to confront that problem directly. And then, like I said, the other problem is we need to actually confront the drone problem.

Shawn Wang

The Ukraine conflict has really unlocked a lot of thinking on that. And now the Iran situation is also unlocking that. I think there’s going to be just this incredible explosion of both drone and counter-drone. Our drones are better than their guns. Let’s also keep it that way.

I think we can sneak in one more question. I’m trying to tie together a lot of things that you’ve said over the years. At the Milken Institute debate with Thiel, which was amazing, you talked about the lag between a new technology and its GDP impact. The other idea you talked about is bourgeois capitalism and how this managerial class was needed because of this complexity. And I think if you bring AI into the fold, you have much higher leverage of people. So if you have Musk Industries and you give Elon AGI, you can run a lot more things at once.

That’s right.

Shawn Wang

And then you have the social contract. I know you retweeted a clip of Sam Altman saying, “We’re rethinking the whole thing,” and you were like, “Absolutely not.” I was at that event with Sam last night, and he actually said that in the last couple of weeks, he felt like people were finally taking that seriously. So I’m just curious how you’re seeing the structure of organizations changing, especially when you invest in early-stage companies, and how the impact on work structure and all of that is playing out.

Yeah. There’s a whole bunch of time we could spend—I don’t know, by the way, we’d be happy to spend more time, but we could spend more time on all that. So, just for people who haven’t followed this: this term “managerial” comes from this thinker in the 20th century, James Burnham, who is one of the great 20th-century political thinkers, societal thinkers.

He was writing in the 1940s and 1950s, and he said that the whole history of capitalism until that point had been in 2 phases. Number 1 had been what he called bourgeois capitalism. Think about it as the name on the door, like Ford Motor Company, because Henry Ford runs the company. It’s a dictatorial model, and Henry Ford just tells everybody what to do.

He said the problem with bourgeois capitalism is that it doesn’t scale, because Henry Ford can only tell so many people to do so many things, and then he runs out of time in the day. The second phase of capitalism was what he called managerial capitalism, which was the creation of a professional class of managers. They’re trained not to be car experts or experts in any particular field, but to be experts in management.

That led to the importance of Harvard Business School, management consulting firms, and all these things. Then you look at every big company today, and most of the executives at most of the Fortune 500 companies are not domain experts in whatever the company does. They’re certainly not the founders of those companies, but they’re professional managers.

In the course of their careers, they’ll probably manage many different kinds of businesses. They’ll rotate around, and they might work in health care for a while, then work in financial services, and then go work in something else or come work in tech.

What Burnham said is that this transition is absolutely required, because the problem with bourgeois capitalism is that it doesn’t scale. Henry Ford doesn’t scale. So if you’re going to run capitalist enterprises that are going to have millions to billions of customers, you’re going to need them to operate at a level of scale and complexity that’s going to require this professional management class.

He said, “Look, the professional management class has its downsides. They’re not necessarily experts at doing the thing. They’re not as inventive. They’re not going to create the next breakthrough thing.” But whether you think that’s good or bad or whatever, it’s what’s going to be required. And basically, that’s what happened, right?

He wrote that book originally around 1940. Over the course of the next 50 years, managerialism basically took over everything. What I’m describing is basically how all big companies run, how all governments run, how all large-scale nonprofits run, and how kind of everything runs.

Basically, what venture capital does is we’re a rump protest movement against that, to try to find the next Henry Ford—or, say, Elon Musk, the next Steve Jobs, the next Bill Gates, or the next Mark Zuckerberg. We start these companies in the old model, right? We start them out as in the Henry Ford model.

We start them out with a founder or a founder with colleagues, but there’s a founder CEO. And then we basically bet that the startup is going to be able to do things—specifically, innovate in ways that the big incumbents in that industry are not going to be able to do.

It’s a bet that by relighting the sort of name-on-the-door kind of thing with a king-monarchical political structure, they’re going to be able to innovate in a way that the incumbent is not going to be able to, because the incumbent is being run by managers, right?

And, of course, venture being what it is, sometimes that works and sometimes it doesn’t, but we’re constantly doing that.

But I’ve always viewed it my entire life as like we’re raging against the dying of the light. We’re constantly trying to fight off managerialism just swamping everything and everything getting boring and gray and dumb and old. Right? We’re trying to keep some level of energy and vitality in the system.

AI is the thing that would lead you to think, “Wow, maybe there’s a third model.” Maybe it’s a combination of the two. Maybe the new Henry Ford or the new Elon or the new Steve Jobs plus AI is the best of both. It’s the spark of genius of the name-on-the-door model, the Henry Ford model, but then you give that person AI superpowers to do all the managerial stuff and let the boss do all the managerial stuff. That may be the actual secret formula.

And we’ve never even known that we wanted this because we never even thought it was a possibility. But you know this: What are these bots really good at? They’re really good at doing paperwork. They’re really good at filling out forms, writing reports, reading them, and doing all the managerial work. They’re amazing at it. So I think the answer very well might be to get the best of both worlds by doing this.

And then the challenge is going to be twofold. It’s going to be for the innovators to really figure out how to leverage AI to actually do this. And then the other challenge is going to be for the incumbents that are managerial to figure out, “Okay, what does that mean?” Now they’re going to be facing a different kind of insurgent competitor that has a different set of capabilities than they’re used to.

This is really going to force a lot of big companies to figure out innovation. I say, figure out innovation or die trying.

Do you feel like that structure accelerates the impact on actual GDP and the economy? If you could space it out, the growth is so fast. Instead of having these companies peter out in growth and impact, they can keep going, if not accelerating.

That’s for sure the hope. The challenge—and, look, the AI utopian view is, of course, that’s going to be the future of the economy, and it’s going to grow 10x and 100x and 1,000x, and we’re entering this regime of much higher economic growth forever, with a consumer cornucopia of everything, and it’s going to be great. I hope that’s true. That’s the current utopian vision. I hope that’s true.

The problem goes back again: The real world is really messy. I’ll give you an example of how the real world is really messy. It requires 900 hours of professional certification training to become a hairdresser in the state of California. Something like 35% of the economy—you have to get some sort of professional certification to do the job. Which is to say that the professions are all cartels, right?

You have to get licensed as a doctor. You have to get licensed as a lawyer. You have to get licensed as a— you have to get into a union. By the way, to work for the government, you have civil service protections and public-sector unions. You have 2 layers of insulation against ever getting fired for anything or anything ever changing.

[laughter]

I’ll give you another example. The dockworkers went on strike a couple of years ago because of robotics. If you go look at a modern dock in Asia, it’s all robots. If you go to an American dock, it’s still guys dragging stuff by hand.

The dockworkers went on strike. It turns out there are 25,000 dockworkers working on docks in America. It turns out they have incredible political power because it’s one of these unified blocks of people. They won their strike, and so they got commitments from the dock owners not to implement more automation.

We learned a couple of things in that. Number one, we learned that even a union as small as 25,000 people still has tremendous political clout. We also learned that the dockworkers union actually turns out to have 50,000 people in it. They have 25,000 people working at docks, and they have 25,000 people drawing a full paycheck sitting at home from prior union agreements.

I’ll give you another great example. There are government agencies—federal government agencies—where the employees have civil service protections and are in public-sector unions. There are entire federal government agencies that struck new collective bargaining agreements during COVID where not only are their jobs guaranteed in perpetuity, but they only have to report to work in an office 1 day per month.

And so there are entire office buildings in Washington, DC, that are empty 29 out of 30 days of the year, that are still operating, and that we’re all still paying for. What the employees do is they’re very smart in this way, and so they figure it out. They come in on the last day of a month and the first day of the next month. So they’re in the office 2 days per 60 days, which means these buildings are empty for 58 days at a time.

[laughter]

You see where I’m heading with this? This is locked in. This is locked in in a way that has nothing to do with, like, people say “capitalist.” It’s anticapitalistic. It’s basically restrictions on trade. It’s restrictions on the ability to change the workforce.

So much of our economy is—I’m describing the entire healthcare system. I’m describing the entire legal profession. I’m describing the entire housing industry. I’m describing the entire education system. K–12 schools in the United States are a literal government monopoly.

How are we going to apply AI in education? The answer is, we’re not, because it’s a literal government monopoly. It is never going to change. There is nothing to do. By the way, you can create an entirely new school system. That’s the one thing you can do. You can do what Alpha School is doing: create an entirely new school system. Other than that, you’re not going to go and change what’s happening in the American K–12 classroom. There’s no chance. The teachers are 100% opposed to it. It’s 100% not going to happen.

So you see what I’m saying? There’s this massive slippage that’s going to take place. Both the AI utopians and the AI doomers are far too optimistic.

[laughter]

You see what I’m saying? They believe that because the technology makes something possible, 8 billion people are all of a sudden going to change how they behave, and it’s just like, nope. So much of how the existing economy works is just wired in.

We’re going to be lucky as a society if AI adoption happens quickly, because if it doesn’t, what we’re just going to have is stagnation.

I know you’ve got to run. You’re still welcome, but it was such a pleasure talking to you. We’re truly living in an age of science fiction coming to real life.

Yes. Could not be more excited. Really, thank you.

Thank you, Marc.

Thank you. Good. Thank you.