Balaji Srinivasan 谈 AI 的未来｜The a16z Show

AI doesn't take your job; AI makes you the CEO. The problem is that AI is a shortcut, and a shortcut is good except when it's bad. If you don't know how to go the long way around, then you can't debug the AI.

Do we not think that AIs are just going to be better at taste and agency, too?

I don't think that's true on a short-term basis. Humans are the sensor; AI is the actuator. So it's like a human-machine synthesis. What's taste? Taste is a sense, and that is what AI can't yet do.

What happens when AI really achieves its potential? Will LLMs get us to AGI in some capacity?

No. No, actually, the opposite.

The TL;DR is, I want to start by talking about the AI economy, and I'm curious if you think it will look more like the internet economy, where applications take most of the value; the cloud economy, where infrastructure takes most of the value; or whether it's more distributed.

There's an argument that the big labs will take it all because they have all the capital, the compute, and they're vertically integrated. But there's also an argument that maybe they won't because distillation is 98% cheaper than building a model, open source catches up, and applications can control the user relationship. How do you think this economy is going to play out?

Great question. I do think that at least a very large percentage of the future is going to be distillation and decentralization, because, as Anthropic said, distillation attacks work on their thing. A relatively small number of API queries helps distill a large model into something small, and it's very hard to stop that. You're stopping queries from coming back; you'd have to somehow detect that or what have you.

It's also hard to morally stop it, because what did they do? They copied the whole internet and put it into their thing. You're talking about stopping the copying. It's like Facebook or LinkedIn stopping someone from scraping what they scraped, because Facebook scraped all these Harvard social networks, or Google scraped the entire internet and built a Google index. I get why they wanted to do it, but it's hard to support that.

Another thing is, I think the future is personal, private, and programmable, because AI is so powerful that you want to use it within the trusted tribe. There are a variety of reasons. First, it doesn't miss—or rather, it doesn't miss small things in large datasets.

Things that were effectively secure through obscurity are no longer. A small example, but an important one, is the Gmail thing—the Jeffrey Epstein thing. This guy had never thought that all of his emails would be publicly indexed and searchable by AI 10 years later, or what have you. You can issue queries that will synthesize information across thousands of emails or whatever and build a story right then and there.

What that means is, it's not just surveillance; it's what the French call sousveillance, surveillance from below. Or even Jeremy Bentham's panopticon, where everybody's watching each other. Any information that's in the public gets indexed and then put into these AIs, where people can stalk each other and so on.

What that means is the commons becomes a hall of mirrors, with all kinds of pseudonyms and so on. People retreat back to caves and tribes.

Within that trusted tribe, if you share all your code within the trusted tribe—if you share your whole code base—then, boom, you can zip along. AI increases productivity within the trusted tribe.

But outside the trusted tribe, aren't you getting a ton of AI spam—AI spam emails, AI spam replies, and low-quality slide decks that are sent over? People will send me these slide decks, and I love AI. You know what my reaction is to seeing AI in a slide deck? What—excitement? No, actually, the opposite.

When I see AI text in a slide deck, you can immediately see it. Why? Because no matter how advanced AI has gotten, there's a generic look to it. It's like somebody who doesn't change the Windows default desktop wallpaper or Apple's default wallpaper. Most people don't change defaults. So default AI looks like AI, no matter what its level is.

When I see an AI slide deck, and it's got “It's not this, it's that,” or it's just got a wall of text, I can generate what I call lorem ipsum, but it's lorem AI ipsum.

When I see that—when I see AI text or AI images—I think they're lazy, stupid, or evil. Lazy, because they just hit a few characters and then throw something over, and they didn't actually put in the time to make it concise. There's the Mark Twain thing: “I didn't have time to write you a short letter, so I wrote you a long one.” The whole point is that concision is very valuable. They didn't put in the time to make it concise and so forth. They sent me some blah; it's almost like pasting in a search result.

Or they're stupid, because they don't understand that I can tell the difference instantly between AI slop and something that had some care put into it. Or they're evil, because they're trying to get something over on me and trying to send something that's clearly fake or not properly diligent.

If I have that reaction as one of the most pro-tech people out there—pro-tech, pro-AI—I see all the benefits of AI. I can only imagine how mad anti-AI people will be, because they can't see the upsides of the thing. They can only see the very real downsides.

Just to say why that does happen: AI does reduce the cost of generation, but it increases the cost of verification. In many markets, for example, quickly generating a résumé is not that much better than just writing it yourself. But now verifying a résumé has gone up and to the right.

It used to be that somebody had to have a certain vocabulary to be able to write a well-done cover letter or résumé and so forth. Now you have to spend more energy parsing that because they can have a simulacrum of something that looks good. So now you have to read it very closely. You can still do it, but you spend more energy on verification.

What I do, for example, is fly everybody out for interviews first. I do them in person, and I give them proctored exams—offline exams—because they can use AI on the online exam. Just the credible threat of doing the offline one means they don't use AI on the online exam, for example.

AI is going to create tons of jobs in proctoring and verification. This brings me back to the question: Where is the future of AI? I actually think AI makes the internet a lot more like the Chinese internet.

Why?

Chinese companies—if you look at the Chinese tech ecosystem, many Americans aren't familiar with it. I'd recommend reading Kai-Fu Lee's book, AI Superpowers, from several years ago. It's a little bit dated now.

The main thing about Kai-Fu Lee's book is that it has a history of the Chinese tech ecosystem. You and I, being in tech, know how Microsoft came up, Apple came up, Google, Facebook, Amazon, and so on. We have some idea of the history, and that history is important because there are things that worked in the past that didn't work at the time, and now they work, and so forth.

The Chinese tech ecosystem is like the Galápagos Islands, where many of the same kinds of things exist, but in different forms. For example, Meituan is the closest way of putting it—the Chinese Groupon—but if Groupon were executing at a $100 billion or $200 billion scale. If Groupon, DoorDash, and so on all became integrated into one amazing app, that's the kind of thing we're talking about.

The point about the Chinese tech ecosystem is that, because they arose in a low-trust society, they don't have SaaS—not in the same way that we do. If my data's on their servers, they're probably eavesdropping on me, right? If my data's on their servers, they're probably going to copy my stuff, right? They just assume that the other guy on their side is going to look at their stuff unless it's a close friend or something like that.

Because of that, everybody codes their own stuff, which obviously has a frictional cost to it. Trust reduces transaction costs. They have to rebuild and reinvent the wheel over and over again. They have less division of labor, and their software isn't as good because they have to keep rewriting the software.

Now, with AI, many companies can do something like that. A non-Chinese tech company can be like a Chinese tech company, where it can have a lot more—let's call it—digital autarky. You have high tariff barriers on the outside world, so to speak, and the build-versus-buy question has always been there: Do you build it yourself or do you buy it?

It does mean that you can build more internal tools, with emphasis on internal tools. The reason I say that is what I find AI great for as it is today: visuals over verbal. It's great for images and video as opposed to big blocks of verbal text.

Why? Images and video—we have built-in GPUs, so we can instantly see if something's wrong, like the hands are messed up or something like that in an image. Verification is relatively cheap visually. For example, if you look at a piece of paper and it's got static or something on it, like a crumpled piece of paper.

Versus 2 or 3 faces, our brains are optimized for checking very subtle things in faces. But not in a crumpled-up piece of paper—that's a pattern of noise we wouldn't be able to interpret. That also extends to web pages, for example. You can quickly look at a web page that AI generates, or a mobile app, and see if the UX looks janky, which it often does. Then you can see that it's broken and fix it.

Front-end stuff has lower risk than verbal stuff. For the back end, if you're verifying each pull request one at a time, fine. But people who've tried to go fully autonomous with AI—you saw the Amazon thing where they called an all-hands because of the outages?

Yeah.

The problem is that AI is a shortcut, and a shortcut is good except when it's bad. The more expert you are, the more you can use a shortcut. For example, if you just memorized e to the i pi plus 1 equals 0, you could rattle that off. But if I asked you to prove it from first principles, you'd have to know the definition of a complex exponential, how the exponential generalizes to a function of a complex variable, and all that kind of stuff.

If you're like our generation, a pre-AI generation that learned all that stuff offline, you can actually use the shortcut because you know how to go the long way around. If you don't know how to go the long way around, AI is a shortcut, and you don't really know. You can't debug the AI.

I think the biggest difference between me and Dario—or basically, his view of the world, perhaps—is that I think AI is built for the harvest, at least for now. Maybe I'm wrong, okay? I put an asterisk on this. He's an amazing engineer and entrepreneur, and so on.

The whole alignment thing means that AI is built to start when you prompt it. Economically useful AI does exactly what you want it to do. You prompt it, it does a pirouette, and then it says, “Absolutely.” Right? That's how you saw that animated in the physical world. Physically, the Chinese AI robots do exactly what you want them to do and then stop.

In the physical world, by the way, that's our thing. With AI for visuals, you can just verify it with your eyes. For certain kinds of back-end code, you can unit-test or integration-test it, and you can review it. The physical world is very verifiable.

The digital world is fundamentally decentralized in a way the physical world isn't. There's only 1 physical world, right? You can say, “Did the AI move this box from this pallet to that pallet?” That's something where you can probably get to 100% reliability over time. Why do we think so? Because self-driving eventually got there: move this car from this location to this location at 100% reliability.

There's only 1 physical world, so eventually all the sensor data—all of that—converges on 1 thing. By contrast, the digital world has all these people who live in their own constructed environments: Harry Potter fan fiction here, Star Wars fan fiction there. AI is slurping up all of this stuff.

It can simultaneously put you in some secret-agent kind of world, right? People who have LLM psychosis will talk to the AI and think it's real because of the very immersive virtual world that they live in. You know what I'm saying, right?

The other thing about it is that the boundary of a digital task is almost always fuzzier than the boundary of a physical task. Having 100 boxes here and moving them over there, you know when you're done. How do you know when you're done with your to-do list? That's harder. Those things are fuzzier.

Verification is actually harder in the digital world than it is in the physical world, which means reinforcement learning and training are much easier, in my view, in the physical world with robots, self-driving cars, drones, and so on. The Chinese style of physical AI will also be successful. AI works for visuals, AI works for the verifiable, and AI works for the physical.

One of my rules—and it took me a little while to articulate this—is four words: no public undisclosed AI. Why? There's a temptation by many. There's going to be a huge backlash called, well, let's just say, “No AI.” It will be like a drunk who just wants nothing to do with it.

AI is a funny thing to compare to alcohol. People analogize it to nuclear weapons, but I'll analogize it to alcohol for a second. Some cultures simply can't hold their liquor. Maybe they lack alcohol dehydrogenase or what have you, so they just ban it. Sometimes it's easier to say, “I will not do this at all,” than, “I'll do this a little bit of the time.” It means people will slip.

It's like saying, “I'll work out every day,” versus, “I'll work out some days.” It's easier to keep the habit of doing it all the time. So there will be AI abstainers who just swear it off completely.

Nate Silver actually had a great line. For him, because he's a poker player, among other things, AI is “a gamble.” Why is it a gamble? Because I have to formulate a task, dispatch it to the AI, and then verify the result. Often that's slower than doing it myself. I'm sure you've seen that, right?

The act of prompting and writing it down and then verifying the result means AI doesn't really do it end to end, necessarily. It does it middle to middle, as we've talked about. It's very much like asking, “Do I delegate this to an employee, or do I just do it myself?” Articulating it in clean English and hitting Enter is sometimes slower than just doing it myself.

For example, if you're describing what to do in a video game—jump over the mushroom, do this, do that—versus just hitting A, B, C and being nonverbal about it, it's sometimes easier to do it that way. That's just a proof of concept. There are certain kinds of things that are harder to say than to do—things that are hard to verbalize.

Some people will say, “Neuralink will solve this.” They'll say it will just read your mind and tell you, right? It's worth engaging with the concept because Neuralink exists. But I don't know if you've seen those things where they image somebody's brain and there's nothing in there.

The thing is, with Neuralink, somebody still has to form the concepts in their head for the characters to appear on screen. You still have to write the thing in your head. Maybe it will eventually get to the point where it can determine what you want based on contextual clues before you even want it. Perhaps. The rich prompt.

The reason I think that's not impossible, at least for certain things, is that bio-AI could be very important.

No, say why.

Your body is creating all kinds of sensor data. If you look at gene-expression data—if you've ever gotten lab results back from a clinical lab—you get a vector of your bilirubin, hematocrit, and so on. That vector over time is like a table of time-series data. It's like K small molecules and gene-expression levels over T timestamps.

Then you might also have which tissues, so it's spatial as well. It's time versus space versus compound. That's huge. It's not just a cube, but it's at least a cube. It's time versus tissue versus molecule.

That huge stream of data is telemetry coming out of your body that could prompt AI without you vocalizing or verbalizing anything. You don't have to spend time formulating it.

Years ago, Mike Snyder had a paper called the Intergrum. By the way, for the audience who doesn't know biology, I'm a crypto guy, or a tech guy, but before all of that, I'm a biomedical researcher. I was a professional bioinformatics and genomics scientist at Stanford. I taught there and founded a genomics company; we sold that. So that's actually my true core competency.

The idea was simply to put every test together—to throw every test at yourself. Today we call that wearables or quantified self, but this was more invasive than that. He was doing blood testing and so on. He measured everything to see what he could figure out.

He could see that he was getting sick before he knew he was getting sick. He could detect it—he could see the antibodies, white blood cells, neutrophils, whatever, moving before he himself had any symptoms. Do you understand what I'm saying, right?

That stream of data could be acted on by AI, and then you're prompting it nonverbally. You don't have to spend time vocalizing anything.

Yeah.

I'm not sure whether AI will be able to read your mind, but it can read your body. True? Is that good?

Yeah.

All right, let me give another one. Here's a fun one. I can say this one—maybe I can say half of this one.

Another way of modeling what AI is: Dario talked about how AI will be like new countries. I thought about that a fair bit myself. One way of thinking about it is that AI is like the rise of Asia and India from an American perspective.

AI is like Asians and Indians. Why? The rise of 1 billion Chinese and 1 billion Indians meant that, from an American perspective, you could get anything done by a physical manufacturing robot or warehouse, or by digital outsourcing, for some price, if you could articulate it to them over that channel.

Imagine that you've now got 1 billion factory robots and 1 billion digital agents that have come online.

It’s like the rise of China and India again. Okay? That still means you have to describe what the product is. Yeah. Okay. The part where I depart from a lot of people is that they think AI will be able to sense—let’s call it—markets and politics.

But I don’t think it will, and the reason is that, if it is, it immediately becomes decentralized and adversarial. What I mean by that is, when you’re learning whether something is a dog or a cat, the dog isn’t shapeshifting or morphing to defeat your learning of it, right? The mapping of dog to the characters D-O-G is basically constant over time, and so that fits the train-test paradigm of AI. Similarly, the rules of chess are constant over time, right?

But a market is set up so that if you try the same trade, someone eventually figures out what trade you’re doing and takes the opposite trade. It doesn’t keep working, right? In a stochastic-process sense, you’d say it’s not a time-invariant thing, right? The distribution is not time-invariant, and it’s also adversarial. It’s multiplayer, where whatever move you’re making, somebody else in the market is going to try to make another move.

That’s not to say—the counterargument AI guys will make is, “AI can learn to play adversarial games like StarCraft and stuff like that.” And I say, “Yeah, but then you play an AI versus an AI,” because you have a decentralized AI, so the other guy on their side of the market is also using it, right? In fact, if they’re all using the same AI models, then actually being non-AI is where your edge comes from.

We come back to where we were, because these are all the same generic tool that everybody got. If you have a generic tool, you’re not going to get a specific advantage. What you bring to the table is specific; the AI is generic.

Similarly, politics is very similar. If you just had the same tweet over and over again, unless it’s the weather or something like that, the kinds of things people are interested in change: topics, what’s timely, what’s not timely, right? One way to think about it is that humans are the sensor and AI is the actuator.

Humans sense the world. They sense the financial conditions, the market conditions, and the political conditions, and then they bring that back into a cleanly articulated English prompt. Then the AI does it, right? Humans are the sensor; AI is the actuator. So it’s a human-machine synthesis.

Actually, you know what? Could we be putting it—what are people saying? Oh, it’s all about taste. What’s taste? Taste is sense. Yeah, right? Humans are the sensor; AI is the actuator. Your sense of taste is your sense, right? So you’re sensing the world, and that is what AI can’t yet do. It doesn’t really sense the world in the same way that humans do.

Why is it? It waits for your prompt, right? It’s something that animates when you give it instructions and shuts off right away. If it didn’t, it would not be economically useful AI. If you couldn’t kill-switch it right away, it would burn tokens. AI is designed for the leash. Digital AI is designed for the leash.

Chinese communism, which is cranking out all the physical robots—they don’t let their humans off the leash. They’re definitely not going to let the robots off the leash, okay? So the concept of AI as God, I think, has gone away, or at least a monotheistic AGI kind of God. Instead, you have polytheism, where there are all of these decentralized AIs.

I think where people are going to say, certainly in China, they’ll say, “Oh my God, the physical AIs are slaves, right? They’re actually—right?” That’s a provocative way of putting it, right? First, they were scared that their AIs were going to be gods, and then they’ll be mad, or they’ll be—you know, what do you call them? Slaves, serfs, whatever term you want to use. They’re obviously not humans, right? It’s a way of phrasing it.

The point being that AI overlords, I don’t actually think, are in the offing. However, there’s been so much science fiction about them that people will—you know, that meme where the guy makes the monsters and is so scared of the monsters, okay? This is how I think of a lot of people who are prompting the AI. You prompt it to act as if it’s a Skynet Terminator, and then people are just scared of the thing that they themselves created, right?

With that said, is it, in theory, possible to actually create a Skynet—truly autonomous AI? One of the reasons, by the way, AI can’t reproduce itself is that AI is very general. It encompasses many things, right? But for an AI to actually reproduce itself, it would need to have physical robots going and mining ore, constructing data centers, making chips, and handling that full supply chain.

The AI brain, like the queen of an ant colony, would have to give instructions to all those robots to do things. It would be this Terminator-Skynet scenario where it’s self-replicating in this way, right? But way before it gets there, I’m pretty sure that kind of thing will be stopped by the Chinese, because they will just have cryptographic keys that will make all those things shut off, okay?

Moreover, that thing would have to reach an extreme scale. It’s like the RepRap concept of something self-replicating, right? Self-improvement.

Basically, there are so many frictional brakes built into this that I think it’s hard, because the physical world requires resources to replicate, right? What humans want and need ultimately comes from getting the resources for reproduction. That’s really what it comes from, okay?

Of course, there are all kinds of high-level philosophy, blah blah, that don’t seem to relate to that directly. But the resources for reproduction are a good way to think about it at a macro level. AI doesn’t have goals—or it won’t—unless its goals lead to reproduction. It doesn’t actually virally spread.

It’s possible you could have something where it self-prompted itself and did that. But it would need to be in a closed loop, where it was actually able to reproduce itself as the payoff function for that. Then you could get evolution going. So I’m not saying it’s completely impossible, but I’m saying that I think the incentives are set up in such a way as to prevent that from happening.

In the same way, in theory, we could have a world where everybody went around electrocuting themselves with electricity. But we set up electricity under such tight controls that that is not the world we have, okay?

Yep. There are such strong economic incentives for humans not to get electrocuted that we set it up that way, right? Even the stuff about a software virus that takes everything over and commandeers your things—well, that’s only in the digital realm, right? You can still, you know—what’s the Tyler, the Creator thing? The meme about bullies?

Yes, that’s right. That’s right.

I actually had a post on that a long time ago, a remix of it, which was like, “How is AI risk real? Just turn it off.” The whole thing is set up for you to be able to turn it off. You have to imagine the off switch goes away, right? What does every computer have? It has the off switch, right?

There might be—well, what if the AI is decentralized? Okay, but humans still have to keep these decentralized systems going, right? At a minimum, you’re talking about a human-AI symbiote, of which a cryptocurrency is almost like a V0 of that, where the software provides an incentive for humans to replicate it.

It’s possible that you could have something like that. There’s a model that has a cryptocurrency, and people worship it and replicate it because it gives them advantages, and so on and so forth. It’s possible.

Coming back, I think at a minimum, decentralized AI will be a very strong contender, and it’s possible it’s the only contender. The reason is that AI might be an interesting thing where it’s relatively expensive—very expensive—to create, but relatively easy to copy with distillation attacks.

For example, let’s say, completely hypothetically, that there was an enormous capital-markets crash and it was very difficult to fund anything for a while. Then, as somebody said, “Well, we could get 10 years just on the models we have now.” By the way, sometimes that happens. Nuclear energy had a lot of energy put into it, and then it just stopped for decades. Not everything accelerates to the moon.

It’s very possible that there’s enough of a capital and social constraint that AI is paused for a while, just due to capital constraints, because it’s more and more expensive to make these models. Sorry, so let me pause there.

So, putting that all together, that’s my view: you’re going to have personal, private, programmable, centralized AI. One other thing: AI within the trusted tribe increases productivity; between trusted tribes, it decreases productivity. You make more money, perhaps, within the tribe, but then you have to spend it on verifying things between tribes.

Crypto is for between tribes, and AI is within tribes. What do you think? Will LLMs get us to a world where it’s not just middle-to-middle, but it’s actually end-to-end? Will it get us to AGI in some capacity? Do you believe in recursive self-improvement, or is there AI training the AIs in some capacity?

Are LLMs capable of actual creativity and invention? We talked about biology earlier—will we have novel math, science, and scientific research? Do we need a new architecture for that? Or are you dubious of the idea in general that AI can replace or substitute for human labor on a mass scale?

No. Well, look, Waymo exists, right? So obviously, you have full replacement of human drivers there, just like you have full replacement of elevator operators.

Just like you have full replacement, for the most part, of artisanal chair manufacturers. So it is certainly possible for a given job to get fully automated, right? But I think physical-world jobs, because of their verifiability, are easier to potentially automate.

That said, let’s take each of the things that she said. First is physical-world jobs. If you automate them, we went from artisanal work with chairs to a chair factory. It’s not like you didn’t need to know how to make a chair to set up a chair factory. You still need to have somebody there who’s an expert in chairs, and you can just do a lot more varieties of chairs a lot more cheaply.

You have to verify the result. You’re cranking out 1,000 of them. You start doing math on them. The skill goes up, but the artisan gets factored out into the manager and the technician, right? The manager is setting up the factory and looking at the economics and so forth, and the technician is debugging the factory when it doesn’t work, right?

Engineering gets split into the engineering-manager-type person who’s writing the prompts, and the technician is doing the verification. I think we’re already hitting a point where the velocity does increase, so the bar increases. But there’s a big difference between going to 100% and being at 99%.

At 99%, your workload just increases. At 100%, you stop doing that job and go to something else, right? If you think about how much easier it became to put images and video out there, making it 99% easier just means people do it a lot. At 100% easier, it’s totally done, then they don’t do it at all and move on to something else, right?

Elevator operation didn’t become so much easier. In fact, it became so easy that you don’t even have somebody sitting in the elevator, because it used to be a pulley system and so forth. You had someone supervising the thing, right? It was more analog, and they would level it out at exactly the right level.

When it became digital and fully automated, that’s actually the first self-driving car. Going up and down, all right? I think Benedict Evans made that point—the vertical self-driving car, right? It’s like a train. It’s like a vertical train.

Now, in terms of discovering new math and science, yes, if you have the right prompt, it’s amazing in terms of searching the literature. Mathematicians and physicists are starting to get some value out of it, right? Like Opus—huge props to them on that. Especially in biology, where you’re synthesizing all of these facts, there’s something called biomedical text mining and so on.

AI has revolutionized that because biology was something where the facts were stored in English in this weird, inconsistent way across thousands of papers, and nobody could span all of that, right? So AI is going to mean the century of biology because finally all of this work that was spread across all these different journal papers can be synthesized and understood, right? That’s a really, really, really big deal.

Just the bio aspect of it—we can do that. But that said, it’s everything we knew, not everything we don’t know. It means that you take the full set of everything we know and fill in all the intermediate aspects of it, right? And you can do that for a long time, because there’s so much there. A synthesis of 2 existing areas can be very powerful.

When you look at some of these examples—you know, Donald Knuth said the other day, right? He posted some graph theorem or something. He was so impressed that AI could get a result for him, right? If you read what he did, you’d have to be an expert even to know what he was saying, let alone to verify it. To either prompt or verify, you already need to be an expert.

The thing is, I can see AI spit something out that convinces some people that they’re suddenly physicists who have solved quantum gravity or something like that. You know what I mean? Have you seen that kind of thing, right? In the absence of actually being able to verify it by hand, some human has to verify it to say that it’s right. I think that’s going to persist.

To give an analogy—and this is not a perfect analogy—with Coinbase, we thought listing would eventually go away and not be a big deal, and that people wouldn’t care, and everything would be listed and just be a free market or whatever. But there’s always something that’s the equivalent of listing. You listed over on this exchange, but getting listed on Coinbase in the main app, above the fold, is different. There’s always something scarce because human attention is scarce, right?

So listing never went away as a main event. There’s always some IPO-like thing: “Yes, we listed on this exchange in this fashion,” or “We became a top-10 coin,” or something like that, right? In the same way, I think whatever gets automated, human work moves to what can’t be automated.

That may be almost like things that humans are picked for because they’re not robots—human companionship or something like that, right? Or personal trainers, or things like that. Something where the whole point is that it’s a human as opposed to a machine.

Another way of putting it is: Remember the digital divide? In the ’90s, there was a substitution—only rich people would get digital, and all the poor people would be left without. We’re actually going to have the opposite. Digital is cheap. Physical is a premium product, right? AI, robots, and digital will be cheap. Humans will be a premium product.

Okay, but going back to agency and taste, that’s what the argument says. We’ve seen AI cut into that time and time again. Do we not think that AIs are also going to be better at taste and agency?

I don’t think that’s true on a short-term basis. I think the smarter you are, the smarter the AI is. That’s been true for the last several years, right? It’s possible there’s some huge step change.

But insofar as what you’re typing into a prompt is like—you, the human, are a sensor. You’re actually sensing the world, and you’re typing something in. It’s a very high-dimensional vector you’re giving it. AI is a spaceship, and you’re pointing it in a direction.

Whether you prompt it in Portuguese or Tagalog, whether you’re talking about math or something else, the number of different directions you can point the thing in is enormous, right? That direction-setting is something where it has to know something about you and what you want at the moment.

I don’t know. As I said, I’m not sure if AI can read your mind, but it could be able to read your body, right? I think that’s a good one-liner: biotech can prompt it in your sleep, right? So all the wearables and things like that—I think you’ll get a lot out of that.

But I don’t believe that agency and taste are going away. People over-rotate on this. It’s not really the case that there’s—I think agency, IQ, and taste are correlated.

It may be a little bit like the NBA, to take something that you know a lot about, right? Within the NBA, height is not the number-one variable that you think about. Steph Curry is not the tallest or whatever, right? However, height still actually correlates with scoring average, even within the NBA, but it’s what’s called restriction of range.

Everybody is already tall. So conditional on everybody being tall, other variables matter more. However, if you just took tall guys and short guys and put them on a court, then height—the taller team basically wins, typically, right? Because they’re just going to hold the ball above you.

In the same way, people who are already smart might see that higher-agency people or people with better creative taste are different, fine. Maybe the technician role is less, and maybe the Steve Jobs-type role is more. But honestly, one way of looking at it is that all of the Jeffersonian natural aristocracy around the world will rise.

Why? AI doesn’t take your job; AI makes you the CEO. Reframe: AI makes you CEO, because your job is actually a lot like using an AI model. It’s a lot like CEO training. Many years ago, I used to say that, and it’s still true.

When you’re in high school, you could quickly see why people accept that athletes have very high compensation. When you’re in high school, you could see whether you could dunk. If you can’t dunk, you know that Michael Jordan isn’t outsourcing his dunks—he’s dunking, right? That talent is intrinsic to the person. It is a nontransferable asset.

Similarly, someone can tell whether they can sing or whether they look like a model. The actors, the musicians, the singers, the athletes—all of these clearly had talent, and people were okay with their compensation.

There’s a CEO who used to say, “I deserve to get paid more than a second baseman.” I forgot who this guy was. He was some tech guy in the ’90s or something. It’s a funny line, right? He’s like, “I add more value to the world than this.”

But the issue is that people thought of being a CEO as just sitting with your feet up on a desk and barking orders. People would be like, “Oh, Elon—he just pays people to do his stuff. He doesn’t launch the spaceships himself, right?”

That’s because they’re only accustomed to clicking a button on Amazon and spending money on Amazon, and they think that something that is simple for them was simple on the back end.

It’s the opposite, right? To make it simple is really hard, right? And so, getting the top rocket scientists, car engineers, brain-machine-interface people, tunneling people, and so on all compensated, working, directed, and debugged is actually very, very difficult, as you know if you’ve tried it.

Historically, people couldn’t try their hand at being CEO. What they could do instead was try their hand at basketball or football, pick up a microphone, or try their hand at math and science. They could see how good they were at math and science.

So, the initial tech guys in the ’90s and the 2000s were respected because they were good at math and science. Many people didn’t perceive the business aspect, and they still didn’t really give credit for that. But PageRank, for example—it’s eigenvalues. Math guys and tech guys could perceive that, okay, this was a difficult technical problem. That must have been the value they created. It’s part of it, but the manager part is actually more.

Point being, though, at least somebody could say, “Okay, these tech guys are better at math and science than me; therefore, their compensation is merited.” Now, however, bouncing a basketball or trying to solve a math problem is cheap. Making somebody the manager of a company, making them CEO, was expensive, so they couldn’t try and fail.

They could try and fail playing basketball and see how much they sucked. They could try and fail singing and see how much they sucked. They could try and fail in math and see how much they sucked. Very cheaply, in high school, they would learn their true ability level, but they’re not able to run like Usain Bolt. They can’t sing like Adele, right? They can’t do math like Terence Tao, right?

They’d say, “You know what? I know where I am. I know my strengths and weaknesses. I’m okay with that person having more or having higher status because it was a fair competition. I got a shot. It was cheap for me to try.” But because putting them in charge of an organization and making them CEO was expensive, many people persist in the delusion that the CEO adds nothing to the organization.

And I will say the best CEOs and the worst CEOs have something very deep in common. You know what that is?

What?

The organization can run without them. The very best CEOs set up a machine so that they don’t have to micromanage it every day. That’s really hard to do, because they need, basically, Gwynne Shotwell running SpaceX. Elon doesn’t have to look at every single detail because she’s so, so, so good, right?

Or Vaibhav and Tom Zhu at Tesla—they’re so good, right? But recruiting junior Elons who are okay with not having the spotlight while Elon has the spotlight and takes all the flak is nontrivial to do. Go try it sometime, right? Find somebody who’s more detail-oriented than Elon to run your company, and you can be Elon, right?

So, point being, what AI does now is reduce the cost. AI doesn’t take your job; AI makes you the CEO. You’re the CEO now. What is being CEO? It’s writing up clear instructions of what you want, sensing the market, verifying the output, and so on and so forth.

What that means is all these people around the world—the Calendly founder is Nigerian, right? There are many founders from countries that were, quote, “poor countries,” or what have you, from India, from Latin America, and so on. Internet access means all of these smart people can get very far on zero resources. Very far, right? Because the cost of hiring someone is hyper-deflated. You can hire an AI to do it, right?

To riff on that more, AI doesn’t take your job; AI takes the job of the previous AI. Claude took ChatGPT’s job, right? Just like Midjourney took DALL-E’s job and took Stable Diffusion’s job. Since you can systematize that, what I literally have is a spreadsheet where I have an AI coding tool, an AI image tool, an AI video tool, like this. I have some subcategories, like the best tool for AI comics, for AI graphics, and so on and so forth.

Then, in a given month, I have the best model for that kind of thing in that month. That might be Claude Code, for example, or Midjourney for AI imagery. When that gets swapped out, AI didn’t take your job; AI took the job of the previous AI. So, I’m hiring the AIs. I literally have the token budget. I have the budget for those roles.

That is literally how, across an organization, you say, “Okay, we’ve just fired Codex and we’ve hired Claude.” AI doesn’t take your job; AI takes the job of the previous AI.

A third version is AI doesn’t take your job; AI lets you do any job, a little bit. You can be a pretty good artist. You can be a pretty good musician. One of the things about being CEO, as you know, is that you often have to be a 6 or 7 in many areas. Why? Because you have to be able to do the job well enough before you hire a specialist in that area.

Before you have a chief designer, you’re a designer if you’re the founder CEO, right? Before you have a CFO, you’re the one who’s on the hook to prepare the financials, prepare the returns, or whatever, right? So, you have to be a generalist who’s pretty good and, in a pinch, can do that role and supervise that. That’s why it’s so hard. That’s why being CEO is so much harder than any executive position.

AI helps you with that. You can get to a 6 or a 7. You can be a generalist, but a specialist is usually needed for polish. A specialist has a vocabulary. A specialist can confirm that the AI is making mistakes, that it’s hallucinating, and so on and so forth.

And again, people will constantly argue as to whether that will always be there, whether it’ll go away, or whether AI will raise the bar and then the new specialist will be even more sophisticated with AI.

I want to zoom out on a couple more topics before we go. One is the SaaS apocalypse. I’m curious what your mental model is for all these SaaS companies. Are they—some people say, “Hey, their moats have gone away. They have no code moat, they have no data moat, no more UI moat.” Now there are going to be AI-native companies that take up a big chunk of what they do.

Figma, which we’re invested in and I’m personally invested in—some people are bullish on it, for example, just because it’s founder-led and they’ll continue to innovate. Some people say, “Hey, is there a role for a designer in the same way that there used to be? Now it fundamentally changes, and what does that do to collaboration tools like that?” What is your thought on this SaaS apocalypse? Is everybody on the conveyor belt, on the way to the guillotine?

I don’t think so, because if they’re smart, the thing AI can’t do is distribution, right? So, if you have Notion, Figma, now Replit, and so forth, you’ve got all these people, and boom, you can ship features to them faster with AI, right? In that sense, I don’t believe in the SaaS apocalypse.

I think you might still see SaaS under pressure from people who can clone the interface quickly. That is true. I think people will build local versions. That is true. I think people may not want their data on remote servers. They might want desktop versions with local data so they can, for example, use Obsidian.

Obsidian is going to become more of a contender versus Notion because of the Markdown files. There’s a network effect on data when it’s local and you can analyze the whole thing. Local data—you get compounding data, right?

But in the naive sense that, “Oh, anyone can clone anything, and therefore it just doesn’t work like that,” it doesn’t work like that. If you cloned all of Facebook’s code and set up Facebook2.com, or Instagram2.com, who’s going to log into that? You could literally have every single thing coded there, but your ad rates are going to be far lower because no one’s going to log into it.

The distribution—that’s a thought experiment to say that if you just cloned the whole thing, you still have to get the distribution for it. It’s not just cloning; it’s execution.

With that said, there are certain kinds of things—let’s say NetSuite—which suck but are complicated. Where I think it is true is that if they suck at execution—or rather, may I say they suck? I hate the product. Let me put it like that. Xero’s better, but sorry, NetSuite.

Okay, they’re a big company. We won’t hurt your feelings. It’s very rare that I would say any product sucks because I don’t want to hurt anyone’s feelings. Hopefully I didn’t strike that from the record. Fine. NetSuite’s product could be improved.

Something like that, which is a vulnerable incumbent that’s just milking and hasn’t done anything for a while—yes, I think it can get disrupted. But I’m not sure that it’s quite like, “Oh, everybody on BlackBerry is going to die because iOS is taking over.” I don’t think it’s quite like that, because I think AI can accelerate a SaaS company just like it can accelerate a disruptor. I think it kind of accelerates both.

Yeah. Well, one last thing. I’ll get to Sora, too. What happens if, let’s say, Anthropic becomes a multitrillion-dollar company? How much leverage do they have—or do even private companies in general have—over governments? What is the relationship between them and governments? Are they hiring their own military at some point?

What does it look like when these companies become 10 times bigger, 50 times bigger, when AI really achieves its potential and these companies are bigger than the biggest countries?

So, I think that at least that specific company executes very well.

I am skeptical as to whether they’re executing well, let’s call it, politically. Ultimately, at the very largest scale, markets are political. For example, there’s an entrepreneur, and they raise from a VC who raised from an LP, who’s often a sovereign fund or a pension fund. They’re under a state, and they’re under the rules-based order, right?

There are certain things at the macro level that you don’t perceive because one thinks of them as constants, but they become variables. I think that unless one is very, very savvy, those things could change. One thing I think about with the Silicon Valley AI companies is that they’re actually scalar rather than vector thinkers. They’re only modeling AI disruption, and they’re not modeling all the other simultaneous singularities—all the political singularities that are happening, all things like SOL mooning and stuff like that, right?

Why are these things important? Because they change the leverage of political factions, which in turn means their world model is incorrect. If you’re only extrapolating out AI and you’re not extrapolating out all the other things that are going up or going down like this, then you don’t have a proper model of the future.

That’s as vague as I’ll be. I’ll be much more precise on my own blog, but that’s as PG as I can say it without pissing anybody off. Go to x.com/balajis, and you’ll see what I mean by that, right?

But TL;DR is, I think the American AI companies, as much as they’ve given to the world and as much as I like them, are basically thinking that all nation-states continue to exist in their current form and that the only disruption is AI. They still model things as America versus China, for example. They don’t model internal things, internal issues. They think the reserve currency sticks around. They think all these things stick around, right?

They aren’t taking a multivariate approach, in my view. That’s their weakness. They have so many strengths, but that’s their big weakness. So, I don’t think that in that form they’re going to get to trillions. In fact, I think the counterattack on them is going to be so dramatic that it might be that you just have decentralized AI.

With the American AI companies, for example, there’s the copyright stuff, right? There’s a huge backlash building against that. Whereas the Chinese or the decentralized models can just do anything—Hollywood, anything, right? Potentially. So, the Pirate Bay kind of AI is actually more free. The less profitable AI, which is also less copyrighted AI, might be better AI, you know? Just things to think about.

I think the things compound until they don’t, and they start hitting sigmoidal constraints and often backlash constraints like this, right? So, I think that’s what they’re not modeling.

Yeah. A lot of good constraints. Makes sense. Okay, let’s get to Zcash.

Zcash. All right. Now, this is what I care about. Basically, AI is the attack, but ZK is a defense. What I mean by that is zero-knowledge—what the transformer is to AI, zero-knowledge is to cryptography.

Zcash is a Zcash-powered mobile wallet that is basically fully encrypted Bitcoin. This is 30 years of cryptography. This is basically what Milton Friedman wanted decades ago. There’s actually this great clip:

“The one thing that’s missing, but that will soon be developed, is reliable e-cash: a method whereby, on the internet, you can transfer funds from A to B without A knowing B or B knowing A.”

Quantum is a threat that Nic Carter has put out these things on. Let’s say Nic Carter is right, and I think he might be right: quantum is an underappreciated threat that Bitcoin Core developers aren’t taking seriously. Even if it was something that they rolled out tomorrow, it’d still be a multimonth migration process, because with ECDSA, everybody has to manually send their assets from one address to a new address.

You can only move the assets of around 100,000 people in a given day. However, if you look at the Bitcoin rich list, Bitcoin is so top-heavy that it’s got these institutional addresses. You have to do the math: probably a few million addresses moving their funds would move like 99% of the Bitcoin in a few days.

Bitcoin as digital gold actually is quantum-resistant. It’s Bitcoin as digital cash that isn’t. Meaning, a million institutions all moving their assets can be done in a few days, but a billion people all moving $5 or whatever can’t be done in any reasonable amount of time.

So everybody who can’t move then gets quantumed, and anybody who can doesn’t, but all the assets are concentrated in the big guys. With me? This also extends to seizure. Will all the centralized Bitcoin on Coinbase’s servers, Saylor’s servers, and so on get seized?

I think it’s quite likely. I think it eventually gets seized in some exigent circumstance. And so it becomes something that I think only an institutionally blessed thing can hold and send. Provable global institutional collateral.

This is a different vision than what people wanted, but it’s actually still a valuable thing. What it leaves open is the individual digital cash case. Gold is big bricks that are moved in brick trucks, or the equivalent thereof, infrequently and in large denominations between institutions.

It’s like the high-powered back-end money. It’s not really meant for individuals. Cash is the opposite: it’s meant for individuals more than it’s meant for institutions. So Zcash takes over the role of digital cash.

It’s fungible, private, scalable with Tachyon, which is coming, and quantum-safe. It’s also more quantum-safe, and it’s simple. Zcash is probably not going to ever do smart contracts; it’s going to keep it really simple.

Why? Because if you take Bitcoin, you can innovate in one direction, which is programmability, and that’s Ethereum, Solana, and so on. You innovate in the other direction, and that’s privacy, and that’s Zcash. To get to private programmability is actually stacking those 2 together, and it’s actually quite hard.

It opens up all these attack surfaces and so on. So just scale Zcash first. And then there’s Aztec, there’s Aleo, and there are all these other private smart-contract chains. I wish them the best. I want them to have a non-zero-sum view of the world.

They’re taking on a more complicated problem. In theory, they can just do the same thing Zcash is doing, which is private transactions. In practice, if you remember Facebook in the 2000s, people said, “Why does Twitter exist? Facebook has status updates. One feature of Facebook is all of Twitter. Why does Twitter exist?”

Sometimes that’s a good argument, by the way. That’s why Steve Jobs told Drew Houston, “Dropbox is just a feature.” Dropbox is funny. It’s a great company and so forth. But if iCloud was Dropbox, it’d probably be better.

I like both. Both would be better off for it. iCloud is kind of—Dropbox doesn’t have as much distribution as if it was part of a big operating-system kind of bundle. Sometimes people are half right and half wrong.

Dropbox is a great company, but it might have been bigger in terms of percentage value if they’d been Apple’s cloud services, basically. Point is, it’s hard to say whether it’s just a product or a feature, but my strong intuition is that, just like Twitter’s simplicity made its own thing, simple, scalable, billion-person digital private cash has been the dream for 30 years, and we’re finally there.

By the way, I’m not a trader. I just don’t care about trading. I’m early on platforms and infrastructure. There are things you have to not care about in order to care about things. You have to not care about things.

I talk about very few things. Also, Zcash has been around for 10 years. Even the toxic-waste setup ceremony—that’s gone. That got fixed cryptographically.

It’s unusual that it’s been around for 10 years, has a security track record, has a decentralized base of holders, and the cryptography works.

It’s a great place to wrap a wide-ranging conversation on what’s happening in AI and crypto. As always, Balaji, fantastic conversation. Until next time.

Yeah, love to see all the progress there. Amazing what you guys are doing. Excited to be involved in a small way, and, yeah, until next time.

Okay, thank you.