AI doesn't take your job. AI makes you the CEO. The problem is 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 also better at taste and agency?
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 the sense. And that is what AI can't yet do.
What happens when AI really achieves its potential? Will LLM get us to AGI in some capacity?
No, actually the opposite. 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, or the cloud economy where infrastructure takes most of the value, or it's more distributed. There's an argument that the big labs will take it all because they have all the capital, they have the compute, they've vertically integrated. But there's also an argument that maybe they won't because distillation is like 98% cheaper than it is to build a model. Open source catches up and apps maybe control the user relationship. How do you think this economy is going to play out?
Great question. So 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 to distill a large model into something small and it's very hard to stop that because you're stopping queries from coming back. You'd have to somehow detect that. And it's also hard to morally stop it because what do they do? They copy the whole internet and put it into their thing. So talking about stopping the copying, it's like Facebook or LinkedIn stopping someone from scraping what they scraped. Facebook scraped all these Harvard social networks or Google scraped the entire internet, built a Google index. I get why they want to do it, but it's hard to support that.
So the other thing is I think the future is personal, private, programmable because AI is so powerful that you want to use it within the trusted tribe for a variety of reasons. First is it doesn't miss small things in large data sets and things that were effectively secure through obscurity. A small example but an important one is the Jmail thing, the Jeffrey Epstein thing. You can query like this guy had never thought that all of his emails would be publicly indexed and searchable by AI 10 years later. So you can issue queries that will synthesize information across thousands of emails or whatever and build a story right then and there.
So what that means is it's not just surveillance. It's what the French call sousveillance, surveillance from below, or even the German Bentham panopticon where everybody's watching each other. Any information that's in the public gets indexed and then put into these guys where people can stalk each other and so on. And then what that means is the commons becomes a hall of mirrors with all kinds of pseudonyms and so people retreat back to caves and tribes. So within that trusted tribe, yes, if you share all your code within the trusted tribe, you share your whole codebase, boom, you can zip along. And so AI increases productivity within the trusted tribe. But outside the trusted tribe, aren't you getting a ton of AI spam and AI spam emails, AI spam replies, low-quality slide decks that are sent over. People will send me these slide decks and I love AI. And you know my reaction is to seeing AI in a slide deck...
No, actually the opposite. When I see AI text in a slide deck, and 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 the Apple default. Most people don't change defaults. So default AI looks like AI no matter where the level of it is. And so because of that when I see an AI slide deck and it's got a wall of text, AI can generate what I call Lorem Ipsum, but it's Lorem AIpsum. When I see that and you know it's AI text or AI images, I think they're lazy, stupid, or evil.
Lazy because they just hit a few characters and then they throw something over and they didn't, like the Mark Twain thing of I didn't have time to write you a short letter. The whole point is concision is very valuable. So, they're lazy because they didn't actually put in the time to make it concise and so they sent me some blah like 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 versus something that had some care go into it. Or they're evil where they're trying to get something over on me and try to send something that's clearly fake or not properly diligent. And the thing is, if I have that reaction, as one of the most pro-tech people out there, pro-AI, see all the benefits of AI, I can only imagine how mad anti-AI people will be, where they can't see the upsides of a thing. They can only see the very real downsides.
And just to say why those happen. AI does reduce the cost of generation but it increases the cost of verification. And many markets, like for example quickly generating a resume is not that much better than just writing it yourself. But now verifying a resume has gone up and to the right. Because it's something where it used to be that somebody would have to sort of have a certain vocabulary to be able to write a well-done cover letter or resume. And now you have to spend more energy parsing that because they can have a similar chrome of something that kind of looks good. So now you have to very closely read it. You can still do it but you have to spend more energy on verification.
So what I do for example is I fly everybody out for interviews first. I do in-person and I give them proctored exams, offline exams, because they can AI the online and just the credible threat of doing the offline means they don't use AI on the online exam. And so AI is going to create tons of jobs in proctoring and verification. This brings me back to where's the future of AI. I actually think AI makes the internet a lot more like the Chinese internet. You know why?
If you look at the Chinese tech ecosystem, and many Americans aren't familiar with it, I'd recommend, it's a little bit dated now, but read Kai-Fu Lee's book, AI Superpowers from several years ago. The main thing about Kai-Fu Lee's book is it has a history of the Chinese tech ecosystem where, for example, you and me being in tech, we kind of know how Microsoft came up, Apple came up, Google, Facebook, Amazon. We have some idea of the history. And that history is important because there's things that worked in the past that didn't work today and now they can work. The Chinese tech ecosystem is like the Galapagos Islands where many of the same kinds of things exist but in different form. For example, Meituan which is like the closest way of putting it the Chinese Groupon but if Groupon was executing at like hundred billion dollar scale. They're very competent, like if Groupon and DoorDash and so on all became integrated into one amazing app.
The point about the Chinese tech ecosystem is because they arose in a low-trust society they don't have SaaS not in the same way that we do. Instead because if my data is on their servers, well they're probably eavesdropping on me. My data's on their servers, they're probably going to copy my stuff. They just assume that the other guy on their side is going to look at their stuff unless it's like their close friend. And so because of that, everybody codes their own stuff, which obviously has a frictional cost to it, because trust reduces transaction costs.
However, so they have to rebuild, they have to reinvent the wheel over and over again. They have less division of labor and so on. 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? And it does mean that you can build more internal tools with emphasis on internal tools.
And the reason I say that is what I find AI great for as of 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 is wrong like the hands are messed up or something like that in an image. So 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 if you look at two or three faces. Our brains are optimized for checking very subtle things off in faces, but not in crumpled up pieces of paper. That's a pattern of noise that we wouldn't be able to tell. And that also extends to web pages. You can quickly look at a web page that AI generates or a mobile app and you can see if the UX looks janky, which it often does. And then you can see that it's broken there and you can fix it. Also front-end stuff has lower risk than verbal stuff. For the back end, if you are verifying each pull request one at a time, fine. But people who've tried to go full auto on AI, you saw the Amazon thing where they've called all hands because of the outages.
The problem is AI is a shortcut. And a shortcut is good except when it's bad. So the more expert you are, you can use a shortcut. For example, if you just memorized e to the i pi plus 1 equals 0, you could just 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 and how the exponential generates to a function of complex variable and all that kind of stuff.
And so if you, like our generation that is a pre-AI generation, learn all that stuff offline and we can actually use the shortcut because we know how to go the long way around. If you don't know how to go the long way around, AI is a shortcut. Then you just don't really actually know. You can't debug the AI.
And I think the biggest difference between me versus Dario or you know basically like his view of the world perhaps is I think AI is built for the harness at least for now. Maybe he's an amazing engineer and entrepreneur. Maybe I'm wrong. So I put an asterisk on this. But 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 and it does a pirouette and then it says absolutely right. Like how you saw that animated in the physical world and physical AI, the Chinese AI, the robots do exactly what they want them to do and then stop.
Now in the physical world, by the way, that's another thing. So AI for visuals you can verify it with your eyes. AI for certain kinds of backend code you can unit or integration test it and you can review it. AI for the physical world is very verifiable because the thing is the digital world is fundamentally decentralized in a way the physical world isn't. There's only one physical world. So you can say did the AI move this box from this pallet to that pallet. That is something where you can get it to probably 100% 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 one physical world. So eventually all the sensor data, all of that converges on one thing.
By contrast, the digital world, there's all these people who live in their own constructed environments. Harry Potter fanfiction here, Star Wars fan. And so AI is slurping up all of this stuff. And so it's simultaneously it can put you in some secret agent kind of world. And people who have LLM psychosis will talk to the AI and think it's real because a very immersive virtual world that they live in. You know what I'm saying?
So, the other thing about it is the boundary of a digital task is almost always more fuzzy than the boundary of a physical task. Like having a 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. So verification is actually harder in the digital world than it is in the physical world, which means reinforcement learning and training is much easier in my view in the physical world with robots and self-driving cars, drones and so forth. So the Chinese style of physical AI will also be successful.
So AI works for visuals, AI works for the verifiable and AI works for the physical. When it is, one of my rules and it took me a little while to articulate this but four words. No public undisclosed AI.
Why? There's a temptation by many. There's going to be a huge backlash called let's just say no AI. It'll be like a drunk who just wants nothing to do with it. And AI is like, it's a funny way to put it like alcohol. People analogize to nuclear weapons, but I'll just analogize to alcohol for a second. Some cultures simply like they can't hold their liquor. Maybe they lack alcohol dehydrogenase or what have you. And so they just ban it. Because 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 just easier to kind of keep the habit all the time. So, it'll be AI teetotalers that just swear off it completely.
And you know Nate Silver actually had a great line where he said AI for him because he's like a poker player among other things. He's like it's a gamble. Why is it a gamble? Because I have to formulate it and dispatch it to the AI and then verify the result. And often that's slower than doing it myself. And I'm sure you've seen that. The act of prompting and writing it down and then verifying the result. AI doesn't really do it end to end necessarily. It does it middle to middle as we've talked about. And it's very much like do I delegate this to an employee or do I just do it myself, because articulating it out in clean English and hitting enter is sometimes slower than just, for example, if you're describing what to do in a video game, jump over the mushroom to this that, versus just hitting A B C and there and being non-verbal about it. It's sometimes easier to do that way. That's just like a proof of concept. Where you'd be like there's certain kinds of things that are harder to say than do.
Okay, those types of things where it's hard to verbalize what it is. And some people will say, Oh yeah, Neuralink will solve this. The difference is, they'll say, I'll just read your mind and tell you, which is actually it's worth engaging the concept because Neuralink exists. But I don't know if you've seen those things where like they image somebody's brain, there's nothing in there. So the thing is with Neuralink, somebody still has to like form the concepts in their head before the characters appear on screen. You still have to like write the thing in your head. Maybe it'll eventually get to the point that 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 by the way, at least for certain things, bio could be very important. You know why?
Your body is creating all kinds of sensor data. If you look at gene expression data. If you've ever gotten labs back, you've done a clinical lab, right? You get a vector of your bilirubin and hematocrit and so on and so forth, that vector over time is like a table of time series data. It's like small molecules and gene expression levels and so on over time stamps. They might also have which tissues. So it's spatial as well. So it's time versus space versus compound. That's this big. It's not just a cube, but it's at least a cube. It's like time versus tissue versus molecule. That huge stream of data is telemetry that's coming out from your body that could prompt AI without you vocalizing or verbalizing anything.
Okay. Years ago, Mike Snyder had a paper called the integrome. By the way, for the audience who doesn't know, biology is actually, I'm not really, I mean, I'm a crypto guy or I'm a tech guy, but actually before all of that, I'm a biomedical researcher. I was a professional bioinformatics genomics scientist at Stanford and I taught there and I founded Counsyl. We sold that. So, that's actually my true core competency. So if you go back years, Mike Snyder, professor at Stanford, wrote a paper on the integrome and the idea was just throw every test. Now today we call that wearables or quantified self, but more invasive than that because he's doing blood testing and so on. And he just measured it and see what he could figure out. He could see that he was getting sick before he knew he was getting sick. He could detect the antibodies, the white blood cells, neutrophils, whatever moving before he himself had any symptoms. Do you understand what I'm saying? So that stream of data AI could act on that and then you're prompting it non-verbally. You don't have to spend time. So I'm not sure whether AI will be able to read your mind, but it can read your body.
All right. Let me give another one.
Here's a fun one. Okay, I can say this one. Maybe I can say this one. I can say half of this one. All right. Another way of modeling what AI is. So Darius talked about oh AI will be like it's like new countries. Well, I thought about that a fair bit myself. So one way to think about it is AI is like the rise of Asia and India from an American perspective. AI is like Asians and Indians. Why? Because you have like the rise of a billion Chinese and a billion Indians meant that from an American perspective you could get anything done by a physical manufacturing robotic warehouse or by digital outsourcing for some price if you could articulate it to them over that channel. So imagine you've got now a billion factory robots and a 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. And the part where I depart from a lot of people is they think AI will be able to sense markets and politics. But I don't think it will. And the reason is or if it is, it immediately gets decentralized and adversarial. And what I mean by that is like when you're learning whether something is a dog or a cat, the dog isn't like shapeshifting on you and morphing on you to defeat your learning of that. The mapping of dog to the character's DOG is basically constant over time. And so that fits the train test paradigm of AI. Similarly, the rules of chess are constant over time. But a market is set up where if you try the same trade, then someone eventually figures out what trade you're doing and they take the opposite trade. It doesn't keep working. In a stochastic process sense, it's not a time invariant thing. The distribution, it's not time invariant and it's also adversarial. It's multiplayer where whatever move you're doing, somebody else in the market is going to try and do another move.
And that's not to say, the counter argument AI guys will say as well, AI can learn to play adversarial games like Starcraft and stuff like that. And I say, Yeah, but then you play an AI versus AI because you have a decentralized AI. So the other guy on the other side of the market is also using it. And 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. And if you have a generic tool, you're not going to get specific advantage. What you provide to the table is specific. The AI is generic. And similarly, politics is very similar. If you just had the same tweet over and over again, unless it's like weather or something like that, the kinds of things people are interested in change, topics, what's timely, what's not timely.
So AI, one way to think about it is humans are the sensor. AI is the actuator. Humans sense the world. They sense the financial conditions, the market conditions, political conditions, and then they bring that back into a cleanly articulated English prompt, and then the AI does it. Humans are the sensor, AI is the actuator. So it's like a human-machine synthesis. What are people saying? Oh it's all about taste. What's taste? Taste is the sense.
