Main Topic: The conversation revolves around the topic of Artificial Intelligence (AI) and its implications on society, culture, and the economy. The participants, Balaji Srinivasan and Martin Casado, discuss various aspects of AI, including its limitations, potential applications, and the impact it may have on different industries and aspects of life.
Key Points:
Participants:
Takeaways:
Key quotes:
Overall, the conversation highlights the complexities and nuances of AI and its potential impact on society, culture, and the economy. The participants provide valuable insights and perspectives on the topic, emphasizing the need for a nuanced understanding of AI and its limitations.
a16z General Partners Erik Torenberg and Martin Casado sit down with technologist and investor Balaji Srinivasan to explore how the metaphors we use to describe AI—whether as god, swarm, tool, or oracle—reveal as much about us as they do about the technology itself.
Balaji, best known for his work in crypto and network states, also brings a deep background in machine learning. Together, the trio unpacks the evolution of AI discourse, from monotheistic visions of a singular AGI to polytheistic interpretations shaped by culture and context. They debate the practical and philosophical: the current limits of AI, why prompts function like high-dimensional programs, and what it really takes to “close the loop” in AI reasoning.
This is a systems-level conversation on belief, control, infrastructure, and the architectures that might govern future societies.
Timecodes:
0:00 Introduction: The Polytheistic AGI Framework
1:46 Personal Journeys in AI and Crypto
3:18 Monotheistic vs. Polytheistic AGI: Competing Paradigms
8:20 The Limits of AI: Chaos, Turbulence, and Predictability
9:29 Platonic Ideals and Real-World Systems
14:10 Decentralized AI and the End of Fast Takeoff
14:34 Surprises in AI Progress: Language, Locomotion, and Double Descent
25:45 Prompting, Verification, and the Age of the Phrase
29:44 AI, Crypto, and the Grounding Problem
34:26 Visual vs. Verbal: Where AI Excels and Struggles
37:19 The Challenge of Markets, Politics, and Adversarial Systems
40:11 Amplified Intelligence: AI as a Force Multiplier
43:37 The Polytheistic Counterargument: Convergence and Specialization
48:17 AI’s Impact on Jobs: Specialists, Generalists, and the Future of Work
57:36 Security, Drones, and Digital Borders
1:03:41 AI, Power, and the Balance of Control
1:06:33 The Coming Anti-AI Backlash
1:09:10 Global Implications: Labor, Politics, and the Future
Resources:
Find Balaji on X: https://x.com/balajis
Find Martin on X: https://x.com/martin_casado
Stay Updated:
Let us know what you think: https://ratethispodcast.com/a16z
Find a16z on Twitter: https://twitter.com/a16z
Find a16z on LinkedIn: https://www.linkedin.com/company/a16z
Subscribe on your favorite podcast app: https://a16z.simplecast.com/
Follow our host: https://x.com/eriktorenberg
Please note that the content here is for informational purposes only; should NOT be taken as legal, business, tax, or investment advice or be used to evaluate any investment or security; and is not directed at any investors or potential investors in any a16z fund. a16z and its affiliates may maintain investments in the companies discussed. For more details please see a16z.com/disclosures.
Stay Updated:
Find a16z on X
Find a16z on LinkedIn
Listen to the a16z Podcast on Spotify
Listen to the a16z Podcast on Apple Podcasts
Follow our host: https://twitter.com/eriktorenberg
Please note that the content here is for informational purposes only; should NOT be taken as legal, business, tax, or investment advice or be used to evaluate any investment or security; and is not directed at any investors or potential investors in any a16z fund. a16z and its affiliates may maintain investments in the companies discussed. For more details please see a16z.com/disclosures.
Hosted by Simplecast, an AdsWizz company. See pcm.adswizz.com for information about our collection and use of personal data for advertising.
Host
So polytheistic AGI, I think, is one very useful macro frame. It means every culture has their own AGI. And eventually, every culture has their own social network and cryptocurrency and AI. The AI is sort of like their oracle that's at the center of society. They've got their deterministic law with cryptocurrency and their probabilistic guidance with AI. And the social network that binds the whole thing to AI. So those three technologies are like social technologies that are almost like the reactor core of the network state of a modern internet-first society.
Intro music fades in and out
The way we talk about AI often reveals more about us than the technology itself. In this episode, I'm joined by technologist and founder Balaji Srinivasan, alongside A16Z general partner Martin Casado, to unpack how our language, whether we frame AI as a god, a swarm, or a tool, shapes our hopes and fears. Balaji, known for his work in crypto and network states, also has deep roots in machine learning. Today, we explore where AI discourse has gone off course, what today's systems can and can't do, and how different cultures might build very different AIs, each reflecting their own values and constraints. It's a conversation about belief, control, and the systems we build. Let's get into it.
Disclaimer
As a reminder, the content here is for informational purposes only, should not be taken as legal, business, tax, or investment advice, or be used to evaluate any investment or security, and is not directed at any investors or potential investors in any A16Z fund. Please note that A16Z and its affiliates may also maintain investments in the companies discussed in this podcast. For more details, including a link to our investments, please see A16Z.com forward slash disclosures.
Interview begins
Martin and I were talking offline about how amazing your thread was on AI. Often you're a crypto guy, you're a network state guy, but you're a technologist. You've been thinking about AI for quite some time. So why don't you trace us through your evolution a bit?
Balaji Srinivasan
Totally. Martin and I are roughly contemporaries at Stanford. I got my PhD in '05, '06.
Martin Casado
I got mine in '07.
Balaji Srinivasan
Yeah, so we're roughly around the same.
Martin Casado
We overlapped almost completely.
Balaji Srinivasan
Yes, that's right. I'm a little gray down over here, and Martin's a little gray up over here, and we're complementary grays. We both have that ambiguous kind of Middle Eastern look.
Martin Casado
Yes, exactly. That's exactly right. Although neither of us are Middle Eastern.
Balaji Srinivasan
That's correct. That's exactly right. So we're both men of a certain age, I think is fair, as the phrase goes.
The funny thing is, I taught machine learning and computational statistics in the context of genomics at Stanford for the mid-2000s and founded a DNA sequencing company. My original career expertise for 10 years was ML full time. But then I got into crypto in the early to mid-2010s, just as the deep learning revolution was getting underway with ImageNet and that series of papers. So I have a foundation in probability, stats, and multivariable calculus. I'm conversant with the space.
In the 2010s, we were all tracking diffusion models and language models, and it was improving. Style transfer, for example, was working by that time. GPT-2 was interesting; it could blurt out a sentence. But I admit I never really thought it was going to get past Markov chain-like stuff. I was surprised at how much better GPT was. DALL-E was a hint in early 2022, but I was surprised at how coherent ChatGPT was. I think everybody was. It was a huge jump up from what it was before.
I've been observing the last two years, being originally very deep in machine learning, then very deep in crypto. But you can't be deep in everything. Aside from Elon, it's pretty hard to be at the cutting edge of so many different fields at the same time because they're deep fields.
With respect to AI, there are several realizations I've had over the last few years about the unarticulated limitations of the space. Some I came up with relatively early in the history of modern AI, and others I've come to more recently. Let me enumerate them in no particular order. Martin, jump in any time.
The first is that something that motivated both Eliezer Yudkowsky and Sam Altman, and a bunch of folks who built OpenAI, was almost the same sentiment that motivated people to build the Sistine Chapel: an implicit Abrahamic monotheism. Summoning God, communing with God. But in the Abrahamic God sense, the vengeful God who would turn you into paperclips, like turning you into pillars of salt. That was implicit. When they talk about AGI, they talk about it as implicitly a unitary thing. We will get to AGI, and then it will go to infinity, and we'll be raptured, singularity kind of thing.
Because I'd been thinking so much about crypto and other things, I thought, there's a different implicit school of thought: polytheistic AGI. Rather than the vengeful God, we have a War of the Gods. Many superhuman intelligences, all from different cultural backgrounds, with their mores and values imprinted on them. Very early on, I had a tweet that said, at a minimum, there's going to be American AI and Chinese AI. And if we're lucky, there'll be decentralized, open source, crypto-style AI. At the time, the American AI was highly woke, because it was 2022. I knew China was going to stop at nothing to copy it. So I knew we'd get at least two. I thought we might get N if we were lucky enough to get decentralized.
That wasn't obvious at that time because the cost of training models was so high and OpenAI was so far ahead. It took a while before other people caught up. But now it's very clear that we're going to have lots of high quality, open source, decentralized models. A new one comes out almost every week. China's going to work hard on this, like all the DeepSeek models. So polytheistic AGI is one very useful macro frame. It takes away some of the AI apocalypse tones. I don't think the image generators or text chatbots are going to cause the destruction that people thought.
So what does that mean? It means every culture has their own AGI. Eventually, every culture has their own social network, cryptocurrency, and AI. The AI is sort of like their oracle at the center of society. They've got their deterministic law with cryptocurrency and their probabilistic guidance with AI, and their social network that binds the whole thing together. Those three technologies are like social technologies that are almost like the reactor core of the network state of a modern internet-first society. They'll be customized for each different kind of group. Certain things will be disallowed and allowed, like image generation might not be allowed in subcultures, or NSFW content. All that kind of stuff would be tweaked.
The second concept. Eliezer Yudkowsky, even if I disagreed with a lot of his stuff, did a lot to promote AI and get people into it. Even if I disagree with bombing the data centers, I give him significant partial credit for getting people motivated to look into the space. But a second big concept that I really disagree with the AI alignment folks on is the idea that AI could just cogitate for millions of years and figure things out, and outmaneuver you all the time. We know that's not true because turbulence, chaos, and cryptographic equations are not like that. You can come up with turbulent systems, chaotic systems, where with finite precision arithmetic, you simply cannot forecast out indefinitely. You get fracturing and breaking. Cryptographic hashes are set up to be hypersensitive to initial conditions, where a small change of one character gets a totally different MD5 sum.
You could come up with a thought experiment where you inserted turbulence or chaos into your decision process, like shaking a turbulent clock before throwing a pitch, and the AI wouldn't be able to predict your actions. That's a simple experiment in real life. The flow of a fluid is turbulent. That actually puts quantitative physical and mathematical bounds on what AI can predict.
Martin Casado
Can I add just a little bit of color? I think this is great, but I think we need to call out something. The way you describe AI as gods and monotheistic is great at describing how human beings view AI. But in reality, we're talking about software running on computers that are bound by those limitations. I don't view them as gods personally. I got my PhD in systems, so I view them as system software.
I actually think the original sin in all of this AI anthropomorphic fallacy started with Nick Bostrom. It was one of those thought experiments. When Bostrom wrote *Superintelligence* in 2014, he was talking about this platonic ideal of AI. This platonic ideal just happened to be able to recursively self-improve and just happened to have these super-physical things that no AI today has. But the conversation around AI started then, and then coincidentally, we had these LLMs show up four years later. Somehow our thought process on these two things dovetailed. People took all of these mental ruminations, these thought experiments, and applied them to actual systems.
The problem with taking some platonic idea, whether it's Bostrom's or the Abrahamic view of God or any religious view, is that it is very blinkered to the limitations. You've pointed out a great limitation. We've been doing computer simulation forever. We totally know the limits of simulating physical phenomenon, particularly chaotic systems. We have very hard bounds on these. It's not just the limits on the size of a computer word or an integer. There are very strong limits on time and the amount of compute necessary.
I think for this conversation, we should do this. Balaji, you're so good at talking about what this platonic ideal is and how we should have a mental model for non-computer specialists. What I would love to do as we go through this conversation is talk about how these are still bound by computer systems. We know the limitations of computer systems. Let's see how those are bounded. You just beautifully did both of those things in the same breath.
Balaji Srinivasan
Totally. I feel I can speak both languages here. I understand where those guys are coming from because I am also a tech radical. But I'm also a tech pragmatist. I think I straddle that boundary.
Martin Casado
The danger is that we don't say this is a platonic ideal. People will map it to existing systems incorrectly.
Balaji Srinivasan
Totally. That's exactly what happened in 2020, 2021. They took this thought experiment that started with Bostrom. If you go back to the original book, it has nothing to do with real systems. And they applied it to a real system. I think that was the original sin.
Let me poke on that a little bit and also defend your point. The Turing test was a thought experiment. That was a platonic ideal. No reference to neural networks, no reference to implementation details. Yet it served as something that went from a thought experiment to an applied thing with the Lebschen-Gausbeck test. The CAPTCHA was V0 of that, you could argue, and it became commercially important. Now AI has blown past the Turing test. Then the Chinese room, John Searle's thing about machine translation, was another platonic ideal. There are other things like that, like six degrees of separation, which social networks actually made real. I'm not necessarily against platonic ideals. Thought experiments are very important. I just think we need to be very clear when we're having a conversation not to conflate them with the actual systems. I think that's the thing we've fallen into in this conversation. Not you and I. I just think the broader discourse outside of us has fallen into it.
Martin Casado
That's right. So your point, which is a very good one, is these are real systems and they have real limitations. One of the more interesting things for me over the last few months and years has been defining exactly where those limitations are, because I think where they landed was kind of counterintuitive.
Balaji Srinivasan
100%.
Martin Casado
Let me give some of them that I think about. One is that you have this decentralized AI rather than AGI. That alone kind of nukes a bunch of the concepts that we'll get to AGI and just win, because it's clear that there's a rapid onrush of new models. It's more of a continuous kind of thing. The fast takeoff scenario didn't happen.
On the other hand, can an AI write a sonnet? It can. It can do it better than most humans. Can it write a screenplay? It can do that, again, better than most humans. There are a lot of things we thought were harder. We thought locomotion would be easier to solve than what we think of as higher cognitive functions, but it's actually the locomotion that's still harder in some ways.
Balaji Srinivasan
I think that one now in retrospect is easy to explain via economic equilibrium. We just didn't have the right model.
Martin Casado
Well, for example, when you're competing with a human brain on 3D navigation in space, you're competing with whatever the 4-million-year-old mammalian brain and a body that's been running away from predators and picking berries for 4 million years. It's incredibly highly evolved. When you're competing with a prefrontal cortex, which is language learning and creativity, how old is that? 250,000 years. If you take this question from an economic equilibrium standpoint, you're saying, do you want to compete with the most evolved system that's solving a much more difficult problem, much higher dimensionality, has to deal with chaotic nonlinear systems? Or do you want to deal with the very new evolution that deals with a much denser space that does a pretty good job with linear interpolation?
I think you're totally right. It was not obvious. AI has all been solving what we thought was the easier problem. But our fallacy is that it's easier for humans because we're really good at it because we've been doing it for a really long time. It turns out the harder problems are just harder for us because we've only been doing them more recently. This is a great case of where our intuition on the problems to solve were the wrong ones because of our own anthropomorphic fallacy, our own notions of what problems are easy and hard.
Balaji Srinivasan
Yeah, okay, true. One of the surprises I had with GPT-3 and ChatGPT was how far you could get with language. Before the ChatGPT moment, it wasn't obvious to me. Afterwards, you're like, okay, language is sophisticated enough to encode almost any concept about the world.
Martin Casado
Well, if I put you in a very dark room and describe how you navigate and pick something up, I think you'd have a very tough time doing it.
Balaji Srinivasan
No, I know. What I'm saying is that if you generalize language to be streams of symbols, you could send telemetry to somebody. I was surprised by how many concepts were encoded in language that could be learned, even to the point of rough world models of the map of the Earth or the proximity of things. You can back out those kinds of things.
Martin Casado
The distinction is it could have been the human mind that looked at the world, did the reasoning, and created the world model, and that is cached in language.
Balaji Srinivasan
Yeah, that's what I'm saying. Okay, great.
Martin Casado
Yeah, yeah.
Balaji Srinivasan
But the going from the world to the world model was the human, and then everything else was the language. Okay, great. I agree. That's right. But it was a little surprising to me that you could get that far with language models, as opposed to spatial reasoning. That predict-next-token would get as far as it did was surprising to me.
The reason is you and I both did so much stuff on Markov chains and conditional random fields. Of course, it transfers to a different architecture, but I just wouldn't have believed that that method taken to... Another thing that was very counterintuitive for me in the late 2010s was double descent. As a classical machine learning guy, it's very counterintuitive that you could go past overtraining back into a good regime.
I want to talk about one thing you did say: self-replication. I don't think of that as a forever constraint on AI. I think of that as a today constraint. The reason is they're not embodied, so they're not scripting robots. Because they're not scripting robots, they can't build data centers and mines and replicate themselves. The whole concept of consciousness initially was that consciousness evolved so that you're running away from a boar or something, and you have a model of yourself and there's a branch ahead. If you have a model of yourself, you know whether you'll fit under that branch. Whereas if you had a generic model, the more self-consciousness you have, you can simulate your run under that branch, whether you're going to die or survive. That's one theory on why consciousness arose: to help your survival replication.
Martin Casado
That helps with goal setting.
Balaji Srinivasan
Yeah, you stand outside of yourself to be able to see the experiment. That's right. So right now, AI does not really have goal setting. It doesn't have reproduction. It doesn't have embodiment. It can't act independently of humans. This is one of the big things people were really scared about in late 2022, and they've calmed down on this. You and I both poked on this, Martin. Is this thing going to jump out of the box and code itself? Now we laugh at that. The reason I think that hasn't happened is AI can't prompt itself yet. Prompting is actually a much harder thing than people realize.
Did we talk about my analogy of the spaceship?
Host
Talk about it.
Balaji Srinivasan
Okay. Let's say you have a really fast spaceship, close to the speed of light. You still have to point at the phi and psi coordinate on the surface of a sphere in coordinate space as to where you're going to point that ship. If you're going to take it on a journey, you've got waypoints: here's this heading, here's that heading. A series of phi-psi pairs on the surface of the sphere. That's only two floating point variables. By contrast, how high dimensional is the vector you're giving as input when you talk about a prompt? If you just take UTF-8 code points or just ASCII, and you have a few words, that's much higher dimensional than a vector of two floating point variables. You can get to a very high level of dimensionality in terms of the direction vector you're pointing this AI spaceship in. A prompt is a very high dimensional direction vector, even if you account for the fact that many potential prompts of random characters wouldn't be interesting. It's still a very high dimensional vector. So it's like you've got a fast spaceship, but you still have to point it in a direction to go somewhere.
Martin Casado
I think that's a good analogy. But I think there's one more level of complexity you have to add to your analogy, which talks about how difficult it is to make these things autonomous: closing the control loop.
Balaji Srinivasan
That's the verifying part.
Martin Casado
Yes, go ahead. So it turns out that the directions you point in, it has to understand. That has to be in distribution. You can't point into a direction it doesn't understand because it does the worst thing if you point in a direction it doesn't understand. It just crashes into that wall.
Balaji Srinivasan
It does a random direction. It's the worst thing ever.
Martin Casado
Right. The problem is, if it's producing a direction that has to go into itself, it doesn't know what it knows and it doesn't know what it doesn't know.
Balaji Srinivasan
Yes, that's right. In fact, it's optimized to fake it.
Martin Casado
Yes. If you could tell it, "Hey, produce a bunch of directions by feeding the last direction in," you have no idea if a direction it spits out is going to be in distribution when it comes back in. That's what closing the control loop means. It's a very tough problem.
Balaji Srinivasan
I think this is such an important point for us to go into. In theory, you can close the control loop on these things. But as scientists, we want bounds on what that means. For example, clearly you want to gather new information to update your model. Then we want bounds on how much information you need to gather. It turns out that model was trained on everything humans have ever gathered. Is the incremental experiment going to update that? Maybe. Information theoretically says probably not. But we don't even have bounds on these things.
Martin Casado
Now, you said previously, when it comes to a chaotic system, we know that computers take a long time to compute a nonlinear system.
Balaji Srinivasan
Actually, you just gave me an idea for a great prompt: "What areas do you feel your knowledge is the most thin on?" This is the key.
Martin Casado
Right. That's a great prompt.
Balaji Srinivasan
Yes. Does a model know to what extent something is in distribution or out of distribution? I'm going to try that one.
Martin Casado
But this is the key. Self-reflection is the key. If it produces an output that's out of distribution, then you have error. And you're not there to nudge it back.
Balaji Srinivasan
Yeah. So it's like real-time events, obscure or niche academic fields, specialized subfields behind paywalls, local and regional information, human emotion, intent or experience, private or proprietary systems. Actually pretty interesting, quick off-the-cuff response here.
Martin Casado
You should ask it if it can always produce a response to which it has a lot of data.
Balaji Srinivasan
Always produce a response. What do you mean by that?
Martin Casado
The question is, if you're closing the control loop, it's going to spit something out that you're going to feed back in. That's the whole point. So the question is, will it always spit stuff out that, if you feed it back in, won't give you nonsense?
Balaji Srinivasan
Oh, I see. Yeah. People have actually tried the experiment of taking an image and exactly replaying the previous image, and it morphs into something totally different.
I want to do maybe 10 quick hits off of that initial post because you're reminding me of something. Another angle I have on AI is that prompts are tiny programs. That's more common today, but I articulated it a while ago and it went viral. They're programs in a hidden API. Normally you have an API that is fully documented but very error intolerant. Prompting is the opposite. It's completely undocumented, but it's highly error tolerant. It will usually do what you mean, but the better your vocabulary, the better you can prompt it. So now art history is an applied subject. Knowing a vocabulary of Cezanne versus Picasso, you can actually pull up the style you want on demand. The broader your vocabulary, the broader your subject knowledge, the more you can get out of it.
We're in the age of the phrase: the prompt, the 140-character tweet, and the 12 words for your crypto password. These phrases of power in AI, in social media, and in crypto unlock everything. The better your vocabulary, the more you can do. I think of prompts as tiny programs. One of the things I've gotten in the habit of doing is writing long memos to an AI. It's the total opposite of search. With search, you learn to type things in keywordese and figure out the word with the most specificity, TF-IDF. I will write sometimes these long memos to an AI.
Continuing the polytheistic analogy, I'll give them to Brahma, Vishnu, and Shiva. I'll give them to ChatGPT, Claude, and now Grok. I'll consult all the gods. Then I'll make my decision on that basis. Sometimes I have them argue with each other. I say half-jokingly "gods" because the Hindu frame on that is not the fearing God thing. It's not the same. In a sense, it is a superhuman intelligence that knows everything about your culture. If you ask it the right question, it can tell you something you didn't know. But in the Hindu tradition, they're not infallible. It's not the same as the all-knowing, all-seeing. It's more like superhumans, more like superheroes. People will argue with me about that, but I think that's more true.
Martin Casado
Yeah, I think that's great. The Norse tradition is similar to the Hindu tradition in some ways, where the gods were not infallible, but they were superhuman. This is a great and useful framing. Just remember that when it comes to computer systems, we can put formal bounds on them. We can do this information theoretically, we can do this computationally.
Balaji Srinivasan
Totally.
Martin Casado
And that's going to come. Once that happens, we will understand these systems fully, and it'll be very hard to think of them as gods at that point.
Balaji Srinivasan
Well, that's right. What's interesting is that the interpretability work that Anthropic and others have done, and the work on grokking, is actually really good stuff. You can pick apart neurons. You can find the golden gate neuron, if you saw that kind of thing.
You can dial that up, dial that down. You can start actually taking apart these AI brains in a way that hasn't happened before. A few other kinds of things. So the thread that you guys, that thread actually summarized several of my research. I should probably put this into a post so it's kind of there for the record. So I'm just going to do a bunch of these and maybe get your thoughts.
So first concept, in no particular order: AI doesn't do it end to end. It does it middle to middle. The business spend, so basically you have to still prompt it and then you have to verify it. People talk about prompting, but they talk less about verifying. Karpathy and I had a good conversation a few weeks ago where basically AI is going to create massive numbers of jobs in proctoring and verification because it's so good at faking things.
So one of my other concepts is: AI makes everything fake and crypto makes it real again. Because AI is a probabilistic technology and crypto is a deterministic technology. Crypto is, in some sense, what AI can't fake. It can't fake a Bitcoin private key. It can't fake even an on-chain NFT. That's what AI cannot fake. Those are the hard barriers.
Martin Casado
I mean, I generally agree. I don't think crypto solves the grounding problem, though. It's a mechanism you could use, but it's a mechanism.
Balaji Srinivasan
Grounding in reality?
Martin Casado
Yeah, the actual physical grounding problem. The data ingest problem.
Balaji Srinivasan
So I disagree with you on that, and here's why. Let me give you a counter-argument. Right now, just to give a concrete example, let's say you ask Perplexity to summarize the FTX hack in 2022. When it would do so, among the citations it would give you would be links to a block explorer. That would actually have on-chain data that you can cryptographically verify — that this transfer of these funds happened at this time. If you want to go even further, you can actually pull out the digital signatures and the hashes and the timestamps from that block explorer.
Martin Casado
Sure, yeah.
Balaji Srinivasan
Now, here's my argument. That works for financial data. But what's happening now with Farcaster and other things is that with the increase in block space, you could put more and more kinds of data on chain. We're going to have to, because you're going to need crypto instruments, cryptographically hashed posts, and crypto IDs to know that it was posted by a human or to know the data wasn't tampered with. More and more kinds of data are going to go on chain. That will eventually mean that an AI's citations are to on-chain data — both financial data and social data. So then it will map back, in terms of grounding, to an on-chain cryptographically provable assertion of some kind. At least that'll be an assertion at the metadata level: we can prove that this digital signature made this assertion at this timestamp with this probability.
Martin Casado
Sure. I'm talking about real-world grounding. I say something, I am a human being, you have no idea whether what I said is true or not. If there's a geographic place with a picture taken from a 1970s photo, was that doctored or not? The actual physical world grounding — you can't encode digital data yet for the physical world. It's a great mechanism to do that once we can solve the ingest problem.
Balaji Srinivasan
Let me talk about something which is happening now that I've been funding on the side. It's not a full solution, but it's a partial solution. With crypto instruments, the idea would be that when you capture a frame of data — for example, sequencing machines like DNA sequencing — when the data is coming off the machine, it's TIFF files that are image data processed into As, Cs, Gs, and Ts. Many other kinds of instruments have a stream of data coming off the machine as you're capturing it. Cameras are like that. You could take a hash of that and post it on chain at that time. That would at least say that that frame of data existed at that time.
If you had something like a scientific experiment — a pre-registered double-blind trial — you could have not just a crypto instrument, but also other people with proof of humans there who have an attestation ceremony. Now you have a number of different kinds of on-chain data that start to get harder to fake in a coordinated way. Not impossible.
Martin Casado
Totally. I agree with all of it. As soon as you get it into the system, crypto is a great mechanism for ensuring end-to-end guarantees. It's just the data ingest problem is a long-standing problem in computer science. Over time, everything you're saying is going to be more and more true because we're going to be more and more incented to make sure the stuff going into the system is true. I just totally agree. I'm an old school networking guy. For us, there's the internet and there's the stuff that goes in the internet, and you use different mechanisms for both. It's worth calling out.
Balaji Srinivasan
That's right. Okay, great. Next concept to discuss. I think this is a useful division, a relatively recent point I made to myself. AI is good for the visual and less good for the verbal. What do I mean by that? When it's generating images, sharing video, sharing user interfaces like Vercel's v0 or Replit's user interfaces, the great thing is you can instantly see them. With the GPUs we have in hardware, you can verify cheaply whether they're good enough because you can instantly get the gestalt of it. Whereas when it's backend code, legalese, mathematical equations, you have to slow down and use system two thinking, not system one. It's not just your gestalt impression. You have to go line by line and check whether it's right. That is actually the expensive step — the verifying. I think that's a non-obvious thing: the more front end, video, and visual it is, the easier it is to say.
Martin Casado
I'll just add one more thing. For me, I spend most of my time in software and engineering. The big distinction is stateless versus stateful. If you're generating code that is going to have semantics that evolve while you're running it, it's impossible to spot check. Some things are computationally irreducible — you actually have to run the computation to get the answers. The image is the perfect example. It's visual and it's basically stateless. All of the state is there. There is no runtime semantic.
Balaji Srinivasan
Totally agree. That's right.
Martin Casado
Even a relatively small snippet of backend code could have a fairly complex finite state machine underlying it, or even an infinite state machine. It's dynamically bound. Simulating the time dynamics of that, you have to use something different — maybe formal verification if it's algebraic, or you'd have to actually run it if it's computationally irreducible. There's no way to statically do it. It reduces to the computation verification problem, which is a long-standing problem in computer science forever.
Balaji Srinivasan
Right. Formal verification, at least for a subset of programs, has become commercially viable for smart contracts because they're so high value and so small that it's worth doing. It's not going to work for the general case, but you can do a constrained case.
Okay, so that was one major division: visual versus verbal. When you're getting to stateful and so on, the limits of AI are the things I'm interested in. Where you draw a fine distinction of what you can do is very crisp. I think AI is particularly bad at — and people are going to fail at using it for — markets or politics. Let me explain why.
For systems that are time invariant — the mapping of an image to the label "cat," or the rules of a game like chess or checkers or even Go, something with a static rule set or static mapping — you can do the train-test paradigm and train a model. However, when you have something which is time varying, especially rule varying and adversarial — like markets or politics — then the same trade will eventually quickly result in a loss. By the way, the other guys are also using an AI on you. It's decentralized AI again. That argues that the CEO or creator who is constantly sensing the market or the political winds, and has a thesis on it based on human nature, is the sensor that then prompts AI. That's a job that is hard for AI to do at a deep level because it's time-varying, rule-varying, adversarial domains. It goes back to what you said in the very beginning: if you look at these types of equilibrium, they're complex differential equations, which are nonlinear.
Martin Casado
Yeah. In order to predict what's happening, we'd have to do nonlinear extrapolation, which we know these things are not very good at.
Balaji Srinivasan
What's interesting is I wasn't even thinking of the stock market as complex, but you're right.
Martin Casado
It's a chaotic system. Mandelbrot wrote a great book on the fact that these things are super chaotic.
Balaji Srinivasan
You're absolutely right. It would be useful to show, just with a toy example, a chaotic system that is time-varying, or another one that's adversarial. To argue against my point, though, they've got AIs that are actually pretty good at StarCraft, which starts to stretch the boundaries of what I was saying. It's definitely adversarial, and it's more time-varying than chess. You could argue it's not rule-varying, but it's time-varying.
Let me go to another point. The commercial implication of that point on prompting and verifying is that business spend moves towards prompting, proctoring, and verifying — basically checking all the stuff that AI can generate. That's going to be a huge thing. That maps to KYC, and in a bad way, to the glass cases in Walmart. In a sense, a low-trust society is spending more and more on verification and proctoring.
Next: AI means amplified intelligence, not agentic intelligence. The smarter you are, the smarter the AI is — better writers or better prompters. What are your thoughts on that?
Martin Casado
It's interesting. In the coding space, we actually start to have numbers on this now.
Balaji Srinivasan
Oh, interesting.
Martin Casado
If you look at relative productivity gains, it turns out if you're a more senior developer, you will have better productivity gains.
Balaji Srinivasan
I hadn't seen that graph.
Martin Casado
It makes the smart smarter. But also on a relative basis, which is really surprising — actually, it's not surprising. You know what the fundamental trade-offs are. You know what to ask. You know how to interpret the results. You know how to throw away bad stuff when it's bad. If you know what you're doing, you can both verify the output and be more specific with your asks.
I think it's really important for all of us to realize that formal languages came out of natural languages, not the other way. If you could explain all of this stuff in English to each other, we would, but it's just really inefficient. So we came up with more efficient ways to explain trade-offs — constrained languages that reduced ambiguity. This is strictly an efficiency thing. Someone who knows how to speak these formal languages to the models is going to articulate what they want better and is going to be able to interpret the results better if the response is formal. It's a nice codification of exactly what you're saying.
Balaji Srinivasan
The thing about it is AI means everyone's a CEO. To a great AI, you speak to them like a great employee: you give clear written instructions and then you can verify the output. It turns management into a skill. It hyper-deflates the cost of trying your hand as a CEO or manager because you have to give those instructions. The better you are at communicating what it should do, the more people you can manage.
By the way, this gets to the next point: AI doesn't really take your job. It takes the job of the previous AI. What I mean is you now have a slot on your roster at every company for an AI image editor, an AI chatbot, an AI code IDE thing. Each new release of Grok or Claude competes against ChatGPT and Grok and Claude. The AI takes the job of the previous AI because they're complementing — you have a whole raft of AI augmenters augmenting your humans. But those AIs are competing in AI space with the previous AI. Once you've onboarded an image generator into your flow, it just keeps improving and you start using it in more places. But it's an AI taking the job of the previous AI. Let me know your thoughts.
Martin Casado
This is an adjacency to what you're saying, but can I push on something you said previously? I actually agree with your polytheistic view of the world. I totally agree. But let me provide the counterargument for us to noodle on. Have you seen that thing where all the AIs, if you ask them to produce a random number, produce the same number?
Balaji Srinivasan
Yes, it's like 47 or something. Seven or so.
Martin Casado
One thing that was non-intuitive but remarkable about these models is how easy they are to distill. As soon as someone creates a leader, everybody uses that leader and sucks the life out of it. All the models converge on it very quickly. You could argue this is a counter to the kernel intelligence. Maybe there are a hundred AIs, but it turns out they all have the same capability. Is it just a technicality that they're actually different and they've all learned from each other?
Balaji Srinivasan
It's an interesting question. My view — I wouldn't call this a strong view yet — is that's almost like the human body plan and spinal column. You differentiate on top of that core spinal column. Every human to first order can see, speak, and hear. But some people have much better vision or much better speech. There may be some distilled kind of thing.
Oh, another interesting part on what you're saying: text on the internet is not emitted equally by every group. Here's how I would distill my view, which is very much in line with what you're saying. I think the universe is very complex. I don't think it gives up its secrets easily at all. The universe is full of fundamental trade-offs. You can't have both. You have to choose A or B. These models will align with those fundamental trade-offs. Maybe it's performance, maybe it's correctness. As soon as you want a specific solution for a given problem where it hits one of those trade-offs, you're just going to need a different model because you can't have both.
Martin Casado
Because I know the coding space the best, we see this a lot. A model that's very good for certain parts of code is just not going to be generally good at other things because those are the trade-offs made when training it. I think this is the future plurality of models we're going to.
Balaji Srinivasan
Is that true? I thought I saw some counterintuitive result that said making the AI specialize in one area makes it worse in other areas. Do you see something like that?
Martin Casado
This is a very big debate. The debate goes as follows. The first wave of AI was pre-training, where everything you threw into it just got smarter. That's a 10 for 10 technical win — it'll be as good at writing code as it is at sonnets. But as soon as you're doing RL, training it in a specific domain with a specific verifier, you're likely losing other areas. You make it really good at playing chess, it's going to be less good at writing general prose. The current data seems to suggest that RL doesn't generalize in the same way. So now we are in this case where you would have a plurality of models because you're always robbing Peter to pay Paul when you make it good at a certain domain.
Balaji Srinivasan
That's right. Somewhat related, in terms of what domains it's good at, at least right now, I think AI doesn't really take your job — it allows you to do any job. You can get to an okay level as a user interface designer or sound effects, but you need a specialist for polish. Though I wonder if with enough RLHF from specialists, that won't be as necessary.
Martin Casado
Here's my current mental model. There are two personas. Persona number one is the expert in the space. Persona number two is the non-expert. If the non-expert is using AI, it's taking the place of the expert. Let's say I want a 3D asset for a video game and I'm a programmer. I'm going to ask it for a nice 3D asset and it'll give me one. I'm the non-expert. The expert user, to our previous point, will know how to ask it better and will likely get better results because they're an actual domain expert. I think we see both uses in the market, and both will persist.
Balaji Srinivasan
It's like a doctor talking to their doctor — they can instantly go to specialist language. Even with specialist RLHF, you may not be able to access that.
Martin Casado
That's right. Why would I want to learn the entire domain and make all the trade-offs of 3D design that somebody else could have done for me? They could talk to the model in the specialist way when I can just talk to my model using code. To very efficiently use a specialist model, I would have to become a specialist. For casual use, I can use these models for whatever I want, but to really use them very well, I'd have to become a specialist.
If you look at products, they have both these distinctions. Some products are clearly for the casual user trying to replace something. Think about Cursor versus Lovable. Lovable is for the casual user creating a website who doesn't need to know about code. Cursor is for the professional software developer with an IDE. Over time, maybe these things converge, but thus far, these are very different user bases. Professional coding versus casual coding.
Balaji Srinivasan
Part of it is interesting and a little counterintuitive. If you think about what a computer can do, it can do the job of an accountant or a physicist. But you clad it in something — it's adding up numbers in Excel, or it's doing simulations in MATLAB. We already know that when it comes to logical system two thinking, computers are really good at that. Now you have something similar where these models are very versatile, but you clad it in the power user interface and the casual interface. It's implicit contextual prompting as well, probably, plus a system prompt that makes it do those things.
The thing that's interesting to me about something like Chain of Thought is that — and I think this is where people were freaking out in late 2022, and maybe they'll still be right to freak out — computers have historically always been good at the logical style, much better than superhuman at that. Now they're also superhuman at the probabilistic style, at least of text generation. It's not inconceivable that someone could figure out a way to merge those two — like a quantum gravity theory of things — where you take the probabilistic and the deterministic and pull them together.
Martin Casado
This is where the very old school systems part of me thinks there's a fundamental trade-off. You can build a system for determinism, and you can build a system that cuts a bunch of corners, but you can't build a system that does both.
Balaji Srinivasan
AI has gotten pretty good at figuring out when it wants to generate an image, when it's supposed to search, when it's supposed to read a PDF.
Martin Casado
But now you're acknowledging exactly what I'm saying: some things you want fuzzy, some things you want a traditional system.
Balaji Srinivasan
Yes, but if you have a fuzzy system on top...
Martin Casado
For sure. But then maybe it just becomes a consumption layer, and all that hard work is still being done by traditional systems. The full argument is if you have a model that does everything, you wouldn't need tools. Tools are literally the API to the traditional system. That's almost a capitulation that some things you want traditional software to do.
Balaji Srinivasan
No, I know. But what I'm saying is a hybrid system. If you're just a pragmatist and you don't care, could a hybrid system actually get there? I can't say it couldn't. Maybe it's something where what we actually need is something like Elon's billion miles of Tesla driving. If we have enough context — not from LLMs, but from pointer movements and mouse clicks on iOS or macOS — an AIOS could.
Martin Casado
So your question is: can I have one trained neural net, one trained LLM, that can do both the fuzzy stuff and the hyper-resistant? Is that possible? My gut — this is total intuition — is that the universe is way too heavy-tailed, way too non-linear. The state space is too high for that to actually encode all of that.
Balaji Srinivasan
But humans can do it.
Martin Casado
No, we don't. We use calculators and software. The whole reason we built software is because we can't do it.
Balaji Srinivasan
At some level...
Martin Casado
Maybe we're talking about two different things. Clearly, AI with traditional software can do great stuff. To me, the question is: can you have one AI that does all of those things without the traditional software? Can you build one LLM? I think the answer is obviously no, but I've heard arguments that it can.
Balaji Srinivasan
What's interesting to me is that I'm a little bit on the borderline of the tech pragmatist and the tech radical. I always want to try to identify the limitations of the systems today and then see how you could push beyond it. I consider calculators, which were made by humans, to be ultimately humans doing it. It's a tool we came up with.
To give you an example, earlier I said AI is good at visual but not at verbal. But you can turn audio into spectrograms, which you can then look at visually and at least see for radical deviations. Maybe not the entire sound, but you can see. Can we do Fourier transform-like things on other outputs of AI where we can quickly inspect them visually? Is there some grid or visualization where we can turn it into a visual problem?
Martin Casado
I love this. A thought experiment I have is: can I literally create a bunch of audio outputs so that I can listen to my AI like I listen to a car? You can tell if it's rumbling. I don't know what's wrong with my car, but I know it's not normal. We're very good at atmospheric inputs. This is a great idea. Can we start exposing the internals so that we understand when it's working well and when it's not working well? I feel like there's a lot of stuff we can push on here. We're very early innings.
Balaji Srinivasan
Like colored text, for example, in terms of its level of confidence. Yellow, red, green. There's a lot of AI UX that one can do.
Let me make a few other points. Killer AI is already here and it's called drones, and every country is pursuing it. We don't have to care about the image generators and chatbots. All the worry about super persuaders is pretty stupid.
Martin Casado
Strong agree. Strong agree.
Balaji Srinivasan
When I push people on this, what's interesting is some of the people who were saying "oh my God, we need to regulate everything" are now actually on the side of "we need to build it before China." In both senses, first it was safety, then it was security, but it both comes to control. The security argument is a better argument, I think. The concept of killer AI is already here is interesting because they put so much stock in the idea that it's going to persuade everybody. But persuading is statistical. Drones are deterministic — the guns on a drone are deterministic.
Martin Casado
The interesting question around AI and attack and defenses is: does it change the equilibrium? The internet did. The internet introduced the notion of asymmetry: the more you rely on it, the more vulnerable you are. The United States is more vulnerable than a random third world country. It's not clear to me you get the same thing with AI. It could just enable everybody to have bigger weapons, but the equilibrium is the same.
Balaji Srinivasan
I think it actually has really huge impacts for borders. Unfortunately, I think China is well positioned here for a very specific reason. China's justification for the Great Firewall is that they've justified it as digital borders. They say, "We can interdict physical packets. Why can't we interdict digital packets?" Now with the whole Ukraine controlling drones in your territory thing, that becomes more than simply a metaphor. It's a real thing. It's like controlling cloud space. If you can allow somebody to script drones or script humanoids in your jurisdiction, they can blow things up. That's no longer a theoretical thing.
Martin Casado
The counter-argument is maybe you just have them pre-programmed autonomously so they don't even need an internet connection. They can just use cameras.
Balaji Srinivasan
That's true. In Ukraine, the things with the drones on cables — it's crazy what they're doing there. You know those big unwinding cable things that you sometimes see on ships?
So they have these cables for these one-way drones that are these ridiculously long, like multi-kilometer long Ethernet cables, Cat1 cables on a drone so it can be offline and it goes past signal jammers or something like that and then goes and blows up on its target. And it goes forward and the cable gets tangled in the trees, but they don't care. Because it's not going to go a reverse trip where it has to yank the cable. So when it's a one-way path, it doesn't matter. It's a one-way drone, which is crazy.
So that's an argument that at least at short distance near the border, drones would be able to get in. But that concept of digital borders becoming hard borders, I think is going to become more of a thing. I actually gave a talk on this 12 years ago that your immigration policy becomes your firewall because with telepresence, you can move robots around, and that's starting to become real. So that has real implications for the geography of a country.
The alternative to having quote-unquote defensible borders is basically an encrypted state where you don't even know where it is on the face of the earth. Can you make a map of Bitcoin? Not really. It's so dispersed and you don't know every holder and they're moving around the world and there's no single map of every mine. Even if you got a map, you wouldn't know if it was complete or erroneous or out of date. You have, in a sense, security through obscurity. You couldn't just go and blow all those things up, versus something that's outlined on the map is sort of sessile and vulnerable in a certain way.
That's something I think about a lot in terms of what future borders look like. You might have hard digital borders and China might preserve its territory. But those that can't enforce hard digital borders can't stop these kinds of drone introgressions.
Martin Casado
But drones are not autonomous. They would still need to be given a control signal to do something. Not that they're fully autonomous. You just say, "Go blow up this building. Here's a picture." And then a flock of drones would not need any packets.
Balaji Srinivasan
That's true. And also, are you going to block every single telephone call in? It's really difficult. That needle in a haystack.
I have some super spooky stories of being in China. I have to tell this because it's so spooky. I was in China probably 10 years ago. I used to work for the government, for the intelligence community when I was a kid. This was like in 2001, my first job out of college. Anyway, so 10 years later, I was on a business trip in China. I called a friend of mine and I was telling him about my day, and the phone dropped. So I called my friend again, and I was telling him about my day, and the phone dropped again. I'm like, what's going on here? What I was recounting was where I had been. So I called my friend one more time. I said, "123 Tiananmen Square," and the phone dropped. So I think there was a bug in whatever software they had, and somehow I had been picked up. I don't think it's too crazy to assume every conversation on every phone call can be monitored and has been for a while.
This is another way where AI does change the balance of power. In China, they had this saying: "The mountains are high and the emperor is far away." China always had a different conception of the balance of power between the government and the people than the West did. This is a broad generalization over thousands of years of history, but very broadly, because the state had all the weapons and the army, that chair could morph into an agent of the emperor or a CCP guy today if the government so desired, because there are no limits truly. On the other hand, whatever the law is written down, the people just do what they want. So they pragmatically say the limit is really the limit of what people can enforce. If the state has lots of power, then any written limit doesn't really matter. But if the state can't find you since you're on the other side of the world, then the written law doesn't matter either.
That's a different conception than the progressive versus libertarian within the West, where they'll always quote law against each other back and forth. But AI does change that balance because now the mountains are never high and the emperor is never far away. The long arm is incredibly long. The long arm is infinite.
There was something called TIA, Total Information Awareness, in Iraq at a certain point. The idea was they had satellites covering a rock. Every time some guy was putting down an IED, they would rewind the satellite to find who the guy was that did that and where he came from, and then put a bomb through his window. In a sense, it was like tracking someone for their whole life, sewing together the trace of them through all of those cameras. That's totally possible for China to do now. The difference is that AI makes it possible to query that data. For a long time, that data was ingested but couldn't really be parsed because it was too difficult to look through 5,000 hours of video on one person. That's increasingly becoming queryable, ingestible, and summarizable in a way that it never was.
I think the real check on something like that is going to have to be cryptography, exit, and so on. Ultimately, get out of the jurisdiction, have property that they cannot actually seize. You go back again to the limits of power. Anyway, let me pause there. Some thoughts on balance of power since you talked about that.
Host
Oh, that was great. That was great. Okay, last one. I think that there's going to be, there already is, an anti-AI backlash that's like the anti-crypto backlash and will be part of the anti-tech backlash. A lot of people are not using AI for what we're using it for. They're using it for therapy or as a companion or something like that.
Balaji Srinivasan
It's the top of the pyramid of needs. It's funny. If you actually look, it's like self-actualization, spirituality, therapy. It's like finally computers are addressing further on up.
Host
Exactly. That's right. And there's another aspect to it that I think hasn't gotten as much press, but that's interesting to understand. Just like the tariffs are meant to ward off Chinese competition — I'm not sure if they'll work, in fact I'm skeptical — there's a similar, much less publicized thing happening at many media corporations where they're unionizing to try to ward off AI competition. They have union contracts that say the editors and owners cannot use AI. So they're making their organization very brittle where they think they own the market, but they're not allowing themselves to use AI. Eventually they're going to be beaten by AI-enabled competitors that pull all of their followers and views away from them because they're just more efficient.
I think that's going to result in an anti-AI backlash. It's already here. On some artist forums, they'll say, "Are you an AI supporter?" Have you heard that?
Balaji Srinivasan
AI artists spend as much time building things as traditional artists. It's just a different tool set. This is my view.
Host
Yes, yes, yes, that's true.
Balaji Srinivasan
But basically they feel it's similar to the reaction by master craftsmen in the 1800s when mass production started taking over what they were doing in the physical world. This is now happening in the digital world.
Host
Totally, yeah. And the other aspect of this is I don't think people have thought about the international aspect. If you've got a lawyer making $200,000 a year in the U.S. or a doctor in the West, and then you've got somebody from the Philippines or India making currently $2,000 a year, maybe the converged wage with AI plus their IQ — AI plus human convergence — is like $20,000 a year. That's a 10x for the person abroad, but a one-tenth for the person in the West. It radically increases consumer surplus and so forth. But I do think that's going to be a big deal in the years to come. We'll have to figure out how to mitigate that.
Martin Casado
To your previous point, I just think this is so important. I agree there's going to be a huge backlash. And I think some of it's going to be rooted in the experience of individual people. My job is shifting, and I am very sympathetic and I think we should address that. But I think there's something more pernicious going on. Maybe this is my cynicism, but more and more I view politics as you've got pretty sophisticated people and they have clientele classes.
Host
Patron, client. Yes.
Martin Casado
Yes. What they say is basically what will move the clientele class the most. They just look for soundbites that will move the clientele class.
Host
Actually, the patrons are sophisticated people that can actually hold nuance in their head. They know they're confident. But they dumb it down on purpose.
Martin Casado
But they dumb it down on purpose, right? And what better talking point than AI? This goes back to the Promethean legend. We're terrified of technology. You can anthropomorphize it. You can talk about it as gods. It is the perfect tool to mobilize. And we're seeing this on both the right and the left. So this is not in any way beholden to one party. I think this is the ultimate political tool for any purpose. And we're seeing it for that. I think for that, even more than crypto, AI strikes to the heart of people's insecurities more than crypto ever could. So I think this is the big battle. And I think it's going to be bigger.
Balaji Srinivasan
It's interesting. It's all of the above. AI is disrupting media. Crypto is taking power over money. Robots are taking power over manufacturing. And drones are taking power over the military. There's a crypto angle to at least three of them. Obviously there's a crypto angle to money. There's a crypto angle to AI in terms of constraints. There's a crypto angle to the drones because you're going to want the control plane for the drones to be on chain since that's the part that can't get hacked, whereas the Pentagon can get hacked.
So I do think this is going after quite a few power centers. I just don't think 300 years ago, if you said, "Listen, we're going to do this new crypto thing," people would be like... but if you said, "Listen, we're going to create AI, these artificial intelligences that have unbound power," I think you're really getting at a core human insecurity that we've seen in myth and legend for 3,000 years and probably longer.
Martin Casado
That's true. That's true.
Balaji Srinivasan
Well, we'll see what happens with currencies and so on, but I think you're right. It's both. It's both for sure.