Seth Rosenberg:
Hi, I am Seth Rosenberg. I'm a partner at Greylock and the host of Product-led AI, a series exploring the opportunities at the application layer of AI
Balaji, thanks for joining today.
Balaji Srinivasan:
Great to be here.
Seth Rosenberg:
So as everyone knows, Balaji is former CTO of Coinbase, author of the Network State, and to me, the first person who made me aware of the implications of COVID before the mainstream media was willing to admit that this was real.
Since then, I've been a huge fan of your thinking, Balaji. It feels like you're always two to three years ahead of the current thing, and also just a very independent thinker in terms of technology's implication on the world. I think it's midnight in whatever network state you're operating in.
Balaji Srinivasan:
Yes, I'm in the Asia time zone, so it's late over here, but I’m always happy to talk tech, so let's do it.
Seth Rosenberg:
I’m very excited about this conversation, and I think a big part of it will be exploring some of the edges around this intersection of AI and crypto – which are two buzzwords that create a third – but I think there's a lot of actually interesting edges here to explore.
So I've heard one of your quotes, which is – and I'm maybe butchering this – but that AI (especially generative AI), is great at making things up and crypto is amazing at verifying what's real. And so I'm curious – thinking about crypto as kind of a balancing force to some of the unsolved issues with proliferation of generative AI – I'm curious to explore that as we get started.
Balaji Srinivasan:
Totally, yeah, and so the quote, by the way is, “AI makes everything fake. Crypto makes it real.”
Again, just so people don't feel, oh, Balaji’s just inventing himself as an AI guy. I actually taught stats and machine learning and neural networks at Stanford in the mid-2000s, and in fact I did bioinformatics and machine learning on large genomic dataset in the 2000s and early 2010s before I got into crypto. So I was a machine learning guy before I was a crypto guy. That was before the era of deep neural networks and the transformer in the modern era. There's certainly a bunch of new tricks, obviously, that have come out since then, but I was at one point very much into the space as a practitioner. I kind of wanted to train a model myself, but it's so expensive to do it, so I'm just messing around with Mistral and other types of stuff right now, but for fun anyway.
So, AI makes everything fake. Crypto makes it real. Again, what I mean by that is even though you can talk about this in a very buzzwordy way, they actually are sort of dual technologies where, for example, AI is probabilistic and cryptography and crypto is deterministic. AI is – at least currently with generative AI – is largely perceived as a front-end technology and crypto is really something where it doesn't produce aesthetics. It doesn't produce amazing visuals like diffusion algorithms or fancy text like LLMs. It's really a backend thing. AI is also something which can bust capture and it can convincingly fake a lot of things that humans do. And so what crypto gives you is proof of human, it gives you actual digital scarcity when AI gives you abundance (and in some ways overabundance, right)?
AI makes everything fake. Crypto makes it real again. AI is probabilistic, crypto is deterministic. Crypto gives you digital scarcity where AI gives you abundance.
And so these are new tools where AI on its own will just flood the web with spam, but crypto can allow for the signed web (on Farcaster for example, where every post is effectively cryptographically signed with your crypto wallet). And what that does, it gives a digital signature and you can map that back to a proof of a human, and you can see, oh, this human actually published this as opposed to a bot. How do you know it's a proof of human? Any individual proof of human and signal can be corrupted, but then you can actually use machine learning once again to kind of reunify that.
So there's kind of a dance between these two technologies. So they complement each other in really deep ways, and so I don't think it's one versus the other, only they also complement each other. The other thing is that when you talk about what can AI do, I am currently – and I may be wrong about this so I'll put an asterisk on it – but I'm currently of the belief that at least the breakthrough that we would need for AI is for it to operate in highly time-varying domains.
So for example, the other day somebody had trained an AI on some game, I think it was some multi-user dungeon game or something like that, and they suddenly found that the score dropped dramatically on some days and then rose on others, and eventually they found that on that day it was a full moon. So every two weeks or four weeks (because this was when the game was played when a full moon was there), the rules changed and the policies that the AI had learned for the last 13 days were no longer applicable on the 14th or the full moon game. And that's a small but very quantitative example of where time-varying rules are something that when it’s out of the training set, it’s something that AI can't learn.
Now the thing is that with generative AI, if you take, and you'll see my point in a second, if you take a text corpus, well, English doesn't change that much over time, so you can learn rules on it. You take an image corpus, the labels don't change that much over time so you can learn rules on them. Or if you take a game like chess, and this is the whole DeepMind school of things, you take a game like chess or Go. Again, the rules don't change in a sense. Even if they can have very complicated outcomes, the rules are fixed. So in a deep sense, there's sort of static environments. Even the laws of physics, they're time and variants that say F=ma, but that law doesn't change with time. Maxwell's equations, the equations themselves don't change your time even if the waves will wiggle and have time to rot us. So you can still train, for example, a drone to fly because the laws of physics don't change with time.
Now what's a domain where the law has changed with time? It’s markets: the same input does not give the same output. In fact, it's an adversarial context where people are constantly trying to change the rules on you so that the same trade will not result in the same profit each time. Maybe there's micro market things like technical analysis type things like Renaissance famously did very well for a long time, Jim Simons just passed away. Maybe there's something there that you can learn that's some universal rule, but if it was disclosed, it would no longer even be a universal rule.
Seth Rosenberg:
But, is it the rules that are changing with markets over time or is it just that they're very dynamic?
Balaji Srinivasan:
I don't know. I don't know how to express it, but I would say… I'd put it like this: I believe they are time-varying in a fundamental way that is different from the laws of physics or from the laws of English that LLM is learning.
Seth Rosenberg:
Another way to think about it is past performance is not indicative of future results, in the sense that the game changes over time.
Balaji Srinivasan:
Exactly, that's right. There's this game called Gnomic. Do you know what that is?
Seth Rosenberg:
Yeah.
Balaji Srinivasan:
It's a game where the name of the game is to change the rules. It's kind of like Congress, right? The name of the game of getting, if you think about it, it's a meta game where an election is a game that elects you to Congress and now the game is to change the rules of the game, including you can gerrymander and change how you were elected in the first place. So that's an extreme meta game where you're not just operating by the rules on the ground, but you're getting elected to change those rules and so on. So these kinds of things like politics and markets, the same input doesn't give the same output. The whole point is also that they're adversarial and people are trying to change the rules on you.
And my view is, I'm not saying it's impossible for an algorithm to do it – AI is actually very, very impressive – but I do think some new concepts are needed maybe, or something like that.
I think that the current, the two or three major paradigms for AI right now, diffusion to algorithms, I don't put in here because it's more on the image side, but the LLM style and the reinforcement learning style of deep mind of the LLM style of OpenAI and others, they don't work in highly time-varying domains like markets and politics. And because of that, that means at a high level, what are you doing as a founder? You are sensing markets and also nowadays politics. So at that high level you will have to sense those domains and then you give directions to your robot army, your AI army, so to speak. So that is at least – I may be wrong about it again, maybe I'll look back in five or 10 years say, oh, that was so stupid. But I think there's something potentially fundamental there, which is because people might ask, why wouldn't it replace the top level as well? Why wouldn't it be able to goal set? And I think there's a top level set of things where you're perceiving very highly time-varying domains that we don't yet have a good algorithm for learning.
I think the current major paradigms for AI right now...don't work in time-varying domains like markets and politics.
Seth Rosenberg:
One thing that's interesting about that thread is one thing that's enabled AI is just the digitization of information over the last 20 years, and I feel like crypto is actually enabling the digitization of markets in a more accessible and fundamental way that creates a training set that is, you could say, analogous to the last 10 years of the internet is the next 10 years as we move to a more cryptographic financial system creates a whole new data set for these AI algorithms.
Balaji Srinivasan:
True. Except the thing is that when you have an AI doing, if you've got an AI trading bot, you also have another AI trading bot on the other side, and now you probably get back to some level of unpredictability because if they're both playing to win, it's like playing AI chess versus AI chess. Now it's back to becoming hard again. So basically the test of this, the Turing test would be the earning test. Can your trading bot just make money? And maybe the answer is yes, maybe that's happening and maybe people are running them and not disclosing it. The cool thing about crypto is you can literally take an AI and have it earn crypto for you if you want to be the next Jim Simons. There's a lot of arbitrage.
Seth Rosenberg:
That’s true. The markets are programmatic.
Balaji Srinivasan:
Yes, completely programmatic. Yep.
Seth Rosenberg:
That's the other interesting layer. Not only does crypto create a new data set – a more digitized and accessible data set for financial markets – but it also creates the execution layer.
Balaji Srinivasan:
Yeah.
Seth Rosenberg:
Which I think is really interesting.
Not only does crypto create a new data set – a more digitized and accessible data set for financial markets – but it also creates the execution layer...I think to actually move to the next level where agents can execute real-world tasks, there's almost no alternative to crypto being the backend.
Balaji Srinivasan:
The thing is, from somebody in 20 years or something in the future, in some ways they'll understand the controversy over crypto as they'll get to it, but another way they'll say, can you imagine that people could only send wires during banking hours 40 out of 168 hours a week? So you go from less than one nine of uptime, 25% uptime or less to 99.9% uptime. That alone is the intranet notification of financial markets and settlements. That alone would justify the whole thing. And then you add the fact that it's got an open backend, and that it's programmable, and that it works on any device, and that you can set up an account in fractions of a second, end to end, all that kind of stuff. It is exactly a total backend change and it changes exactly how you do things.
I mean, it's really just… if you think about the internet itself, it's taking all forms of information and putting them online movies and music and books and newspapers, what the blockchain does, it takes every form of scarcity and puts it on chain, which is currency of course, but it's also stocks and bonds and so on and so forth.
And now that we finally have some of the political breakthroughs, that's I think going to be a very important moment where we finally start to get true crypto equities and all the efficiencies you can get out of it when the old system just concedes. Somebody said something very good, which was that the anti crypto movement is disproportionately concentrated among older representatives, older senators, older congressmen that just don't get it right. They're just not internet-first enough.
Seth Rosenberg:
It does feel like in this world of AI agents that are autonomously executing tasks, you need an open protocol for finance in order for it to actually do things. You have an open protocol for email, for example, so agents can read documents and send emails, but I think to actually move to the next level where these agents can actually execute more real world tasks, it feels like there's almost no alternative to crypto being the backend.
Balaji Srinivasan:
Yeah, that's right. Because the fundamental thing is with crypto, a program can hold money, and here's an analogy: Before the internet was in widespread use, you had to have a human assigned to a telephone number. The first order you had a phone book and every telephone number had a human name next to it in the white pages, or you had a business name, but basically there was a human who owned that business. Afterwards you could get a machine that had an IP address and for example, you could have a shared computer at an Apple store that many people used, and it wasn't just assigned to one person, it wasn't.
And that concept of an IP address that a machine could just have, we take it for granted, but imagine if every single server in a server farm had to have one human associated with it in order to send or receive messages. It couldn't just be handed as a machine from one party to another. That's the same as going from a human-associated bank account where there must be a human name on the account to a crypto address that a machine can send and receive money to autonomously, or an algorithm can send and receive money autonomously. You've actually unhooked it in a very fundamental way, and that enables new things. For example, you can have a program spin up a hundred different addresses in a fraction of second, whereas to open a hundred bank accounts would take weeks, months if they even let you do it. And so just that alone is a complete transformation of what's possible on the scale of going from physical mail to email.
Just to give you a small example, why would it be useful to open up a bank account quickly? Well, think about how helpful it is to be able to just quickly set a website to test an idea, right? Small example, I saw something, you saw how Phantom wallet the other day is now above Square cash? Even that's consumer crypto. That is the number one free app financially in the app store and Phantom is just a very well-built app (I have no stake in or anything like that), it's a great company. I probably should have invested,\. But the thing about Phantom Wallet, it's so easy to use that there was a bake sale, not exactly a bake sale, let's say a bake sale equivalent, a kid's thing that was at the other day, and all the kids could just boom like this on their iPad, set up Phantom Wallet or their iPhone, set up Phantom Wallet and just like a bake sale, just start receiving payments right then and there.
Now to send the money out, you need some SOL. Okay, fine. That's still kind of a pain, but to receive it, you don't. And if you think about how much ceremony is involved with setting up an account to receive the legacy financial system does not make it easy to just quickly receive payments. You have to have an entity and you have to get a bank account or go to the bank and then you can receive things through a stripe thing and so on and so forth. That fiat system is just so slow here. This is back to the future. This is put-out-a-box, accept cash. It's that simple, except it can also now be done from the other side of the world or from the other side of the room and an arbitrary amount of money and then all the digital benefits.
So just instantly being able to set up a bank account, that alone is another. It's like the uptime thing we just talked about going from less than 25% uptime to four nines uptime going from days or weeks to set up a bank account to fractions of a second. There's many different ways in which the crypto system is not just like 10 x, but multiple origins of magnitude better than the legacy system.
Seth Rosenberg:
It’s almost binary. These agents almost can't operate in the financial system, which unlocks so many real-world use cases, without crypto as the backend.
And I think a third thing, which is similar to many areas of crypto, is just beyond the technology, but just the coordination and incentive system. There's no technical limitation for a bank not to accept money 24/7 or not to have open APIs, et cetera, but it's an incentive issue, which I think likely won't be broken unless there's a parallel system that's built.
Balaji Srinivasan:
Yeah, this was something people used to say back in the day, they used to say, oh, the banks will just adopt all the crypto innovations. And I'm like saying Blockbuster will adopt all of Netflix's innovations. First of all, they should have, but it's like saying, I mean every incumbent will adopt what the disruptor is coming up with. Christensen talked about that with the innovator Shama, the late great Clayton Christensen. There's a reason that the incumbent doesn't do it one, and number two is one of the reasons that incumbents, especially in the financial space can't do it, is that they are hamstrung by all these rules and regulations and also they don't have the upside for doing it, right? When you take a risk on doing something, it upsets somebody's apple cart. Somebody often doesn't want that level of real time or that level of transparency or that level of uptime.
Seth Rosenberg:
Or breaking the monopoly.
Balaji Srinivasan:
Exactly.
Seth Rosenberg:
When you custody, your equity is at a certain bank or brokerage, they have a monopoly practically speaking on the financial services they can give you because no one else can collateralize those assets.
Balaji Srinivasan:
Exactly. That's right. So why would they want you to even be able to withdraw it? Why would they want any of this? Right? So it's actually, I mean, cryptocurrency and what it really represents is so amazing in terms of a genuinely bottom-up grassroots form of the financial system globally, right? It's actually amazing when you really zoom back. I mean, just to contrast it, obviously mobile is amazing. I'm a fan of everything that Google and Apple did with iOS to Android, but the thing about mobile, or even AI, is they came to some extent top down, right? In the sense of mobile with iOS and Android came from Google and Apple, which had huge name brands by that point. And AI came top down from, at first, the research labs at major universities and then the big tech companies.
Seth Rosenberg:
Although OpenAI was a good startup, I know they had a lot of insider support, but yeah.
Balaji Srinivasan:
True. But I would say it wasn't as grassroots as I would say social and crypto were more grassroots in terms of how they started. Of course, Facebook eventually became a big company, but in the early two thousands, social was not something that was pushed top down by Microsoft. So social and crypto I think came bottom up and mobile and AI came more top down, which is fine, just different immersion modes.
Seth Rosenberg:
I think this is an interesting thread of crypto as an accelerant to AI, as we were just talking about new data sets and enablement on the execution layer for anything financial or asset related. But then crypto also has a balancing force to AI. And so going back to proof of human and identity forecaster as an example, what's preventing just focusing on crypto as a proof of human mechanism? What's preventing that from being rolled out more broadly? Because obviously phishing attacks are one of the risk vectors that AI enables. Where I call you, I pretend to be your father asking for a hundred thousand dollars wire for an emergency medical bill. It turns out I'm actually not your father. Maybe cryptography can become cheap and accessible enough on that phone call that it can actually verify that I actually am your dad. So what's preventing us from that world?
Balaji Srinivasan:
A few things. First is, one thing I'd say is I'm not a shoe horner. That is to say if a technology doesn't work in an area, I wouldn't do crypto plus AI for the sake of it. So with that proviso, two or three things. First, there are several hundred million people who have crypto wallets. I would say we would have even more if literally the government hadn't been fighting it continuously for the last three or four years, but it's an amazing achievement to have as many as we do. So I think first of all, post legalization – when I say legalization, of course legalization is gradual in many ways. This is a replay of the crypto wars of the nineties where the question was could strong encryption be in the hands of individuals? And that was a hard fought battle in the nineties before you or I, that's old enough that Anderson lived through that. Of course we lived through it, but he's conscious, a professional actor. The point being that now with greater legal support with now both presidential candidates falling over themselves to be pro-cr, we're going to see a lot of things change very quickly. And I think you're going to see a lot more companies step in with the water being warm. I mean now that BlackRock and every major Wall Street hedge fund is announcing, they've got stakes. This whole thing about, oh, it's illegal or or whatever. It's funny, a few months ago when the Bitcoin ETF came out, one of my friends, not friend's, colleague, whatever, just a guy I know from overseas whatever, was like, oh, hey, B, did you see Bitcoin is legal now? And I was like, okay, LOL. This kind of person who's like, what kind of Bitcoin should I buy? And you have to remember anybody who's outside of space, it's funny to us or whatever who are within tech, but it's kind of like.. I don't know..Okay, I don't care about watches. And if somebody said, what kind of watch should I buy? You could probably name a Rolex. There's all these fancy watch brands or whatever. If you said, what kind of Rolex should I buy? The watch aficionado would say, oh my God, you're such a Philistine. You're so ignorant. So of course I cut people slack. If they're not interested in a space point being though the entire space is being held back in many ways, especially in terms of distribution and integrations with a legacy system by its perceived risk or legality or whatever, that's all going to change quickly over the next, well, knock on wood over the next few years if all goes well, but it looks like the momentum is good right now.
I'm not a shoe-horner. If a technology doesn't work in an area, I wouldn't do crypto-plus-AI just for the sake of it.
And if so, then you start getting, I mean the prerequisites have been put in place.There's end intent encryption on many, many devices. Apple has something that is extremely crypto in terms of its vault feature, that’s on-device, right? That's something where you get the private keys, right? One password and others have crypto-like things with the pass phrases. And in a sense, if you take all of the stuff for identity, I mean AI, keys, passwords, crypto, private keys, your literal keys to your car, your passport, your driver's license, your login into your email, all of those things, your Google login, all those things kind of combine into a hardware dongle or set of crypto private keys that become your digital password into everything.
If you look at Cpass, some of the more advanced countries that have digital passwords, you can kind of see where that's going. So then when you have that, well, that solves a lot of the AI problems that are arising where with Cpass, for example, or Estonia thing, you can sign into a website with your identity if you want.
Now, of course this opens up a new issue, which is, wait a second to beat what happened with the item? Am I giving up all my privacy? And now you need another crypto tool, which is called zero knowledge, and it's actually all in your list of things to talk about. And so what you could do is just prove you are a human, for example, without proving anything else about yourself or prove you are an Estonian or a Singaporean without proving else if the prover circuit is set up correctly. So we can also get privacy back with crypto.
Seth Rosenberg:
I'm trying to think of the different barriers for a zero knowledge, cryptographic identity step for things as simple as email or Twitter or Facebook or Instagram or even the telephone one is distribution. What are some of the other kinds of barriers before these more centralized services will reach a point where they actually require this, right? Because even if it exists, we're still kind of bottlenecked by decentralized services requiring it.
Balaji Srinivasan:
But I mean, the thing is a lot of them are crypto adjacent, like Twitter flirted with, they put NFTs in their thing. Instagram had NFTs for a while, Google Cloud and AWS and Azure all have crypto support. They're all aware of it. And it's something where the way the crypto kind of works is it surges, then people try something, and then it crashes and they ignore it for a few years and it comes back and we're on our fourth or fifth cycle depending on how you count. And often things that just are barely there at the peak of one cycle, they kind of get polished or what have you, and then they come back again three or four years later, like CryptoKitties with NFTs and then that crash, and then it came back with a huge NFT boom and that crash, and now that's coming way back with micro NFTs and with Salona making it really cheap to send lots of little NFTs.
So in the same way, I think a bunch of companies that have sort of flirted with crypto before will integrate it more. One reason a long time ago, I dunno, this sounds like forever ago in crypto years, but 2018, 2017, when ENS was first coming out, I knew that eventually crypto logins would work for one very simple reason. You know what that was? Because when you log into a website with your ENS, you also have a balance.
Seth Rosenberg:
It's the combination of a login and connecting your credit card, which is obviously valuable for the application.
Balaji Srinivasan:
Yes, right? If you think about that, sometimes you can know that a problem is hard when you realize it hasn't been solved yet. So Google and Facebook, they have, whatever, a billion numbers of people who log into websites every day with Google login and Facebook login. These are very sophisticated commercial actors, and yet it's just the email account that you log in with and that's the only thing you can give to the site. They have somehow not managed to be able to link Google Pay or Google Wallet or Facebook credits to the site on the other side. Something is holding that back, whether it is regulations, whether it is the difficulty of moving money, something has made it difficult for them to have a thousand Facebook dollars or whatever in your account around the world. There's an unobserved, gigantic boulder in the way. These very sophisticated commercial actors not having a balance associated with those accounts.
So just that, I mean Apple is the closest to this where they have Apple login and an Apple pay to pay. They're probably the best version of this, but even still, it's hiding behind this credit card thing and it's the best version of the old world, but it's still not fully integrated. It's not like you can load the balance into a site and so on. You can do it, it works well for shops and stuff, one-off payments, but it's not like zing, zing, zing. I can pay as quickly as I can tweet or something.
So another thing that's sort of like that, by the way of the unobserved observable, this is a point I made 10 years ago, but still relevant. How many websites have you gone and put your credit card into? A lot. A lot. Okay, those are websites you're paying. Okay. Now, privacy.com, by the way, is really good tool for guarding the zillion credit card billing events or whatever, right? Okay. Now here's a question, a second question, which is how many websites have you put your bank account into to get paid out?
Seth Rosenberg:
Very few.
Balaji Srinivasan:
Yeah, exactly. So I dunno, five, or 10, right? But that's actually kind of surprising. Every website you're paying in, but payouts seem to be very hard on the internet. Well, for one reason, if you put your bank account into a site, they can actually debit from it. So people are conscious of that.
Seth Rosenberg:
Which is crazy.
Balaji Srinivasan:
Which is crazy, right? It's just insecure by design. The whole thing was designed for an offline world. So just again, we talked about uptime going from sub 25% a week. We talked about the speed of opening a bank account. How about the security of it? Basically your account, your crypto address is more like your email where people can send to it. They can't just debit from it, whereas your bank account is insecure by design. And just another example where crypto is just not 10%, but 10,000 x better, okay? Point being something, there's some force, unobserved force that is making it difficult to execute payouts on the internet. Those companies that have managed it have become massive companies. That's, for example, Apple and Google with the App Store and Google Play, that's Facebook with the platform to some extent. That is Amazon with affiliates that is Uber and Airbnb with drivers and hosts, and that is now Twitter with its payouts that's using Stripe for. But in general, payouts on the internet are actually pretty challenging to do. Even still, crypto makes that easy because you can instantly set up a new account on a site and get paid out there.
Seth Rosenberg:
Yeah, that's interesting. I mean, the other dynamic obviously is that businesses want my money and they don't want to give me money.
Balaji Srinivasan:
Oh, true. But crypto means instant jobs. The thing is with Uber, Airbnb, you're doing a job for them and then they give you a cut of it. Why isn't there more of that? The answer is it's actually really difficult to just quickly pay somebody with low overhead,
Seth Rosenberg:
Especially across borders.
Balaji Srinivasan:
Especially across borders. So now you can go and click and set up a new crypto address as quickly as you can set up a new account on that site. So that's why, connecting it from before I was like, you can set up a hundred accounts programmatically. Why is that useful? Well, I can set up a customer account for every site with one click automatically on the backend and then accumulate funds there, and when I want to withdraw to my main account privately,
Seth Rosenberg:
Not only does that enable human labor marketplaces, but of course it enables agent agent payments.
Balaji Srinivasan:
Yes, that's right. So that's another area where AI and crypto really love each other, where if you're an intelligent agent, something I wrote about, somebody actually wrote a paper on this as well after an API application programming interface, it'll probably be like whatever we call it an AI API, which is how the AI interfaces with a website as opposed to – and the difference is that a normal client where it's a Python script or you're curling, that is something where every step is deterministic and it's got a very low error rate and it's meant for bulk operations. An AI API is like an inquisitive human that could make errors, but it's going to be a lot faster than a human. It's like a third interface. It's not your normal web interface. It's headless, but it's not your normal API interface because you're going to want to not allow the AI API to execute bulk deletes or bulk sends.
Seth Rosenberg:
Right, have some guardrails around it. It's not deterministic.
Balaji Srinivasan:
Yeah, exactly. That's right. Because probabilistic, you want to queue all those actions and then bring it back to the human who will click yes, I will execute this series of actions for you. So it's like a golden retriever. It's a dog that accumulates a bunch of things, a stack of actions brings them back and you click them. And now crypto comes in extremely handy here because that's how this thing can execute. Boom, it goes and searches a bunch of sites and the new robots text is like a crypto store version of all the items there that the AI API can see those prices and it can spend crypto to buy them once a human gives them the go ahead.
Crypto means instant jobs...Why isn't there more of that? The answer is it's actually really difficult to just pay somebody quickly with low overhead, especially across borders.
Seth Rosenberg:
Yeah, it's super interesting, the agent to agent payments and also information disclosure, how all that will work is super fascinating. Even, for example, if I ask your agent what your food preferences are as my agent's planning dinner for us, I'm actually using compute resources for you. And so there's even micropayments like that. That may be interesting. I mean that's probably too small to matter. I may ask your agent to do a bunch of research for me that actually may require a smaller micropayment.
Balaji Srinivasan:
Yeah, exactly. That's right.
So just to enumerate certain things. We talked about how AI makes everything fake. Crypto makes it real. Again, we talked about AI's, probabilistic, crypto's, deterministic. We talked about AI agents will want to pay for things, and crypto actually allows to have money and to make those payments. We talked about –
Seth Rosenberg:
Identities and proof of humanity.
Balaji Srinivasan:
Yes, exactly. It means that it breaks all captchas and crypto restores identity with proof of humanity. And also AI has lots of spam on the web, and so crypto allows for a signed web where like Farcaster, which is a working example, everything is signed basically with your wallet. You've got BS ETH or Vitalik doeth, and I should look at the exact details of the latest version of the forecaster protocol by notes on the roadmap to have everything having an effective digital signature there.
There's a few more points. One big one is that people have talked about how AI is killing the open web, but crypto’s going to bring you back. Why? Because blockchains are a gigantic public data structure. Something like, for example, a forecaster protocol or lens.
Seth Rosenberg:
When you say AI is killing the public web, you mean because you don't need to actually navigate to an app or website because it will just deliver the information for you?
Balaji Srinivasan:
Well, it's several things. First, a lot of the old web 2.0 companies are making it harder and harder to spider and scrape their website because people want to index it for models. For example, Twitter famously (under Elon) made it much harder to just view tweets logged out. It used to be much easier to do that, but right now you have to log into view all tweets.
The second is that sites like Stack Overflow or Reddit, people are just trying to game the site by having AI bots post answers and get up votes, which sort of corrupts the training data or what have you, right? The third is that AI is very good at generating very good spam, and so Google indexes are not set up for that. And so what you go to is I think the crypto web, which adds essentially fixes for all those things. You have an NFT gate that gets you into a site if you need it. B, you have a signed web where you can tell where their people are human. See, you have open blockchains and open backends where people can index them. And in many ways it just sort of compensates for the kinds of things that AI is doing to the old web.
Seth Rosenberg:
Yeah, it brings scarcity to abundance.
Balaji Srinivasan:
Yeah, exactly. It's got a new sort of armor plated. I mean, in many ways crypto is all about ‘don't trust, verify’, and AI is about essentially a synthetic human, which you need to verify the heck out of.
Seth Rosenberg:
What role do you think crypto will play in changing the incentive structure for training dataset? Right now it's binary, right? Elon either allows the Twitter data to be trained by his competitors or he locks it all down. There's a scenario where he just sets a price using crypto where any model that wants to train, he could somehow detect that is being used for that purpose and he actually gets compensated. Do you think there's a world where people in a decentralized way are incentivized to put content, high quality content out into the open web that can then be used as part of a training set, but they actually get compensated for it?
Balaji Srinivasan:
ScaleAI is a good company that does something like this, but people are literally paid to generate training data, which is more work than it is posting on social media. And so I think that that just becomes the new digital blue collar basically. Huge numbers of people in India and other places will just be employed, generating training data, driving around, picking things up, jumping up and down whatever is needed to train the algorithms. That's digital blue collar, and there's an enormous number of things that you need that kind of training data for, and there's a huge appetite for it. And you could have a million people generate a million points of data a day and still, well, okay, a million people generate 20 million points of data a day and you'd still want more depending on what it is you're doing. So I think that is going to be a huge way in which AI creates jobs. And again, crypto makes that easier because you can pay all these people, the million people in different countries, low amounts of money more easily.
Seth Rosenberg:
Going back to how zero knowledge proofs further extend the way that crypto enables AI, I've heard you talk about how zero knowledge can enable AI to trade on more sensitive data sets. Can you say more about that?
Balaji Srinivasan:
Zero knowledge proof is a very technical subject, but the fundamental concept is, let's say that I wanted to prove something to you without proving anything else besides that thing, I want to prove that let's say Seth is in California without giving any idea of where he is in California, that is like a zero knowledge proof, for example, where you're just proving the thing you want to prove and nothing else. It's often it used for proving set membership online, for example, prove that you are a US citizen so that you can be XYZ, but not anything else about you,
Seth Rosenberg:
Not your name…
Balaji Srinivasan:
Not your security, not your name, exactly, not your social security, not all that stuff. So you don't necessarily have to show the full passport. You could just show that, right? Basically it's similar to when you go and do Google Auth. Sometimes a site will say, here are the permissions you want to grant. So imagine it's like the minimum necessary information. You're only checking these fields and granting only that information, nothing else.
Why is it helpful? Well, if you take medical data, for example, medical data, it's medical, so it's private, but it's statistical, so you want to aggregate it. And so that fundamental paradox has been at the heart of biomedicine for many years, how do you do privacy preserving aggregations? Now, one answer is you just trust doctors and they just do this centrally. But another answer nowadays may be with your knowledge, you could have open public participatory data sets that are still privacy preserving. So you could put in an anonymized version of your genomic data and get out an anonymized version and you'd only get the statistical information and nobody would know what was in there in the mix, right? It's very, very interesting and powerful where these things people have talked about for a while, but the cryptographic primitives are becoming feasible now.
Seth Rosenberg:
And have you seen examples of that actually happening today?
Balaji Srinivasan:
Yeah, there's people like Dawn Song at Oasis working on this. I should go and check out the very latest, but I know there's a bunch of people who are working on this. There's another group called XX Park. It's a play on the words of Xerox Park, so you can look at their work as well.
Seth Rosenberg:
And what role do you think crypto will play in the governance layer of AI, like the theoretical kill switch or even decentralized regulation?
Balaji Srinivasan:
Well, I think crypto is going to play a role in the governance layer of the world in a sense. You remember the Facebook movie in 2010?
Seth Rosenberg:
Of course, the Social Network.
Balaji Srinivasan:
Okay, so what's remarkable by its absence in that movie – if you go back and rewatch it, remember this came out like 2010 – what's remarkable by its absence, if you go back and rewatch it, nothing about politics.
Seth Rosenberg:
When I joined Facebook, it was a totally different narrative. It was like a fun startup empowering individuals to give a voice, things like the Arab Spring, empowering democracy. Only until the media blamed Facebook for electing Trump did the narrative change.
Balaji Srinivasan:
But here's the thing. It's funny you say that, because the movie (I was going to make that exact point in a slightly different way), which is the movie only talked about the party and socializing aspects of Facebook. It was almost as if Zuck had founded Universal Studios Florida and people came and had a good time and he just made a lot of money doing that or whatever. It was almost like a theme park, a digital theme park is how it was presented.
But the thing is, as you mentioned in 2010, the Arab Spring had already happened, and that was actually for the people in the Middle East, a very political thing, but we observed it as giving people a voice and so on, and what was the outcome of that? There's a lot of chaos. Let's call that a messy situation. But the thing is that what people didn't project forward was if you had said at that time, even for somebody who understood 500 million people are using social media, it's grown faster than anything we've ever seen. It's multiple billions of dollars. Even people who are in the middle of it would not have believed that in 10 years, the number one most important issue in the entire world for a few days would be whether the president of the United States could tweet. Whether they could post this. Crazy to think.
And the reason is, I mean, we think about what a startup is. As a startup, you want people to use your product. You want important people to use your product. Oh my God, a sender is using their product. Wow, they logged in once. Oh my God, they posted a tweet.
Seth Rosenberg:
Yeah, you never think, ‘how do I prevent that?’
Balaji Srinivasan:
Yeah, exactly. So when you think about the state interfacing with the network and the network is so powerful that it decides to de-platform the state, that's an important flip in many ways. 2020 was the year that we became internet-first. Everything now happens digitally and then it gets printed out physically. If it's a really important example, you have a meeting online and if it's really important, you meet in person or you go through 20 things in a store, and then if it's really important, you print it out by clicking enter, and that shopping cart comes to your door. You have some cad cam thing, and if it's important, you hit enter and you print it out.
Everything is internet-first, including the governance layer, right? Politics happen on Twitter first, and in this case, why do I make this point? The point is social media is now in a sense, the Agora of the world. It's basically where all of the discussion is happening, right? It's like a giant Congress. And the reason that that's important is that it wasn't really built for that. It was built for discussion, not, but now we kind of find ourselves in a similar time, and here's why. Over the last 10 years, it became obvious that all politics became social media. I think over the next 10 years, all politics becomes crypto tribalism. So just like the mainstream media was turned into just one poster on social media, dollar holders will be turned into just one crypto tribe. It's already kind of happening where arguably there's more Bitcoin holders than there are of most other major currencies, certainly more Bitcoin holders than pound holders or yen holders for example, and probably even more than Euro holders are comparable. That's huge actually, if you think about it, that's a massive new thing, and they're probably more passionate about it than most Euro holders and they're growing. So Ethereum is not that far behind. It's like a hundred million, 150, 200 depending on how you count it.
And so the point is that you're already starting to see political people come to power and hold power in part on the basis of their crypto policies. But Kelly of El Salvador is one of the first, and now of course you're seeing this become a huge issue with the Republican Democrat and so on.
Here's my point. If in 2010, even after the Arab Spring, it was implausible-sounding to say in 10 years the president won't be the president unless they can post on social media in 2021 – even after a sovereign held Bitcoin in 2023, even after BlackRock holds Bitcoin, it still will sound implausible to you. Perhaps if I say in 10 years the U.S. government will be out of money if it doesn't hold Bitcoin. And what that means is when you ask about governance, that means Bitcoin and crypto represent a government of governments. They're above governments in the sense that they constrain a propagated state because the money can't be printed, or at least Bitcoin can't be printed. So it goes back to the gold standard where the gold standard used to constrain states because if they didn't have the gold, they couldn't fight wars. There was a constraint on what they could do. Now that's going to be as big a deal as social media was. All the convulsions around social media in the 2010s are just a prequel of the craziness that's going to ensue around actually going to limited budgets and so on.
Seth Rosenberg:
So what does that mean in terms of actual policy influence?
Balaji Srinivasan:
Oh boy. Bitcoin is a strong enough force that I think you're going to get three classes of states, what I call Bitcoin atheist, Bitcoin monotheists, and Bitcoin. So a Bitcoin atheist state is a state where possession of Bitcoin in private hands is outlawed. It's like China, for example. So I don't think they've outlawed it in private hands, but they made it hard to mess around with it. Right? Hong Kong, they're allowing it, but mainland China, it's still not. So that's something where in a bankrupt state, they will also probably want to restrict Bitcoin in private hands. It's like cheating the government or something. It'll be called something like that. I don't think that's what it is. Just like
Seth Rosenberg:
You can't buy U.S. dollars in countries with breakaway inflation.
Balaji Srinivasan:
Exactly. That's right. Or it'll be…, you'd call it guns. You'd call it speech. It's similar to a few other things. Sometimes it's banned. Right? Okay. Those are Bitcoin atheist states. Then Bitcoin monotheist states or Bitcoin maximal states will allow Bitcoin in only Bitcoin, no crypto, no anything, right?
Seth Rosenberg:
No fiat.
Balaji Srinivasan:
Well, the fiat would have to be backed by Bitcoin. So gold backed or either gold or Bitcoin only. Really hard money. These will be states that have just recovered or have just gone through an extreme currency crisis and as a reaction to that, want extreme simplicity in everything and are willing to trade off upside of having functional financial markets and so on, for lack of downside by just having everything be really simple and reduce complexity and so on. So those will be like Bitcoin, monotheist, Bitcoin, maximal estates, and then finally bitcoin poly. The estates I think will be like the UAE or Singapore or financial capitals that have enough prosperity that they can tolerate complexity in the lab, Bitcoin and other coins alongside. But I do think that that sort of trico atomization will be as much of a factoring of the world as, I mean, I dunno, guns or speech or other kinds of really important things where you have very different legal regimes and very different results as a consequence.
Seth Rosenberg:
I wanted to get your take on truth. I know you talk a lot about citizen journalism, and it’s this interesting dichotomy in some ways where we have more information available to us than ever before, and yet our shared understanding of what's true and what's not is also more divergent than ever before.
Now first of all, I actually want to get your take on if that's actually true or if that's just the perception, meaning that in a world with more centralized information, there was more cohesion around bullshit potentially. And so there was just as much or maybe even less truth, but it was just that we all believe the same lie. First of all, do you believe that there's more truth now or less truth?
Balaji Srinivasan:
I think that there's more upside and more downside. That’s to say, when the mainstream media, for example, will look at the internet, they will see QAnon, but they don't see satoshi on, they see a downward deviation, which has worse evidentiary standards than the mainstream, but they don't see an upward deviation that has better evidentiary standards than the mainstream.
Seth Rosenberg:
Yeah, it's like learning about Covid on Twitter is actually more truth available.
Balaji Srinivasan:
Exactly. And for us, for investors, VCs, people who are essentially, for lack of a better term, professional researchers, I think we're reason (I'm going to qualify this with many asteris), but I think on balance, we get better information by taking in the feed and figuring it out from there and then acting on it than we do from just having it sort of spoonfed to us, and especially spoonfed to us by people who are just really insistently stamp foot that they are the institutions.
Seth Rosenberg:
Who have a political bias and are not the ground truth.
Balaji Srinivasan:
Yeah, exactly. And especially during Covid: “Masks don't work.” Before, they do.
Seth Rosenberg:
That was so ridiculous.
Balaji Srinivasan:
Ridiculous. And there's so much standing on prestigious science, and the problem is, and this is a complicated thing to talk about, but basically there's at least two failure modes, and then there's the third success mode. So failure mode one is to say, science says you need to listen because this study just came out last week and so forth. That's one failure mode. And the second failure mode is say, oh yeah, well screw science, screw scientists, screw all of that. I'm just going to go to my crystals or make or whatever, right? And so the third road, but I think the road we need to take is the road of decentralized science, which says, science is not a prestigious citation. It's independent replication, right? Example, what happened with the superconductor thing was good because there was no political line on that the institutions are actually silent. There was no long standing doctrine that superconductors mean that you are on the right or the left or whatever. It was silent. It was not a political issue, unlike a lot of other things, right?
Seth Rosenberg:
And everyone just tried to replicate it based on the paper.
Balaji Srinivasan:
Bingo. That's independent replication rather than prestigious citation. Nobody cared about the reputation or the institution of the guy. They just care about the algorithm or the process to cook the aluminum or whatever it was to that superconductor and then test it,
Seth Rosenberg:
And then there were prediction markets going on. That was an interesting week.
Balaji Srinivasan:
Exactly, and that's awesome. That's a glimpse of it.
Another glimpse – more than a glimpse – is Satoshi and Bitcoin and crypto show the value of decentralized truth, what the blockchain is. It literally is something where whether you're Indian or Pakistani, whether you're Japanese or Chinese, even warring tribes agree on who holds what amount of Bitcoin, like the state of the Bitcoin blockchain, and that's the kind of thing that people have absolutely killed each other over in the past. You're talking about billions, but billions and billions of dollars.
So that's actually really important where Bitcoin shows that you can get decentralized truth anonymously, if you can agree on one bite in a highly adversarial environment, you can agree on a string of bites. If you can agree on a string of bites, you can agree on many other kinds of information, at least some of the basics (who put this on-chain? And what did they put on-chain and when?) from the digital signature and the hash and the timestamp.
In general, by the way, with a lot of these things, I think we just need the internet-first version. What's the internet-first version of science? Well, it's decentralized science. What's the internet-first version of journalism? Well, it's citizen journalism where everybody's a participatory journalist and rather than a professional journalist class. What's the internet-first version of democracy? So I'm writing an article in this crypto democracy where what you're doing is you're saying, well, actually internet voting already exists. You have the low stakes voting of Reddit and Twitter where you're up voting things. You have the medium to high stakes voting of something like snapshot.org where you're voting on DAOs for the allocation of hundreds of millions of dollars in funds, and you have the ultra high stakes vote of, for example, an Estonia, which actually has cryptographically verifiable hardware wallet like things, and you're making decisions on war and peace, and they may even have their former prime minister running nato. So you're talking at the very highest levels. Internet voting is already working.
So crypto democracy would basically say rather than just yelling, we're protecting democracy and so on. You defend democracy with cryptography, the same technology which allows you to send one Bitcoin allows you to send one vote.
Seth Rosenberg:
Yeah, exactly. And then you don't have this unfortunately emerging disillusion of truth even in America.
Balaji Srinivasan:
Exactly. That's right. Now every vote counts because you can count all the votes, you can get a zero knowledge receipt of all kinds of stuff. The whole system can be set up to be audited, to build trust in the system. And do I think we can easily retrofit the existing system, but do I think that we can do it on the outskirts and make it work just like Estonia on the outskirts of the west and Singapore and the UAE and places like that? Do I think we can make it work in our cloud communities, in our network states and startup societies and DAOs? Yes. And then it diffuses back in from outside.
Seth Rosenberg:
In some ways, we went from centralized truth to decentralized chaos to now decentralized truth.
Balaji Srinivasan:
Yeah, exactly. It says like centralization, decentralization, recentralization, because really decentralization is sort of, it's like hit the fire alarm and everybody scrambles out of the building, but then they need to regroup and then rebuild the building.
Seth Rosenberg:
You can't have decentralization without a shared understanding of truth.
Balaji Srinivasan:
Yeah, exactly. That's right. Decentralization. There's a lot that's good to it, but in many ways it's sort of a break glass of emergency temporary state, and then you need to rebuild civilization. I understand where crypto anarchy comes from, and I'm sympathetic to it in some ways, but I prefer crypto civilization.
Seth Rosenberg:
On that note of decentralization, what's your take on open source AI? Whether that will win, be a primary part of how these AI models evolve and what its implication is on societal risk and regulation?
Balaji Srinivasan:
So I hope it wins, and I think it might turn out to be Android/Apple, where you have sort of open source and closed source kind of neck and neck, but open source AI really should exist because it's something which at a minimum, a bunch of large companies that are big enough to fund something but not big enough to fund a full model should be able to get together on something like Linux Consortium and train an open source model that they all use, or else they'll have to pay huge API call costs to whoever has the big closed source model.
I do believe that over time we're going to see every sufficiently large community, whether nation, state or network or network state or what have you, has its own social network, cryptocurrency, land messaging gap, like a bunch of these things sort of full stack. That's kind of what defines a digital society.
You can't have decentralization without a shared understanding of truth.
Seth Rosenberg:
I know you're in general, obviously very pro-free markets, but is there some level of AI regulation that you think should exist or how do we protect against the downside of our foreign adversaries getting access to state-of-the-art AI, and even just individuals using it for affairs purposes?
Balaji Srinivasan:
I'm not against all states. I'm just very skeptical of this government, and the reason I say is, let's just take the last disaster. Okay, let's take, it's now coming out, which people kind of knew that the U.S. government and the Chinese government seem to have funded this gain of function research, and Fauci has covered up, like S has talked about this. A lot of books have been written on it. It's now kind of coming out that this seems real, right? With the discovery and the emails, and so how's the U.S. government going to protect us from existential risk? It's causing the existential risk, it's funding the gain of function research, right? Number one.
Number two is when it comes to killer AI or something like that, as I mentioned, I tweeted about this, but the killer AI is coming. It's called drones. It's here. It's here, right? So because the question I always ask when people say, oh, AI is going to kill us all. I'm like, okay, exactly how? It’s a digital life form. If it's a life form, how's it going to kill us? It's going to need actuators in the physical world. It's not going to scare you to death through the computer. It's going to hit you with something. What is it going to hit you with? Either it's going to be so persuasive – and this is a big thing for them – it’s so persuasive that it controls the mind of somebody and turns 'em into a bot, which I don't actually think is that likely, unless that person is themselves kind of like a am Enrique type cult member or whatever, if their Heaven's Gate or something, right? They're crazy themselves and the AI has changed them into that. They type something into the prompt. They're like, I can't believe it said this to me, and then they go off and run, but they're crazy in the first place.
More plausibly, the AI is scripting a drone or a robot or something like that. That's possible. I could see that the Skynet scenario playing out, but guess what? Governments aren't going to stop that from happening. They're going to do everything they can to accelerate it because everybody's all in on drone warfare. So the thing is that, yeah, there's a national security argument, there's an AI argument, but they combine, I think, in different ways in the narrative. The problem is right now we're sort of on track. I think we can probably fix this, but with nuclear physics, we ended up getting sort of the tragedy of nuclear weapons, but not enough nuclear power. Right? Too many nuclear weapons, not enough nuclear power. So it'd be a tragedy if we got all the killer AI–
Seth Rosenberg:
Which you're saying is going to be funded anyways because nation states have an incentive to create it?
Balaji Srinivasan:
Exactly. That's right. So the killer AI is, and –
Seth Rosenberg:
None of the AI doctors or AI therapists or exactly that just accelerate everyone's quality of life.
Balaji Srinivasan:
Exactly. Exactly. So people are spending, I had this kind of one-liner or whatever on this, right? Once you kind of just remap it as killer AI is drones. It's just something where the state isn't going to stop with regulation. It's going to be the number one funder of AI weapons. Maybe that's a necessary evil, but then other AI regulations, unnecessary evil, so reject all the compute bands, right? Yeah. It's like AI safety. It's bizarrely focused on the AI guided by prompts rather than the AIs that are guiding the bombs. People want to regulate AI image generators rather than AI target locators.
Seth Rosenberg:
I think it's an interesting point. And I feel like this is a classic mental error that people make, which is to underestimate the status quo because people only detect change, meaning that you're not making the argument that open source AI is safe. You're just making the argument that people underestimate the risks of centralized AI and centralized regulation.
Balaji Srinivasan:
To go against open source in the name of safety, I don't think is smart. So I will also make the claim that in general, if we're going to get, first of all, I like a lot of the people involved, but it to be said, the people who are saying, “Let's close down – “
Seth Rosenberg:
Doing it for their own benefit.
Balaji Srinivasan:
They have large interests in companies that want to close, and the thing about this is I think they're wrong. Even their logic is we have a short-term advantage with this closed model. Let's stop all future developments so we can get a few more billion out of this company or whatever. I mean, come on, that's such a negative sum attack on the commons to try to shut down all of the research so you can eke out a few more billion with one company. I mean, that's not right. Number one.
Number two is it's something where we have such a track record already with Linux and scientific publication and so on and so forth, all the analogies of, oh my God, AI is like nuclear weapons, and we wouldn't want that to be open source.
Seth Rosenberg:
Pushback I would have is that you could say that a lot of this originated from the money printing that happened to offset Covid, which was kind of a black swan event, and then you could argue that didn't necessarily have to happen. You didn't have to, but because meaning the shutdowns, but because it did, there's this black swan event which was then corrected, and now we're more in equilibrium and these banks aren't fundamentally bankrupt. It's more of a duration issue, and now that kind of trust has been restored with this backstop. It kind of stopped the bleeding.
Balaji Srinivasan:
I understand that argument. I disagree with that argument because I think among other things, they treated a coronavirus like a financial crisis. It should have just literally been, well..
Seth Rosenberg:
They created the financial crisis and then they fixed it.
Balaji Srinivasan:
Yeah, I'm not sure they fixed it. I really don't think the general populace appreciates how bad bank balance sheets are right now. I don't think most of the banks understand what's on their balance sheets, so I basically expect something significant, and I'm not the only person, obviously significantly worse in 2008, which is pretty bad, and so my view is hold crypto, hold it locally, things are going to get cray. That's my prediction for the next couple of years, two, three years, but we'll see. I may be wrong. Hopefully, I hope I'm wrong.
Seth Rosenberg:
I mean, the thing is crypto does well out of the way if the economy rips or if the economy collapses. Yeah.
Balaji Srinivasan:
In that sense, that's true.
Seth Rosenberg:
Okay. Well, this has been fascinating. Thanks for doing a late night call in the Asian time zone, and it's always super fun to catch up and debate these issues.
Balaji Srinivasan:
Awesome. Thanks Seth.