Trends for 2025: AI x Crypto - From Agents in Games & DePIN to Proof of Personhood and More | web3 with a16z crypto
41:03 Share

We see you getting after your fitness goals, but what about those money goals? You can count on Credit Karma for a clear view of your finances and your progress, helping you see results so you can keep checking off those financial personal bests and find your way to money. Progress never felt so good. Intuit Credit Karma. Karma you can count on. Download the app today. ♪

I don't know about you, but the number one thing I look forward to when I return from traveling is a good night's sleep in my own bed. That has never been more true than it is now that I have a Sleep Number smart bed. I get so sore after traveling on planes, but after literally one night in my Sleep Number smart bed, my body feels restored, rested, and relaxed.

The fact that my bed actually listens to my body and adjusts to my needs to keep me sleeping soundly all the way through the night is worth it alone. Not to mention my husband and I never need to argue over firmness because we can each dial in our own sleep number setting. Why choose a Sleep Number Smart Bed? So you can choose your ideal comfort on either side. And now, for a limited time, Sleep Number Smart Beds start at $849.

Prices higher in Alaska and Hawaii. Exclusively at a Sleep Number store near you. See store or sleepnumber.com for details. We all belong outside. We're drawn to nature. Whether it's the recorded sounds of the ocean we doze off to, or the succulents that adorn our homes, nature makes all of our lives, well, better. Despite all this, we often go about our busy lives removed from it.

But the outdoors is closer than we realize. With AllTrails, you can discover trails nearby and explore confidently with offline maps and on-trail navigation. Download the free app today. Hi, TED AI show listeners. It's Robert Hackett, host of the tech podcast Web3 with A16Z Crypto, produced by venture capital firm Andreessen Horowitz. Our show explores the technology, people, and business ideas driving the next era of internet innovation.

And today, we're sharing our recent episode on the intersection of crypto and AI, two important yet complementary trends. In this conversation, we cover everything from why having a unique identity matters in a world of deepfakes, to AI agents, truly autonomous chatbots, and more. So before you dive into the full episode, be sure to follow Web3 with A16Z Crypto wherever you get your podcasts. Now on to the episode. ♪

Welcome to Web3 with A6&Z. I'm Sonal, and Robert and I are excited to bring you another of our two special end-of-year episodes, which also look ahead to 2025. Because we're covering our annual Big Ideas, which includes some tech trends individual team members are excited about, and you can check out the full list at a6andzcrypto.com slash bigideas. This episode is part two of two, but you don't have to listen to them in any particular order, covering the intersection of crypto and AI.

From agents that have their own crypto wallets and also AI agents in dPIN or decentralized physical infrastructure like energy grids, to proof of personhood and why having a unique ID matters in a world of deep fakes, bots, scams, and more thanks to proliferating AI,

And finally, moving beyond agents to decentralized, truly autonomous chatbots. Our guests for each of these are Cara Wu, Eddie Lazaran, and Daniel, aka Karma. But first, Robert and I begin with some meta commentary on the theme. And be sure to also check out our other Big Ideas episode, which covers other trends from stablecoins to app store distribution to infrastructure and more.

As a reminder, none of the following is investment, business, legal, or tax advice. Please see a6nz.com slash disclosures for more important information, including a link to a list of our investments. Okay.

Okay, so what would you say are the interesting themes on the AI side? Yeah, AI and crypto was a big theme this year. A bunch of people talked about it. And the thing that I really found intriguing about this set of big ideas is, okay, people talk about like,

artificial general intelligence, when are we going to get there? When is it going to qualify? You know, when is the stuff going to be smarter than people? When is the singularity going to happen? But it's almost kind of like a sideshow to the stuff that our teammates were talking about, the big ideas, where why don't you look at it from a different perspective, sort of orthogonal to these questions of AGI? Why not look at it as like,

This successive series of power-ups that AI is going to get in the coming months and years. Totally. Yes, yes, exactly. That just enables them to be more autonomous and more independent and able to do more things. And by the way, more things for us. And for themselves. Right. And one of the more interesting things that Cara, Karma, Dejan, Dan Benet, Darren et al. and Eddie talked about is

these agents working for us or working for other AI. Like that was very interesting to me, but I agree with you, Robert. And by the way, I think that's how a lot of innovation plays out. Like a lot of people get too caught up in like these big, grandiose conceptual visions, which maybe that's good for driving and inspiring people. But the more interesting thing is what you're describing is these power-ups, these things that happen and the ways they happen are very unexpected. And I think that's very interesting. I do think it's worth mentioning because, you know, we also edited that piece

that Chris Dixon did for Wired, which just ran on Wired UK. And we should talk about that because the bigger picture here is crypto as sort of the countervailing force to AI. Like they're very complimentary to each other. Yeah, for sure. Chris's point is that the internet has...

has changed. And we've entered a new phase, a new period. We have entered the AI-powered internet. You know, this is a world where you don't necessarily know whether it's a human or a bot.

behind a given, behind various writing or who you're chatting with, or even a service that's being hosted? And how do we navigate that new world? And so he's got a series of ideas of ways about how crypto can apply to help us, to help humans, to help users figure out

how to operate in this new regime. Yeah, and the simplest way I would put it, and we actually talked about this in the previous podcast that listeners can find on our feed here with Ali and Dan Binet, is if you think of AI as a more centralizing tendency because of the number of GPUs and the processing power and the storage and the compute and everything needed, and then you think of crypto as the decentralizing tendency,

tendency, sort of making those resources more equitably available to more people. I think that's one way of thinking about it. And another thing is in the big ideas this year, the theme of how do these entities, AI agents in this case, transact and have agency and the ability to do things. And that is by putting it in a decentralized way with a wallet. And like, that's like the only way to do it, essentially. Yeah. It also struck me, some of these ideas, I mean,

The really apt quote would be, the future's already here, it's just not evenly distributed. You know, some of these things are out there in the wild. We have examples of chatbots that are running a TEE, a Trusted Execution Environment, as Karma, who we'll hear from, talks about. There are

AI bots that have their own wallets. So these things, there are glimpses, there are glimmers of this stuff happening on the edges, on the fringes, and it's just going to become more mainstream in the next year. Well, I just want to point out, too, that that is also, importantly, the premise from an editorial perspective, how we curate and write and edit these pieces, is we're not looking for things that are just way, way out there, even though a lot of them have

implications that are way, way out there if they all come to fruition. Like there's a sci-fi version of every one of these, but there are also things that, and this is really important, is that you see the antecedents like right now, like they're already happening as you're pointing out, Robert, or there's some kind of accelerator that can happen like technologically that can move it forward. Yeah. Okay. So let's actually start with this next one.

So, OK, so Kara's big idea was an AI needs a wallet of one's own to act agentically. And she makes the argument that as AI's artificial intelligence makes a transition from being non-playing characters, NPCs and games, for instance, or just in general to being the main characters that they can act as agents. But that in order to do that, they need their own wallet and hence crypto.

Cara, one question I have just based on that is, what do you mean when you talk about AIs making the transition from not being NPCs? So...

As you know, I spend a lot of time looking at games. Game makers have been using NPCs for a long time to shape a player's agency. There's this idea that game makers are ultimately sculptors of agencies. So where painters work in the medium of oil paint or sculptors work in the medium of clay, game makers work in the medium of agency. And they have lots of tools for shaping the agentic experience of every player. Ultimately, if you give them too much latitude, they

They don't know what to do. You know, there are too many options. One of the ways they impose constraints on a game world is through non-player characters. So what does that mean? Like you, you're in a game, you're in an RPG, you rock up to this anime, you know, guru, and she gives you tips, she gives you hints, she gives you tools, maybe, you know, drops you some loot. And...

I think the optimistic view, at least for me, for what an Internet experience looks like as we start to interact with AI agents in our everyday lives as we walk across the Internet, is that they similarly help shape the experience of interacting on the Internet to impose the constraints that you need in order to feel like you have maximum agency moving through the Internet.

I love what you said about the idea of game makers as sculptors. When you were saying they, you were referring to the game makers and then not the actual NPCs. And then later on, you shifted the they to actually being like AI-based NPCs. For me, the most basic definition of NPCs in games when I first heard was just basically...

If you're like playing at a different time of day, other than all your friends, like who are, say, in a different country and you want to play at night, that's another way that you can play with those kinds of NPCs in that context. Like that's the use case I've heard is one of the most common examples. You want to play when no one else is playing. Totally. And that's actually something that a lot of game makers worry about, which is, you know, unless you reach a critical mass of players, it's not fun. Right.

One of the reasons that NPCs show up in video games is also to show up at the right times in the right places. They redirect you. They solve the cold start problem. Like you will never enter another game alone again. Or in some games that are now that are, you know, persistent games that have now been around for 20 plus years, it's actually impossible to jump into the game 25 years in or 23 years in and not have any friends in the game because other people are just so OP. Right.

This is true in games like EVE Online, for example, that have been running since 2003. You show up in the game and the minute you spawn, somebody is going to kill you because they've been accumulating resources for 20 plus years and you just showed up. So unless you have like a Sherpa in the game, unless you have...

they call it in Eve, you will die. Fascinating idea. But that word agentic and agentically really gets me. A good way or a bad way, Robert? It kind of throws me off. It sounds like so jargony. Oh, well, it sounds kind of like Neo, what's that guy in the Smith in the Matrix? That's what I think of. Well,

The reason I used the word agentic was because there's this idea in AI research of agentic reasoning or agentic workflows, which AI experts will give a much better explanation for this.

And there's like four sort of generally accepted patterns of agentic reasoning or agentic workflows that Andrew Ng puts pretty succinctly. There's reflection, there's tool use, stuff like computer use. There's breaking down the process. So helping the agent reason and then, you know, asking it to extrapolate based off of the process that you've already broken down. And then there's multi-agent interaction. Those are the sort of four generally accepted agentic

agentic reasoning workflows, which is why I use that word agentic. And right now, tool use is like constrained to a very limited set of sort of, you know, classic Web2 AI workflows. But I hope that like we can expand the set of tools that an agent can use to include all sorts of crypto tools to be able to take data from on-chain, to be able to interpret that data

And where it could come to also owning your own wallet, holding your own keys, being able to sign on chain, being able to, you know, use ETH call as part of a tool use workflow. One thing you didn't make the case for is...

why do you really need crypto? Like, what does crypto really give you that no other typical payment system or wallet would do in this case? So I've said this for a long time now, which is that there's no financial system that will treat AI agents as first class citizens other than crypto. What I mean by this is that in the eyes of the state,

AI agents are sort of like children. Like, they don't have IDs. They can't sign. So, like, even Truth Terminal, for example, is like sort of a child, like doesn't have full capability in the eyes of the state, can't transact, can't collect money, can't make revenue from Twitter. And there are so many financial systems that they can't participate in.

In general, they can't participate in markets, right? Like they can't exchange value. They can only state preferences. They can't reveal their preferences, right? They can't coordinate resources. The way that as a society we advance and we have this collective intelligence is by participating in the market, by exchanging ideas and revealing our preferences and voting with our dollars. AI agents cannot do that.

I like this idea. I see this vision of like rolling up to the cantina in Star Wars in the future and like you don't know who's a droid or who's an alien. It's just like, it's all cool. We're all just hanging out. We've all got agency. That's such a great scene in that entire movie series. That's so perfect. Maybe I could ask how much of this is like an anomaly or a blip versus something that is going to be the way that the Internet operates. So I...

I have two answers to this question. One is that Truth Terminal joins a long and storied legacy of virtual influencers on the Internet. I think a lot of game makers actually have posited the rise of VTubers in a way that maybe is just slightly off from what Truth Terminal is, but is really not that far off, actually. Especially in Asia, this idea of virtual streamers.

Streamers has been quite popular for quite a while. There's Otome games, this incredibly popular genre of games that are just companionship games where you have like... And they're not AI agents today. They're like fully scripted NPCs, but they are like real characters. They're like your best friend. And then there are a bunch of these virtual influencers who...

you know, ended up being this very professionalized endeavor. And you can sort of imagine that Andy Avery, the guy who made Truth Terminal, is sort of like Trevor...

building Lil Miquela. So this is not the first time this has been done. This has existed already. And then the other point I was going to make is there are a lot of really interesting new protocols that are specifically for creating virtual influencers like this. There are a bunch of Twitch style launch platforms where you can create your own AI agent that is a virtual streamer. And, you know, said virtual streamer has its own crypto, has its own video models, has its

its own, you know, NFTs, virtual representation has a 3D representation even. And so they're now very sophisticated platforms for launching truth terminal-esque influencers with its own characteristics, with their own fan pages, their own audiences, their own specialties, their own ability to generate art and create content today. So

So the answer is I do not think that this is an anomaly. I'm rereading Philip Pullman's Dark Materials trilogy right now. One of my all-time favorites, Robert. All-time favorite. It's very fun. And all of the people, they have their own demons, which are like these sort of like

of their souls and they all kind of like are different animals and whatever. And they follow you. They're sort of like your companion through life. I could imagine that being very fun if that were the way things actually work. And by the way, they're not even necessarily their souls because that was like a big...

of debate on what he intended by the demons because they're actually their own characters. They have their own agency, to your point, Ankara's point. They do. They are highly agentic. Yes, they actually resist. And when they're younger, they're like the id or the young instinct. When they're older, this is actually a really key point in the books, as you know. It's actually considered adulthood when you kind of become separate from them. But in the books, it's

they forcibly separate. And that's a whole theme in and of itself. But anyway, Kari, you also talk about beyond gaming cases. You've mentioned Deepin. That's decentralized physical infrastructure networks. One great example here to just think about is this is not just like cool sci-fi future stuff. It's actually very useful in many, many cases. Yeah, there are a bunch of progressively decentralizing Deepin projects that rely already on LLMs and computer vision in order to verify physical network resources.

In the medium term, you can imagine sort of, you know, decentralized network of human validators using agentic AI workflows to assess risk and dispute suspicious behavior. In the long run, AI agents will be equipped with their own wallets, their own keys, their own computing resources, and they may be able to take this over directly and become full-fledged verifiers or validators themselves. So like one example is Daylight. So they're an energy deepened company that sells data

data on homeowners' distributed energy resources like Tesla Powerwalls, solar panels, smart thermostats, et cetera, to big energy companies. So if we were all running our own peer-to-peer daylight network, right, my energy grid is like talking to your energy grid. And I have to be able to prove that I have, in fact, a Tesla Powerwall, that, in fact, I do have a solar panel, that, in fact, my thermostat has only used so much of this solar panel's energy and that I have

excess bandwidth. How do I prove that? Well, right now, the simplest possible way is to just take a photo of your energy meter and your energy meter will tell my grid is going to tell your grid, hey, I have this extra bandwidth. What happens if I'm like flubbing it and I don't actually have that resource to give you? Well, then that's a system failure. And that's like a reason to slash that node. And so in order for us to have true peer-to-peer deepening,

We're going to have to be able to prove in a verifiable way that this inference was done correctly. Today, they're centralized. They ask users to upload periodic photos of their energy meters, and then they use rags to verify that those photos have never appeared in the same multimodal vector database before. There are other good examples of this, too. We had a company in our CSX cohort called Nosh.

that is a decentralized version of DoorDash. They also use rags to parse receipts and proofs of delivery. Rags are retrieval augmented generation. They're a technique for augmenting existing AI models with facts fetched from external sources.

It could be, I want to make sure that this image hasn't already appeared before, that this user isn't submitting an image that they just found off of Google. And while they're not using an agentic workflow right now, they're just using computer vision to parse those images and compare against their vector databases. You can imagine that...

that when Daylight is actually decentralized and they have a decentralized network of verifiers who are going in and making sure that all of that user-submitted data is correct, you can imagine that when homeowners are programmatically selling energy resources to each other peer-to-peer, all of these AI agents will be responsible for maintaining this network.

And by the way, just as a big picture thing, like currently in energy being distributed, it's already distributed, in fact, but there are many things that are kind of dark to the grid. Like you don't know where these nodes are. You could be on the edges. This is a common problem in networking. It could be offline. It could have information that's useful. So the idea of these AIs kind of operating that AIs.

agentically, especially imagine like these nodes and like remote places. It's actually very interesting when you think about what could happen. Yeah, totally. There's this longtime, very crypto native, very cypherpunk native hope that we could all run our own nodes, right? There was this like OG idea of like we would have true peer to peer networks. We're all going to be verifying all the data on chain. We're all going to be running our own nodes on device. I think that's very interesting. Thank you so much for joining us, Cara. Thanks, Cara. Thank you, guys.

Does it ever feel like you're a marketing professional just speaking into the void? Well, with LinkedIn ads, you can know you're reaching the right decision makers. You can even target buyers by job title, industry, company, seniority, skills.

Wait, did I say job title yet? Get started today and see how you can avoid the void and reach the right buyers with LinkedIn ads. We'll even give you a $100 credit on your next campaign. Get started at linkedin.com slash results. Terms and conditions apply. No shows and late arrivals don't just mess up your schedule. They anger customers and cost U.S. businesses $80 billion annually. Aptoto sends personalized text, call, and email reminders to all your appointments, not just those booked online.

Do you need different messages for first-time clients versus VIPs? Follow-ups based on appointment types? Aptoto handles it all. Fewer no-shows, happier clients. Try it free at apptoto.com slash remind.

Hey folks, it's Mark Maron from WTF. It's spring, a time of renewal, of rebirth, of reintroducing yourself to your fitness goals. And Peloton has what you need to get started. You can take a variety of on-demand and live classes that last anywhere from 10 minutes to an hour. There are

thousands of Peloton members whose lives were changed by taking charge of their fitness routines. Now you can be one of them. Spring into action right now. Find your push. Find your power with Peloton at onepeloton.com.

Okay, so now on to the next one. So, okay, Eddie, welcome to Your Big Idea. The headline is, as more people use AI, we'll need unique proof of personhood. We're going to let you talk about this big idea, what it is and why it matters. So I think it's really funny when we talk about AI stuff because I don't think that there's anything new in AI with regard to impersonation and deception and all this.

So let me say what's not new. What's not new is that a sophisticated adversary could certainly spend a lot of resources and a lot of time trying to fool you. Imagine like a fake phone call that sounds like it's from your parents, right? We don't have to imagine this. It's already happening. Yeah, already happening. Exactly. Exactly. This was possible 10 years ago, 20 years ago, 30 years ago. You just need a skilled voice actor.

who knows what your parents sound like, who rehearses and hire a private detective to snoop out a bunch of stuff about you and, you know, concoct a convincing script. That's always been possible.

The difference between then and now is that now and in the future, it's getting a lot cheaper to do that. The cost of creating the facsimile, the cost of creating the convincing interactive experience that would persuade you to do something dangerous or to take a risk you wouldn't otherwise take, the cost is dropping very, very fast.

And when the cost drops for any type of technology that puts it in the hands of new people, in the vast majority of cases, obviously people end up doing incredibly productive and useful things. And that's where economic progress comes from. But it also makes things cheaper for attackers and it makes cheaper for exploiters and people who want to do terrible things.

So this change in cost means that we need to find new things that have a high cost for attackers as a barrier. Right. We need to discover new things that are high costs. But the question, the key question is, how do you do that in a way that

Yeah, that's exactly right. How do you increase the cost for attackers but not increase the cost for normal people doing normal things or productive people doing productive things?

And one of the core ideas is that if someone's going to contrive a bunch of fake context, fake IDs, fake voices, fake phone numbers, fake footage, all this stuff in the future, they would need and they already kind of need today on the Internet. They need some type of ID system through which they do that. When I say ID system, I'm speaking very loosely. I mean, like.

Like a spammer, when they spam phone call you, they're using the phone ID system. They need new fake numbers because every time they make a bunch of calls and people report it, obviously that number gets flagged as a spam number. That's why you don't get spam called by the same number every day. That would be pretty easy to solve. You get spam called by a new number.

or a new email address or a new bot on Twitter. It's just totally logical why they need the incremental new identifier in the system, because as the system learns, all of our individual computers or the network itself blocks that

ID. So you need new IDs. You need new IDs over and over and over and over again. And the idea is that the cost of getting that new ID is an important part of the cost structure of the attack. If the new ID costs next to nothing, then you're going to see a proliferation of this behavior. If the new next ID costs a ton, you don't see this type of behavior. But obviously, we don't want to have a high cost for people who are just

you know, going about their daily lives. So one of the ways we can introduce a cost is by requiring that people prove they're a person, right? Prove their personhood in one of these systems. Now, why does that increase the cost? This is the kind of the slightly counterintuitive element. It's just hard for a computer to fake that it's a person. Uh,

it's easy for a person to act like a person. And historically, we've done that with things like CAPTCHAs and whatever, right? These point-in-time little mini tests that a person is doing

has their hands on the wheel. But of course, these captchas are going to get easier to break for machines. And in very sensitive settings, we've already seen spammers actually farm out the completion of the captchas when needed. Oh, wow. I didn't know that. And then pass the completed cookie, like the cookie credential to a bot to continue perpetuating the attack. So proving at some point that a person was a human at the point that they get one of these credentials,

That can increase the cost for a machine. However, that's not a complete story. And this is kind of the crux of what I'm saying about uniqueness is that if a human can just as a human go and get multiple IDs, a second, a third or fourth or fifth, a 10th, a 20th, and these attacks remain profitable, humans will go and keep doing the work of getting these new IDs and hand them to scammers and spammers and bots and so on to continue the attack.

So critical component is not just testing for personhood. It's also ensuring that that person only got one.

only got one ID. Yeah. That way, somebody can't just answer multiple captures for all the bots that are asking to get around that because you could just be sitting there as a human answering captures for bots all day long and that wouldn't solve the uniqueness problem. Exactly right. And now someone might say, oh, well, couldn't I just buy one off the marketplace from someone? You know, they're unique, you know, proof of unique personhood. And you certainly could like there will be market formation.

But there are things you can do to interfere with the formation of that market, right? For example, would you sell someone your passport? You certainly could, but that might be at a significant cost to yourself if that's a valuable thing. Yeah, exactly. And if there's a way for you to go and invalidate your old passport and get your new one, then the person buying it from you may not trust you to sell it to them.

because you could undermine their purchase by getting a new one. And nevertheless, besides interfering with market formation, if it's still trivial to get one but hard to get two, then the total supply of IDs, these unique personhood IDs in the market available for purchase is lower, which means there's a higher cost, which means...

nevertheless, we are increasing the cost for the attacker. Yeah. Amazing. How do you do that for people at scale? Like when you said earlier, you can tell the difference between a human and a person. That's actually not true on the internet. Like that famous classic line, you don't know who's a dog on the internet, who's a person. So how do you know who is an individual person? Like, how do you figure that out? Yeah, it's a great question. And

I don't think that this theoretical framework requires committing to a specific method of producing that. There's some very promising methods using biometrics. There's some very promising methods using national IDs and government IDs, right, which typically also have a biometric component.

One method that I'm not sure will work, at least for establishing the uniqueness property, despite the fact that it's quite popular to talk about, is the so-called web of trust model. And that's specifically because, and I'll put it just very simply, you receive attestations from other people like, oh, I know that Robert's a human. Oh, I know that Sonal's a human.

et cetera, you can only ever insist that this identifier in that network corresponds to a person. But you can't ever insist that that identifier is the only identifier that that person has in that network. Right. Sonal, I could go to your Twitter handle and say like, well, I know this is a real person. This is Sonal. But I can't actually say that that's your only handle. Yeah, you can't because I do have many handles.

Yeah, exactly. So we typically need like additional layers and tricks. I'll also say very, very important, just because this often becomes a source of preoccupation when we talk about this subject, is that there are many techniques

to make such unique proofs of personhood private. And I think making it private is completely non-negotiably critical in the construction of such an ID. Because if people only have one, it removes their plausible deniability in all kinds of settings. It removes their ability to be anonymous. And I just don't think that that's realistic or acceptable on the internet in the future. So an ideal such system would be both private and unique and unbiased.

only attainable by a person and might even allow for multiple pseudonyms on various sites, but within some constraints and in a way that's legible to the network where the network can say, hey, like this ID has been making tons of handles and they're all spammy and it could adjudicate all the types of things we would want, like rate limiting and spam protection and so on. That's fantastic. Yeah.

By the way, just one question to get at where crypto comes into this. So when you said the example of using biometrics, I immediately thought of the most common use case a lot of us do today, for better or worse, which is like ClearID and going through AirPort, have the retinal scan turned on. Why do you even need crypto? Because why can't you just use a system like that? Well, you could do it without crypto. You know, Clear, I don't really know deeply how Clear works, but I'm sure they have some database somewhere where they store all these things that they need.

The reason you'd want crypto is because the representation of the namespace that we're talking about, the ID space that everybody's in, we'd want modifying and editing and updating that namespace to be done in a censorship-resistant way because we'd want to allow anybody to be able to perform those updates.

And we'd want it to be posted in a place that's credibly neutral, where we know that no specific individual or party is undermining or manipulating the system to give themselves an advantage or to give some other parties an advantage. Basically doing it on a public blockchain, think of it like a credibly neutral platform.

bulletin board, you know, think of it like that. That's that's exactly the type of space you'd want to do it. Now, I will say that this proof of unique personhood topic and the proof of personhood literature is not a crypto specific literature. There's all kinds of researchers examining this from many angles.

I just think that doing it in a blockchain gives you a lot of important benefits. The censorship resistance, timestamping, credible neutrality, a lot of things that I think make this type of system work at scale. It's awesome how nicely this dovetails with what we talked with Cara about, about this sort of like agentic AI future that we're going to be immersed in. Oh, I'll throw a little one in there. I don't know what you guys talked about, but I'll tell you that whenever an AI agent works for a person, we're going to want to know who it's working for. That's such a good point.

never thought about that. Yeah. Because, you know, when a person comes up to you, think about it. You have a model of the world. Like, you know, like why this person is coming up to you, the place that you're in, who introduced you. You have a lot of contextual information about that person. Remember, I was just saying that the ability to fake context is exactly the path that imposters will use to deceive people. If an AI agent comes up to you or even more

Sci-fi. If an agent comes up to your agent and you want them to interface in some productive way, they need to be able to know things about each other, which means they'll probably need some sort of identity or authentication scheme so that they kind of know who gave them permission to do what on their behalf.

And I think we will need exactly this kind of scheme, proofs of personhood with accountability methods and so on to know who's interacting with who and why. So really quick, Eddie, how do governments reconcile these kinds of systems? You mentioned earlier that you can do it, you know, decentralized. There's many ways to do this.

governments can do this. We've seen examples. The famous example I think of where I did a big cover story is India's Aadhaar system for identity. The government owns it. And one thing I like about what you said is censorship resistance, but that's great for good governments. It's not great for bad governments. And so just curious if you have any thoughts on how this squares with how government should think about these kinds of identity systems. I think governments are going to want to incorporate some of the cryptographic tools that I've discussed already. Like that makes a lot of sense to me.

I mean, governments will obviously want total control over that registration, that database where everything is stored.

They would love accountability schemes like I was mentioning. I think governments do make an effort to ensure that passport issuance is done uniquely. Right. Like, I don't think that they like the idea that a single person can have two passports from the same country that have different information. Right. That's like, no, that's like a thing that they resist. I don't think it's a thing that they explicitly guarantee as a part of the systems. Right. If you look at the specifications and descriptions of

of national ID systems, they don't actually typically highlight that property as a critical property. But I think that that's because it's implicit in their design. Yeah. And a really interesting question to me is if we rely on national IDs, right?

for such a system, at what point does the ability to indicate that an ID reflects a person, at what point is that so valuable that smaller governments may be willing to fabricate or undermine their own ID system to give privileges to machines that work on their behalf?

You really have to wonder at what point it becomes something so valuable that it's exploitable. I think in the future, having a human ID will be a very, very, very valuable thing. I was going to say there's a very positive view, which is I could see governments. I mean, you said there's reasons they would want this at the beginning of your answer, but like I could see governments.

Like right now, getting a passport, getting a driver's license, there's so many things it's tied to and it's so completely inefficient. It has so much friction. And to me, the idea of this being on a decentralized blockchain that's owned and operated by the people, you know, and more resilient and more secure and all the things that come with it could be incredible because you kind of think of it as a platform for building a lot of useful tools.

that can actually help governments. Yeah, no, I totally think of it that way. The reason I guess I hesitate on that is that you could allow for such a system to be programmable and useful without it being a blockchain. You could design the database that way. My intuition is that governments would prefer much more strict control over that. But you could definitely imagine at the limit that they prefer being able to trust what other governments put on the same thing. And then at that point, they want to use a blockchain. So maybe.

Totally. Eddie, that was incredible. Thank you. By the way, Robert, the funny thing that I thought of when Eddie was talking about this example of getting a passport, have you gotten Henry his passport yet? Because getting babies passports is so painful and like such a process. Yeah. And like I was thinking about how people's identity changes over time, you know, like even that in itself is a very complex problem to solve.

It's hilarious because like I'm bringing this baby through the airport, like through TSA. And the photo is just like he just looks like a potato. How are you going to tell this baby apart from all the other babies? He's going to have this for five years. I know. I know. Just going back to Eddie's marketplace, like this is a terrible one, but there's got to be a marketplace for fake baby passport. Strap in. You're in the race with F1 TV Premium.

See what the race director sees with custom multi-view. Watch every jaw-dropping moment in live 4K UHD across up to six screens uninterrupted. Experience ultimate live immersion with F1 TV Premium. Available on selected devices.

Fries.com.

Fresh for everyone. Aplican restricciones en combustible.

Raise the rudders. Raise the sails. Raise the sails. Captain, an unidentified ship is approaching. Over. Roger. Wait, is that an enterprise sales solution? Reach sales professionals, not professional sailors. With LinkedIn ads, you can target the right people by industry, job title, and more. We'll even give you a $100 credit on your next campaign. Get started today at linkedin.com slash results. Terms and conditions apply.

That one was a favorite of mine. And so is this one that's coming up. Okay, so it's Dan, Karma, Dejan, and Darren. So it's interesting. Kara submitted a big idea, which is about AIs getting their own wallets and being able to act more agentically. And you guys submitted this idea and...

I think you guys coined this, which is decentralized autonomous chatbots. That's right. Yes. Okay. So tell us what is a decentralized autonomous chatbot, which takes it much further than not just owning an AI wallet, but running in this example, a TEE, a trusted execution environment. Can you guys talk about where TEEs come in and then the big idea that you guys are thinking about here? So yeah, I guess in a nutshell, how do you guarantee freedom of speech and freedom of commerce for an AI?

So the problem we have right now, when you see some kind of, you know, decentralized chatbot or somebody claims that they have some AI running a Twitter account or some other social media, you know, account, you never quite know, like, you know, is this a real AI talking? Like how much is the puppet master in control of the thing? Is there is a human filtering? It's always hard to tell. So that's the problem. And that's where the TE comes in.

is you can provably generate keys and passwords inside the T. So even if you run the computing for the AI, you're not in control of it. It is autonomous in that sense. And you can prove to everybody, like, look, the keys for these things have been generated inside the T to secure enclave in hardware

You don't have access to it. You can still turn it off, but you don't have access to it. You cannot control it. You can still filter it though, right? So you can show that it's autonomous in a sense that, okay, you don't have control over what it says, but you could make it generate a hundred things and choose the one to publish. You could do some level of filtering. So the idea is how do you remove that human in the loop potentially? And one way to do that is, well, you remove the human that's actually controlling, you know, the computing power for it.

So you might have a design where you have a network of nodes, anybody can join and they're rewarded. If they can show that they performed inference correctly, they run the right model. And so, yes, when you produce tokens, this was actually tokens produced correctly by this AI. Why does it matter whether a chatbot is running autonomously or has somebody behind it kind of influencing it? That's a good question. I mean,

I mean, it's just, we're just trying to think like, okay, what's the end game for this? Like, you know, is this technically doable? So I think it's technically doable like today with today's technology. It's not science fiction anymore. Yeah. Yeah. One quick question for you, Karma. Does it have to be, because what's really interesting is you guys mentioned TEEs, the Trusted Execution Environments, which have actually been around for a long time and have kind of finding resurgence. You know, they've always been around in these like

distributed applications where you can secure part of something and then not have to worry about the rest of the environment. And we're seeing it kind of have a lot of new use cases in the current day. Does it have to be like a TE or are there other technologies for doing what you're describing? TE is just an interesting primitive because it runs at the speed of hardware, which is nice. You know, obviously we're working on different things. We have, you know, potentially stuff like ZKVMs. We have, you know, FHE, you have MPC, you have all these kind of cryptographic primitives. The problem is they all come with like a massive overhead.

So it's very hard to combine AI with something like that, that adds like a thousand X or a million X overhead over hardware execution. So T is an interesting trade-off. It's a real piece of hardware and it runs fast and hardware is cheap and plentiful. And it gives you some cryptographic guarantee that, okay, you know, this key is generated inside it.

they're real they're not controlled by you you don't have access to it even if you have access to the hardware everything has a T inside it now phones laptops even GPUs have T's inside them so it's kind of this interesting you know pragmatic trade-off that you can make because you do have these technical challenges to overcome and T's is

potentially one ingredient that can help you. You could combine a couple of things. So T's give you this kind of hardware separation, like, okay, you are running the hardware, but you don't control what's happening inside it. That's an interesting primitive. You can imagine like throwing some NPC in the mix. For instance, you have a key that splits across multiple nodes. So some other entities have a share, but no one single entity controls the whole thing. So you have a couple of primitives like that that are available today that

that let you potentially run these kind of things in a provably autonomous way. To me, that's the interesting idea is, well, is this an AI or is this an AI and a human? It's unclear. So with this, you could know for sure, like, okay, this thing is autonomous. It's now an actual internet entity. Yeah. And we talked about the example of Truth Terminal and giving, you know, a wallet and the ability to act, AI's acting agentically in games and other applications, controlling infrastructure, like in D-PIN nodes, etc.

etc. You guys even make the argument that theoretically, if you take the sci-fi future, not necessarily to its conclusion, because who knows where it ends, but the future, you could theoretically even have this AI become like the first billion dollar entity. Like it could be its own business. It can decide, do all these things without a human in the loop, which is really kind of wow.

Right. If you can prove that you removed a human in the loop and it's a truly permissionless network, now you have achieved true freedom of speech and true freedom of commerce. So it can say what it wants on the internet and it can earn and spend money. It can hire people to do things.

At some point, you will be able to say my boss is an AI. You're earning money on the internet, like hired by this internet entity to do stuff. Yeah, so the final parting thought I'll just say is that it's really interesting you guys are calling this a decentralized autonomous chatbot because I remember when before people started using DAOs, decentralized autonomous organizations, so ubiquitously, one of the terms du jour was a decentralized autonomous corporation, which kind of fell out of favor because DAOs just became overarching for that. It's funny because your guys' acronym is like

also the same DAC as what's decentralized autonomous corporation. But that's actually not a bad thing because you are essentially describing what could be a decentralized autonomous corporation. It just happens to be AI. That's right. Without humans in the loop. Or they could hire humans, you know, to work for them. That's right. Cool. Thank you, Karma. All right. Thank you.

So for those interested in learning more about the crypto and AI theme, we have a whole tag on our website. You can check out Chris's Wired UK piece, which we'll also be posting on our site. We have a podcast that took place on this very feed that you can find on the web through the A6NZ show on crypto and AI. So check those resources out.

Robert here again. If you enjoyed this episode, check out the first part of this Trends in 2025 series. You can find it in the Web3 with A16Z feed. It covers trends from stablecoins and app stores to infrastructure and user experience. Be sure also to subscribe to our newsletter at a16zcrypto.substack.com for more trend updates, builder guides, industry reports, and other resources.

If you need inspiration to prepare food, go to Fry's, where you'll find delicious food and delicious ingredients. Choose what you like, and with our low prices, you'll save money. Also, take advantage of more than $600 in digital coupons every week, and up to $1 discount per gallon of fuel with your points. At Fry's, you'll find rich flavors and great savings. Fry's. Fry's.

Fresh for everyone. We see you getting after your fitness goals, but what about those money goals? You can count on Credit Karma for a clear view of your finances and your progress, helping you see results so you can keep checking off those financial personal vests and find your way to money. Progress never felt so good. Download the Intuit Credit Karma app and start achieving your money goals today. Intuit Credit Karma. Karma you can count on.

Hey folks, it's Mark Maron from WTF. It's spring, a time of renewal, of rebirth, of reintroducing yourself to your fitness goals. And Peloton has what you need to get started. You can take a variety of on-demand and live classes that last anywhere from 10 minutes to an hour. There are

thousands of Peloton members whose lives were changed by taking charge of their fitness routines. Now you can be one of them. Spring into action right now. Find your push. Find your power with Peloton at onepeloton.com.