I Dismissed Web3 + AI. I Was Wrong | Paul Gebheim - Glass AI

Hello everybody. Okay, everybody that's been sitting and standing and listening, come over here. Everybody out these chairs, stand up a little bit. Let's get a little movement. It's 300 p.

m. We all need to remember that we have bodies. We're not just agents yet. I hope. There you go.

Get a little stretch in. We're going to do this. Okay. So, my talk today is called Web 3 Plus AI isn't dumb. Um, one of my co-workers suggested that I give it the tagline of I am, but I don't think I'm dumb.

My name is Paul Gabe. My title is the knower of things. At some companies, I'm a builder of things. Um, I've been building for the web since about 1996, which is crazy because I'm 42 years old. So, for those of you that were building in 1996, you know that we didn't really all know that this was going to be possible.

And I think that we're in a similar shift now what's happening with AI. We're in a similar shift that's happening right now with web 3. And I think it's kind of interesting. Um I'm going to start off by just saying like I am not an expert. Um by the way, Shitz Creek is awesome and we have a recent passway here.

But um I'm not an expert, but I I was not an expert when I started coding for web in 1996 either. Um in 1996 I wrote Pearl CGI. I don't know if anybody in the room has done that. Um, I also then went on and worked on video streaming before anybody was an expert on video streaming. I worked at Justin TV and now I'm streaming on Twitch again.

Um, after that I worked on mobile stacks before anybody was an expert in mobile. I built a beer social network at one point. Um, and then I moved on to blockchains in the mid mid2010s. Nobody was an expert there either. And I don't think anybody is really an expert on what it means to build AI.

Now I know we have the people that know how to build models etc. But like what does that actually mean for application development? What does that mean for a platform? We don't have that yet. Um I'm just going to start a little bit with my skepticism.

So when I started working on this problem, I was looking around at what we were calling the AI web 3 space. Does anybody remember what this looked like at the beginning? Do we have any answers? Any any memory? What were people trying to do?

It was basically just decentralized compute marketplaces. This was the first thing that everybody like picked up and they said, "Okay, web 3 plus AI must mean that we like tokenize a GPU or something, you know, and out of this we do get some cool stuff, right? We do get Prime Intellect and they're doing a lot of cool work. we get a lot of interesting stuff. But my skepticism is that I think it's extremely difficult for these sorts of platforms to compete with the centralizing forces which are the scale and power density that's required in order to build models.

And so I looked at this problem and I just think I don't think this is really where we're at. And then the other piece is that at web 3 in the way that we talk about it now, we're really good at vibes, right? But I don't think anybody's going to beat Nvidia with vibes. we have to beat Nvidia a different way. Um, so yeah, first thing here is that the economics are really tough.

If we're just talking decentralized AI, which is why I'm using web 3 plus AI, not decentralized. Um, the scale is very difficult, capital expenditure to build models, to build large scale infrastructure, very difficult, latency, very difficult, integration, terrible. So I think what we need to answer if we're talking about why is web 3 and AI important and this is where I really came to converting myself is that the real questions of what is important here are questions of power, control, and trust. the the world that we're walking into. Everybody just listened to Tamas talk about how we're going to have agents drafting EIPs that will make proposals that will change a core protocol that we're all using to trust our money and then those will be reviewed and then blah blah blah.

That whole piece, right? That's something that we're talking about that seems kind of far out and kind of crazy for Ethereum. But this falls into in the corporate world, we just call this change management. Who here's worked at a corporate job, ever needed to deal with change management, right? The real problems that everybody's going to be trying to solve in enterprise as we go is this change management.

And they're questions of power, control, and trust. And power, control, and trust are actually things that we know how to do very well in web 3, right? We have primitives, we have tools and we have techniques that we have developed over 15 years at this point which allow us to bring into production and into the world interesting ways of composing things for instance governance. So one of the major things that we're going to see in the across the AI world is the problems of governance. So I'll give you an example.

Uh today you guys we're all working with models that take a very long time to train for the most part right? So you get GPT 4.5 and or whatever opus 4.5 and then 4.6 and those even at those boundaries we have this problem of like does my system still work.

So if I'm a company that's deploying real product into production and it needs to use these models in order to get things done. There's a huge change management problem of how do we do upgrades? How do we actually make this work? Right? this is going to get worse and worse and worse as the problem uh gets faster.

And so in the world that's coming with reinforcement learning where we've got models that are being AB tested against a bunch of agents and all this kind of stuff that will be happening, we actually need ways to start thinking about controlling upgrades, how to do forks, how to do consensus around things that are happening. And I believe that a lot of the tools that we've talked about in web 3 become useful in this case. I'm not going to go into details. This is more about like if you are a builder in this and you've thought of hard problems, try to think about how it will work in the case where you have truly autonomous systems. Um, another one is privacy.

This one is like [ __ ] terrifying to me, frankly. I'm not not sure if I'm supposed to swear. Um, but like the reality here is that oh yeah, I declare privacy. This is basically what's happening right now. um across the whole AI space, right, is that there's a bunch of companies that are running literally spending hundreds of billions of dollars to build an infrastructure and they are hoping that they can outrun the hell that is there.

Um for users, this matters a lot. This is privacy of my inputs. What data have I put into it? For AI companies or model companies, it's also what are the weights of my models? How do I protect my IP as somebody that spent hundred billion dollars training this thing?

or if I'm an agent company, how do I actually provide value to a user that's not just completely clonable? Um, so I think privacy and all the tools that we've been building, I know we've had big token bumps last year from Zcash and all the things. I worked on Snark, I worked on the ZK stuff a lot at Polygon. Um, the ability to do Z to to do private computation, verifiable execution will become critical in the world where we're really scaling these models out. Um and then the last big case for me is coordination.

And so sci-fi story, let's start with the sci-fi story, right? So the story that everybody's telling all of us is that uh exactly what Tamas said. We're going to create this world of autonomous agents and these autonomous agents are going to do lots and lots of little things. So let's take a particular version of that case. Let's imagine that like we really get what everybody said and like next year we have autonomous robots, right?

So we've got a 100,000 autonomous robots that are walking around San Francisco just doing [ __ ] for people, right? I say like, "Let's go get me some groceries." And it goes and it picks up the best deals at Whole Foods for me or whatever, right? Kind of does that thing. We've got all these mobs on the street.

How do they decide how to charge themselves? I really mean charge themselves. Like power in their butts. How do they decide to do that? There's a coordination problem, a realworld resource coordination problem that will occur there.

And the things that we've built in the blockchain world are actually fantastic solutions to it. Imagine if they could run uh autonomously created real-time auctions for resources that they need at that moment and coordinate dynamically in that case. Well, we still need censorship resistance. We still need civil protection. We still need stable coins or some way to transfer value.

We still need double spend protection. This is what we've been building in this space. So I think that when we look at agent coordination problems in the real world, in the digital world, we have done a very uh a lot of the leg work that's going to be necessary in that system. The last one here is also an adversarial context. Um for people that aren't familiar with building in like the enterprise world, usually your adversarial context in that world is really sliced down, right?

Your CISO provides a bunch of rules at which you operate and then you kind of operate within some trust boundary where you assume that people aren't going to try to screw you too much right within this trust boundary. The problem in the agentic case here is that that trust boundary is going to shrink dynam shrink a lot. Right? So if we're building autonomous systems that are powered by artificial intelligence, there is going to be a huge adversarial problem that is on the hook for any piece of software that exists. We've also been building systems that are resistant to adversarial problems.

Some of this is via algorithmic game theory, via tokconomics. Some of it's via security principles. some of it via snarks and all these kind of fancy technologies. Cryptography, fully homorphic encryption. These are all tools that will need to be used in order to bring this world into the future.

Um, and we have the tooling. We have tooling for verifiable execution, right? Uh, over the last couple of years, succinct um, these teams have brought forth the technology on ZK in order to make verifiable execution possible. We have amazing teams like Zama that are working on fully homorphic encryption primitives. We're doing some work on that too.

Um there's a lot of mechanism design that's existed in the system that I think is going to be workable. And of course, everybody in their mother has talked about X42 and payments and how that works in this system. Um but we have this tooling and we can apply it. I think that if we all get together and are creative, we're going to find a lot more places where this will work. Um verifiability obviously in payments.

Um and then I'm just going to give you example. I think everybody's probably seen this, but if you haven't, you should. Uh there's a few versions of these, but uh in a world where we have the ability to use generative AI to create any image that we want, the ability to capture truth via camera becomes important. And an existing tool that we have developed that was funded fully by the the blockchain community's commitment to ZK is a tool here that allows you to do verifiable image editing. So, for instance, if you use a camera that has a private key built into it, you could take a a photo, you can then do image editing on that and have that all verifiable and you know that it's actually back to the source image.

So, if you have to crop it for a photo that's going on a newspaper or something, you can actually get provenence back to the exact camera that created the image. This is going to be incredibly important as we the generative uh world kind of like really takes over. Good example. Um, so in conclusion, AI is going to create a bunch of autonomous actors. It's going to be scary as [ __ ] Web 3 enables autonomous systems to work together.

So the future is not Web 3 off in its own thing. Web 3 will be everything. Thank you.