The Network Layer AI Forgot - Austin Federa

Captions as YouTube publishes them, usually machine-generated, so product and people names can be misheard.

1. 0:09

   All right, good afternoon. Uh, I'm Austin Federa. I am one of the three co-founders of the D0ero protocol. And D0ero is a purpose-built network for high performance distributed systems. And when we say network, I don't mean layer 1. I don't mean layer 2. I don't mean anything even blockchain related initially. I mean a physical infrastructure network. So physical fiber on physical cables running under oceans, over mountains, beneath cities.

2. 0:40

   And this is a deeply physical network. Uh we we are talking again about cables that sometimes sharks go after, not as much anymore, but they used to go after them more. And physical infrastructure devices. So this is an OSI layer 1, two, and three network. It's far lower down in the stack than pretty much anything in crypto operates except for like Bitcoin miners. And this network is powered under the hood by a ton of cryptographic principles. It is powered

3. 1:06

   by blockchain and it is run by multiple independent contributors. But we'll start off with kind of the first question of why do this at all? And so this is a graph of the internet. The internet we all know and love. Uh what you see from this though is there's two properties of the internet that the internet's particularly bad in. So this is a track uh over about 24 hours between LA to Singapore. And the

4. 1:31

   variance in the red and orange is what's known as jitter. This is how much the ping time varies from moment to moment across a system. So if you grow up playing video games, you know you literally live and die based on your ping time. The jitter is a measurement of how much that ping time changes. And so what you see is the public internet while it's a marvel of reach. I can reach any server anywhere in the world

5. 1:54

   for like 40 bucks a month. That's incredible. It is not built for performance. It's built for cost reduction. And so what you see here is you can actually see right around the 1600 mark uh when the cable system that this particular ping test was being routed over changed and it moved to a worse system and it went up from about you let's call it 300 250 milliseconds up to more of a base of 350 to 400 milliseconds and then you can see later in the day when it moved back down to that slower cable uh faster cable system

6. 2:27

   again this is totally fine if you're watching YouTube or if you're sending telegrams or IME messages. If you're running a high performance global network that is carrying financial transactions, AI inference and all those other sorts of activities, this really isn't suitable. And so this blue line that you see at the bottom, which doesn't even look like a scatter plot, but it technically is, uh, that is the zero network. That is consistent 24/7

7. 2:51

   365 with functionally zero jitter and a very known latency amount. This is the type of thing that we see all the time in companies like this. So you may think of these as all public internet infrastructure companies. None of these use the public internet for the majority of their work. Even eBay has a global private fiber network for moving data around because the public internet is

8. 3:14

   too slow for their needs. So when we think about this question, it's like why is Facebook involved in private fiber infrastructure? Well, it's because the public internet is too slow. And if it's too slow to show uh reels to people, so you keep scrolling, so they keep showing you ads, why are we using it in blockchain for all financial infrastructure, all transactions, and all data? And this is the journey that every successful software system has been on. Facebook started ostensibly in a dorm room, but then eventually you move on to they built their own servers.

9. 3:46

   They built their own server racks, they built their own data centers, then they built their own subc cables. initial eventually to increase their TAM they had to wire up the Philippines so they could show people ads and crypto has finally reached the point of adoption where we're actually starting just barely and just in a few areas to truly compete with traditional finance but crypto today is too slow and too unpredictable to be able to do that and

10. 4:11

    the same types of problems that are affecting the financial side of crypto also affect the AI side of it and so we have a training problem and we have an inference problem with AI data workloads as well. Let's start with the training problem because that's one of the more fun ones. What we have today is we have data stored largely in centralized databases and vertically integrated cloud providers and other types of

11. 4:34

    providers and we need to get that data to where the GPU farms actually are. So look at something like core. They built a huge GPU farm and they said okay great we're ready to take customers. The customers said where do we stick all our training data? They said okay. They went back and they built a bunch of storage arrays bolted onto that uh data. They said, "Great. How do we move like 10

12. 4:56

    pabytes of data from an S3 bucket into your system so we can actually access and train your GPUs?" And they said, "Oh, okay. We got to go build a lot of fiber infrastructure." And unlike the other two, fiber infrastructure, it's physical infrastructure in the ground.

13. 5:10

    Governments have to permit it. You have to like dig holes in the ground. It takes months if not years to install a lot of this fiber capacity. And so when OpenAI comes in and they're building a data center, one of the first things they're doing after they figure out, can we get enough power is they figure out right of ways and access to put in fiber to link all these different data centers together. And the zero network being this global network that has multiple

14. 5:35

    contributors to it is able to be sort of this fabric layer that allows data to move more quickly than the public internet. The second problem is the inference problem which we don't necessarily think of as a problem in the traditional sense but most data that you're consuming from a centralized web 2 entity it comes from close proximity to you I live in New York if I'm accessing Google docs that Google doc is in New Jersey if I'm watching Netflix that's coming from a server in New

15. 6:00

    Jersey the distance nowadays between a web 2 consumer and a web 2 service is actually not that geographically large in most of the world but all these AI data centers the one limit factor on where you build them. It's not zoning.

16. 6:13

    It's not permitting. It's where can you get enough electricity? And you can't get enough electricity in the areas where we have most data centers today. They're in rural parts of the country. They're in areas where the power grid uh allows that type of expansion. And so when we think about the data movement problem, Claude's got to pull all your code off your laptop and send it up somewhere. And it's not going to uh a data center in New Jersey. It's going to a data center in like Lach, Texas. It's

17. 6:40

    going to all of these places that historically don't have great connectivity. And so when we think about the decentralized problem of AI, it's not just the training. It's not just all these other components is that we have to move much more data as a society than we've ever had to move since the advent of the cloud. And that's a whole new problem for crypto. It's a whole new problem for AI. And it's something that the zero network is really in part built to address. So if you look at this, this

18. 7:06

    is a map of the world as we should be uh familiar with and this is the zero network today. If you go toxyz/dashboard, you can see this uh whole network in real time because it's all powered by blockchain under the hood. It's about 8.8 terabytes per second of global connectivity and capacity. But one of the really cool things about this Oh, look, it moves. Great. Um, one of the really cool things about this is it's a

19. 7:31

    contributor network built by 15 independent contributors. This is the first high performance fiber network in the world run by multiple independent contributors. So each of these contributors, they admin their own boxes, they admin their own lines the same way that each validator is run by a different independent entity and not by the near foundation. We've taken this approach that for so long has been the underpinning of blockchain compute and

20. 7:56

    we've now applied it to connectivity and bandwidth. And so we have censorship resistance. We have alternative pathways into systems if one provider misbehaves. We have staking on link rewards uh to make sure that if that link misbehaves, there's a financial repercussion for it just like slashing in blockchain. And so the entire purpose of this is to allow these networks to exceed the performance of what the public internet is capable of. And that's holding back a ton of

21. 8:23

    innovation today in the blockchain space. Our first use case for this was the Salana network where we have about 50% of validators running on it today because they built a validator client, two validator clients that were limited by the speed of the public internet. And instead of saying let's bring the cluster down to just a few nodes located in a few data centers, it was can we keep this thing global and can we exceed the pos the speed of what's possible on

22. 8:46

    the public internet. So again this is a like layer 1 2 and three in the OSI stack very lowlevel down network. This is blockchain agnostic. You can run nonblockchain systems on this as well. But we think that the crypto community where I come from like particularly understands the importance of this in a way that like traditional software companies don't yet or at least they have different needs and requirements from. So what does this look like? We have core contributor teams. We have network contributor teams. We have users

23. 9:16

    that are all accessing it. And what we build is that software protocol that meshes it all together. One of the coolest things we can do on this network that you can't do on the public internet is multiccast data. And so multiccast is this really cool system where it's hardware acceleration for packet replication. If you have one validator on in New York and you have a thousand validators in London in a normal model

24. 9:38

    you're sending a thousand copies of the data across the Atlantic. With multiccast you send one copy and that data is replicated as close as the endpoint as possible. This is not possible on the public internet. The technology simply is not supported. uh it's baked into the double zero network and this is one of the main features because what is blockchain doing sending a lot of the same data to a lot of different nodes all over the world and

25. 10:01

    so by using this we've been able to reduce I'll skip through cuz we're running a little bit on time here reduce latency and block propagation even within Europe on Salana by over 100 milliseconds we've also been able to reduce the bandwidth usage on this shred propagation system by 99.96% using multiccast. This is approximately the difference in water between these two pictures. Um, so what we're really looking at here is a whole new era for high performance distributed systems. A whole new era for distributed and

26. 10:33

    decentralized AI and the idea that decentralization doesn't have to be a clutch. It doesn't have to be something you pay extra for because you believe in the value of these systems. It can be something that can make these systems even more competitive. So if you're interested in learning more,00.xyz, lots of info there. and it'll be around later as well. Thank you.