How 0G Labs is building a modular AI chain to compete with OpenAI

0:01I'm Ashen Addison from the cryptocoin show and today on blockchain interviews with Michael Hinrich co-founder of Zerg Labs Michael welcome to the show and thanks for taking the time yeah nice to be here thanks for having me Ashton pretty excited about uh being on likewise excited to dive into uh the interception of AI and blockchain and you know AI is growing so quickly

0:22it's hard to keep up I know your team has been super busy and I want to dive into the test net road map to the main net uh everything that's going on with zero G but I'd love to start off our conversation first of all uh with just a little bit on yourself uh in business and blockchain and what got you to starting zerog G labs and then we can dive into

0:41everything yeah it's uh funny my first uh run in with technology really happened when my family so my my dad my my mom my sister myself moved to Silicon Valley and I I was kind of put I was around 13 I was put into totally new environment and I was a little bit bored at high school and so I would just kind of show up to my dad's uh office after school and just you know utilize the

1:07fast internet and kind of the technology that sap Labs had and uh one day like a manager just approached me and he was like Hey I've seen you like a few times uh you're interested in making yourself useful and I'm like uh like what like what do I do he's like how about how about you learn how to program and uh you know help me write some classes and

1:27I'm like I have no idea how to do that he's like no problem I'll send you to a Visual Basic fot camp and you know then you help us out and so four weeks later I guess I became the the youngest programmer there and because I was so young they couldn't pay me uh in cash so they gave me like free hardware and I was like totally in I was totally in heaven I had like the latest like uh

1:49thinkpads and like laptops and all the cool Hardware gears uh at that time um and so that's that's how I came into engineering and and building some of my first software programs and I stuck there throughout uh high school and college and then eventually also worked at Microsoft and did some technical product management there and then from there moved over to the business side I

2:11worked for uh daning company um Consulting Fortune 500 technology and finance companies uh from there then I moved to uh the East Coast uh in Connecticut and worked for Bridgewater Associates on the portfolio construction team then I moved back to graduate school at uh Stanford um and built my first Web Two company it's a company called Garden which I scaled to about

2:37650 employees at its Heights uh about 100 million of contracted AR and so it became a unicorn and top white combinator company and then late uh 2022 one of my classmates from Stanford uh Thomas called me up and he was like Hey I invested in this company five years ago called conflux uh two of the co-founders want to to start something new more global scale are you interested

3:03in chatting with them and so I said yeah of course Let's uh let's connect and um six months later of co-founder dating I came to the same conclusion like wow these are the best Engineers I've ever worked with we need to start something and so those were really the origins of uh zerog G and and our journey it really started with the core of the team yeah

3:25that's amazing it's it's such a cool backstory of growing up right into it and I have something similar with my father being one of the early innovators in the internet and and and here we are and I feel like there's so much Innovation still in blockchain it's the next version of the internet uh to be able to be building something that I feel like it's going to grow uh

3:43humongously still especially with the Now new capabilities of AI incorporating that into blockchain uh it's it's an amazing time building so um maybe you can talk about what you saw you know you met with those new Founders and then you have that entrepreneurial Journey you've been through this transition of uh electronics and the internet as it continues to develop and looking at

4:08where blockchain is at right now in AI what was the problem that you saw that needed to be solved with inventing zeroch Labs yeah we wanted to do something at the intersection of a what are we passion to what is within our domain expertise and then three um what does the space need in order to really be uh pushed forward and so after a lot of conversations in the

4:33ecosystem uh with projects like polygon and arbitrum and and various AI projects it became pretty clear that uh data availability was one of those key unlocks and so what we saw in the space is that the kind of uh standard was somewhere in the 1 to 10 megabytes per second of data availability capacity but if you want to truly do let's say onchain uh AI model training we really

4:59need to be thinking about tens of gigabytes per seconds especially for really large models and so utilizing our backgrounds and I actually haven't introduced Ming or fun yet um Ming uh spent about 11 years at Microsoft research he uh uh wrote some of the first AI algorithms for Microsoft's uh being search engine uh he's a distributed storage and um and

5:24scalability experts he wrote some of the key papers in the space like Tango and curfew and then fun uh was originally his intern actually at Microsoft research before they started conflex together but he equally impressive was a MIT uh computer science PhD has two Olympic gold medals in informatics and um also University of Toronto uh professor and so um with that kind of

5:49domain expertise around scalability AI distributed storage we were then able to architect uh a really high throughput system and our core insight was we need to basically segment data into a data storage and into a data publishing Lane so that we can create an aspect of horizontal scalability within that system so uh and we made a lot of uh improvements in the actual architecture

6:17and implementation level specifics and so net net each one of our nodes on our Network can generate about 10 megabytes per second and then per consensus layer if you add about 5,000 notes in a consensus layer we can get to about 50 gigabyte per second in throughput um per consensus layer and that then gives us uh an advantage of also lowering costs by about 100x compared to current

6:41systems like Celestia and Igan da and so on which then means you know the whole preamble to get to the like final goal of we can truly build things in a decentralized context just as well as in the centralized context same performance same costs uh the only limitations now are physical limitations like speed of light and so on for distributed systems

7:05definitely yeah there's always that speed of light issue um which it would be great to get there for for data availability um maybe you can explain a little bit further for those people who they're using the blockchain a little bit they don't really understand the inner workings of it um what the the inner workings of like data availability what is needed um say it's for AI or

7:28even just for a traditional search search engine is it getting the information when you ask uh chat GPT something it needs to look for that information and bring it back or find it in a certain location can you just explain the data avity availability a little bit more yeah and using AI as a example so uh if you're interacting with something like open AI it's all in a centralized

7:52server form so you make a request you know there's a bunch of stuff that happens in the background with the model around that server Farm bunch of like inference task essentially you have a lot of gpus kind of strung together in a centralized setting and that's very different from a blockchain blockchain is a completely distributed system you have uh often permissionless ways of

8:12interacting with that system so anybody can join anybody can generally become a a verifier and a kind of proof of stake system so then uh you take something like ethereum and it has kind of what's known as a monolithic blockchain it's got multiple functions so you've got the execution layer you have the consensus layer you have the data availability

8:34layer and then finally you've got the settlement layer as well and so a monolithic chain does all of these functions and most of the transaction costs go towards the data availability layer because basically it's verifying that a particular block has actually been published on the network and so by pulling out that specific layer we can then hyper optimize that part and make

8:56it uh significantly cheaper to publish data and significantly more performant to have full throughput on a particular blockchain by making that modular and and pulling that out of something like ethereum and so in effect that's what we're doing we're hyper optimizing this one one function and part of a blockchain that's very cool and so then yeah go ahead yeah yeah and so then

9:23when you have a AI specific uh use case the way it would then work is uh let's say on an execution layer you have a specific inference task that inference task then goes to to to us as a DA layer we then uh pull all of the right serving infrastructure together find the right like GPU locations load the load a particular model that's stored into memory and then kind of serve that to

9:47the end user so it's it's quite a different like backend process but it ends up going to the same kind of user experience mhm and that's good to know and if you were to compare the open AI experience that people using right now as you mentioned it's centralized using multiple gpus to something that is decentralized or utilizing blockchain in some way um is that something that's

10:12viable today and with zerog Labs technology or where exactly is that at where we can have that those options freely available where hey this is just as good as open AI but also utilizing blockchain yeah so today what's possible is you can uh utilize open source models like a llama 3 or mistol which are definitely rivaling uh open Ai and its capability um and there's a consistent

10:38kind of we want to one up each other um open AI is a fully closed Source system so you can only utilize the chat gbt model if you're within their framework and you're utilizing their fine-tuning tools and so on now the issue with that is if you imagine running an AI agent that's for example steering a traffic control system or running a big administrative process in the company

11:04like you know we're going to match candidates with the right jobs um when you you're in the open AI framework then the challenge is you don't know what the model was trained on uh where the model is sitting what version of the model you're getting what the kind of weights and biases were in that particular model and so you're have no control over that

11:25but with kind of an open source model that let's say you can train on your own uh type of training data you can know at any point in time you know what version of the model you're running what decisions are being made by the model it's completely publicly verifiable and that's really the power of combining Ai and uh blockchains together and so you've got full Provence full

11:46verifiability total transparency in fact you know you you then have an open AI you know pun intended yeah definitely and do you think that as uh AI continues to to grow there will be more of a discussion around the the the biased unbiased part of it you know when in traditional search results you can search and you can scroll down and see you sort of different results but with

12:12open AI it's just like here's the answer and in at the quotes yeah it may not be accurate but like it's sort of like this is the end all be all truth um but what are the algorithms or the what's the closed Source software that's providing you that information and and sort of who is the person who's saying that this is the truth exactly and who's deciding what bias to

12:33remove from the training data and who's deciding what the alignment of that particular model is with human interest that's all happening within one company and so how as a society can you then have insight into what that one company is doing in fact I think I recently just read that the head of alignment for open AI quit because he wasn't comfortable

12:54with the practices of open Ai and you know that's an issue we're trusting this big entity to do what's right for an entire Society shouldn't we as a society really dictate what's right for us as a society and so how do we really become part of this governance process how do we make AI a public good and that's that's what really we stand for yeah it's a big feat um and but AI open AI

13:18has grown to be a huge company and there's a lot you know if you're going to become the next Google search engine and more then there's got to be a lot into that um I i' would love to learn more on how your team has been and doing this how long has it been in the works for and what stage are you guys at exactly right now yeah we've been uh working so I I

13:39went full-time in in May of last year so it's almost exactly been a year which is uh kind of fun crazy to see uh we released uh we came out of stealth about a couple months ago we uh uh made the announcement that we uh raised the 35 million uh precede and then with that started to do a number of initi so we've also recently announced our public test

14:02net a few weeks ago and then we're going to announce public test net V2 quite soon and then move to main net hopefully sometime in July very cool and with that public test net is there ways that you know the end users or developers can interact and test it out what exactly is possible in the testnet yeah so you can uh run your own storage nodes you can run your own

14:31validator you can deploy um different kind of chains to utilize the data availability property so you can fully utilize the system in its test net uh State uh of course because it's a testet there's certain things that we're still upgrading in V2 of the test net we're going to uh improve some of the kind of storage mechanisms and also the interface uh with interacting with our

14:56test net but all of the most of the functionality is live there's some on the da side that we're still um kind of testing out and working on as well um so for example we're we'll be moving um from a data availability sampling technique to a quorum based technique um just for kind of Speed and Performance gains and so all of that's going to happen in testet

15:18V2 great looking forward to it and in terms of connecting into the the blockchain world is this something that other blockchains can be cross-compatible with or they can interact and and use the applications on as well or is it sort of a standalone with uh zero g no absolutely uh and for different depending on where you are on the stack as a blockchain we can have different

15:43use cases so if you're layer one uh we can support you with doing uh State snapshots for example or or you can fully offload your state to us and we can manage that for you so if you have a you know a blockchain that can't handle a lot of data through put then you can just offload it to and then we can deal with it if you're Layer Two then you can utilize us as both kind of storage layer

16:07as well as the data availability layer and then the same thing for layer 3es and app chains as well and so yeah we're we're very modular and you can plug Us in how however you need that's amazing so would you be able to potentially lower the ethereum transaction fees uh for all these people that are paying too high fees especially when it starts pumping uh well yes so that's that's the

16:29whole idea behind the layer twos on ethereum is basically like optimism and arbitrum uh all of those kind of low fees uh one of the reasons is you're offloading um kind of the execution to a different layer and then you're batching transactions together and then if they Implement a better da system then you're starting to talk about fractions of a fraction of a penny in terms of

16:50transaction costs so that allows completely different possibilities like you can now have fully onchain gaming fully onchain AI because you're almost have gasless uh fees at that point um and so yeah that's that's kind of the next evolution of blockchain so as as these layer twos and layer 3es and app chains adopt da Solutions like us then you can create completely different

17:16experiences really looking forward to that and you know I saw some recent updates in the AI world uh with the newer versions of open Ai and I feel like AI in as an industry is just growing exponentially and I'm guessing that it's going to be using a lot more data as hopefully you know we start generating videos and and real time uh you metaverses and all of this kind of

17:40stuff uh do you have any predictions on you know what we might see in AI in in the next year or even six months um and and how much more data availability we might need as it continues to grow it's uh it's it's always hard to predict the future but some of the things I think that are going to happen in the next let's call it five years is uh definitely more specialized agents

18:08for particular types of use cases so in the blockchain world instead of for example you know waking up in the morning and saying oh I want to buy this particular coin that's on this particular chain and I have to go through this Dex and this bridge you can just tell your agent like go buy this from me um and so that's that's one you know key application there and then uh

18:29other things in the non-blockchain world so for example there's going to be full self-driving mode from Tesla I've already kind of tested some of the data out and it's it's pretty amazing um it's still has some ways to go but we'll start seeing AI agents really interacting with the real world as well and then we'll see a world where we start having humanoid robots also

18:49interacting with our world and they're they're going to have ai models in them and then we just need to ensure as a society that there's certain features and safety mechanisms built in so that uh full alignment with human interest is is ensured and so I'm starting to see a world like that where AI is part of the fabric of society whether it's digital

19:11or physical yeah it's it's a pretty crazy future um and and I feel like it's gonna uh it's going to grow quickly um so it's good for for us to be you know involved in following along and hopefully capitalizing on the opportunities as they come to grow um is there any insights into uh the main net uh you said it's coming in the near future what has to be done uh and what else can

19:37people do to follow along to make sure that they get involved with it or learn more about it you know on day one hopefully yeah absolutely so the mainnet uh we're we're working super hard um basically the technology team or um like Ming and and font teams are basically working at 100 % to get there as quickly as possible and so we're hoping sometime

20:03July end of July for main net phase one which basically gives you the basic uh features and the stable environment around storage to data availability components and so uh you can definitely utilize our test right now to do the things that I mentioned which are you know running a storage node running a validator node deploying app chains uh layer twos to utilize our da property um

20:26and so on so definitely definitely do that uh you can follow us on uh x uh so just uh at zeror laabs and my personal one which is at M Hinrich um and you can see the latest updates there and if you're interested in building in the space we also recently launched an accelerator together with one piece labs and those applications are live till midle of uh June so definitely apply if

20:53uh you want uh to build part of the kind of biggest AI meets blockchain community in the world we're super excited about that that's very exciting and I can leave the link to the socials for Zer and yourself if anyone wants to connect in the show notes below and to the platform and the test net details and hopefully people will be paying attention I would love to also have you

21:15back on uh right around the mainnet launch if possible I know you guys are going to be super busy um but it's a good thing because you're you're helping change the world so I really appreciate your time today Michael uh all the best with everything zerog G labs and let's definitely follow up in the near future yep thank you it was a big pleasure being here