ETH Denver AI Summit Opening Keynote

[Music] I think it's working hello hello all right um hello everyone um really excited to be here uh to kick off the uh crypto AI track at eth Denver um to start off I'm going to be kind of like the warm-up act before the real Builders and um stars come on um my goal with this talk is to kind of give you guys a general introduction to crypto a but more importantly to kind of give everyone a reason why you should be so excited about this entire track and why I think it's going to become a pretty huge deal so some of the things I'm going to cover in this talk is kind of what is this thing in the the first place um and what this thing is not and I think that second point is really important because I think a lot of people think that crypto AI is like this huge thing and and actually right now it's actually kind of tiny and it's going to grow into something bigger um I'm going to talk about what that tiny thing is so what exists today and then get into uh where we're going in the future but before we get into it I just kind of want to make the point to the few of us in this room that is able to step outside of their crypto bubble every now and then the kinds of things or the rate of innovation that's going on in AI outside of crypto is pretty insane so get this in the last two weeks um X or Twitter released a an AI model which outco competed the top AI model open air by open air 03 or I think it's 4 right now um seven days later so that was literally like 3 days ago anthropic released another AI model which beat grock so the rate of innovation of these AI models getting smarter and smarter has now come down to like a weekly basis um in addition to that not one not two but five life-sized human helper robots were launched uh that can do a range of things like buy your groceries put those groceries away make a meal for you and a bunch of other household items and I'm not saying that's Cutting Edge but it's pretty scary and very Black Mirror esque to kind of see these things happen and then the third thing is um all the top AI model providers so open AI clawed anthropic all that kind of stuff just released AI agent products and the reason this is super interesting is that they currently allow you to do a bunch of really cool things if you're a software engineer now you have the best coding assistant for whatever project you're trying to build right um if you are a business or entrepreneur or founder you can now create an app in a single prompt as if you're typing to chat gbt so these agents are going to become fully more immersed in your life but what about in crypto well in crypto for the last year an interesting stat is the Mind share of AI has been number one so so on this screen here the on the left side of the screen is a screenshot from Kito which measures sentiment analysis in crypto and you can see that AI is commanding 33% I took I took a screenshot of this this morning uh of total mind share in crypto and as you know crypto markets are very narative driven um but literally a month ago that block was 70% so AI is consuming a ton of M share in crypto and I'm wondering kind of at this point why like why is it so interesting why do people even care well the obvious thing is or at least obvious to me is I think crypto and AI have a lot of synergistic compatibility between themselves um just off the top of my head the way that um AI is helped by crypto is cheaper Financial rails the ability to create assets and trade and do whatever you want kind of autonomously manage your situation and then of course people are like well can I trust this AI in the first place blockchains help with the verification side of things right not to mention decentralization of model ownership and all that kind of stuff now on the uh how does AI help crypto the number one thing is ux abstraction so um I mean I don't think it's any um surprise to anyone here but crypto is kind of hard to navigate and use no matter what we tell ourselves the ux between a million different chains and l2s is kind of crazy now ai can help kind of abstract all of that away by just talking into a simple chat bot or or whatever that might be so there's a lot of synergistic compatibility and if you're thinking well what is crypto AI there's two buckets that you can kind of think of on the right on the left side you've got the resource stack so think of this as like uh infrastructure layup so hm how can I get data how can I get compute and how can I build these AI models that's kind of like the boring infrastructure stuff which is actually really really important blockchains can help with that and more on that later on the left side we've got the agent stack think of this as like the application layer in my opinion I think these agents are going to become kind of like the new versions of websites and so if you're interacting with these agents how does it differ from websites like do I move around do I click or do I just speak to someone or an agent and it just does a bunch of things for me it instantly understands what I'm trying to save so I think the Nuance between these two is is is super important to understand the not so obvious though is probably a lot of questions that well I've been hearing but I'm sure you guys have as well which is okay well what do these things look like how do I even interact with these things and what are they going to do for us and are they autonomous you know it's the number one question I get like hey EZ you talk about these agent things every week but like I don't think they're actually agents are they autonomous which is pretty valid feedback my kind of vision of all of this is twofold I think agents will be the number one users of blockchains and that's a pretty well it used to be a very non-consensus thing to say and the reason why I think this is think about it for the last 10 years we've built an alternative Financial system right but it's built on open source code and toolings now who do you think is going to be the better user of that software stack is it going to be a human that sleeps eight hours a day that has mood swings that probably is more error prone or is it going to be a super smart kind of human machine hybrid my bed is going to be on the latter and I already see a lot of this agentic activity happening today in its most rural form which is Bots right Bots run a ton of activity on blockchains already the second fold here is I think blockchains will create key resources so with Bitcoin with ethereum we've seen amazing Resource Management around developing new apps uh or you know uh constructing a new uh value for money or monetary asset right I think blockchains are going to be used in a similar way for AI resources like data like computes stuff like that and so I think one thing I always ask myself when everyone comes up to me and they're like hey um I want to add AI to my product and I think I could do a and c I kind of like tell them to step back and say well let's just look at like what we've been building so far like think about it we've been building all these blockchains to kind of account for humans using it right so we're like okay how can we make humans less error prone we've got smart contracts it's deterministic etc etc and we've built a bunch of different chains and now there's different tribes you know ethereum people love ethereum Bitcoin people love Bitcoin salana people love salana and some of them overlap between each other right but I'm wondering in a world full of Agents when they're the users or when AI is like spread across everything and people interact with these things in a very different way it'll probably be chain agnostic right so does it matter if we have 500 l2s maybe maybe not I don't have the answer or do we have like one to five major l1s and they do all of it for us regardless of that I think we need to focus on the low hanging fruit which is let's make one of our main sectors easy to use which is Defi and simplify a lot of the onchain experience which I think crypto a is going to help a lot with and at the bottom layer I'm representing all the the resource stack that will kind of feed into all of these things going forwards um my number one bit of feedback is EZ all of these agents aren't autonomous they're just BS and what I will say is nothing um Innovative that's gone from 0 to one has been its perfect instantiation right at the end it it's almost ridiculous to assume that we're going to have the perfect layout of all these things I think autonomy is going to be pretty Progressive think of like AI as like this baby toddler and that's the reason why people love it so much it's very human right but it gets um mistaken sometimes it makes mistakes it hallucinates sometimes and so we're kind of like guiding it we're kind of like go taking it through its like early days is it's teenage years to try and figure out what it actually looks like and how it interacts with us um and then in my opinion focus on like where the fundamentals are evolving for these machines right we're like look at things like reasoning analysis efficiency try and figure out whether these agents or these crypto AI things are getting more human if the answer is no you're probably on the wrong path there um and the data shows I mean I don't know if you guys can actually make out that tweet but it basically says that um of the top AI developers in the world pretty much much 43% are using agents to actually construct the apps that they're building and of those agents I think 34% of them use tools so they use other agents to interact with that's a pretty high hit rate okay so what actually exists right like you know stop talking about the theory the truth is there's a hell of a lot of Agents out there um they bracket across personal assistance writing content constructing um code for you or running your defi protocol or whatever that might be there's a there's a ton um and I think the one that's captured most people's imaginations within crypto specifically has been hey these like Twitter kol agents right and people are like all they do is they make funny jokes and they've kind of gotten kind of boring and I completely agree with you most of those Twitter kol Bots or agents are going to they're not going to make it but you got to ask yourself for the agents that are able to aggregate a really large community or following what are they going to do what's their next step well we're starting to see those steps by um agents like axt and for those of you who don't know it's a agent that basically gives you really good Financial alpha or tells you where to look in terms of like investing your money within crypto and it's really quite good because its tone is pretty accurate and it it like ingests a lot of Twitter data and um onchain data and what they released recently a few weeks ago was this Bloomberg terminal kind of product where if you own its token you get access directly for any kind of query that you might have for it and it's really interesting how they've like constructed um different perks based on whether you link your Twitter account or whether you have certain tokens Etc so I think a lot of these agents are going to dog food uh to their hundreds of thousands of followers a product so it's gone kind of like from a meme coin to something that has actual value and on the deep end side of things or on the AI Network the resource side of things we're seeing a lot of blockchain networks being spun up that hit some of the core components that you need within AI compute data coordination and models I've given some examples up here you'll note under compute um which I think is actually one of the most important ones right now you have a difference of training and inference and these kind of protocols basically help uh funnel into all of them I'm not going to get into them but feel free to find me after if you want to talk more about that um also I just want to point out on the coordination layer um crypto is really probably most useful there because we use tokens to incentivize different types of behavior and creation so um those two Protocols are pretty cool at least from my opinion and then on the agent side of things um I think you know going back to my earlier point I think we should focus on maybe to start off with we're seeing some really cool things appear so we have these agent execution engines which you can kind of think of as like chat gbt but for anything defi related so you can kind of like type into a search bar I think this is a screenshot from wayfinder one of the teams building one of these things and you could say hey I have these tokens on my ethereum mainnet wallet could you shift some over to Bas buy this other token and stake them for me and it'll do it all in a matter of seconds which is really really cool to see and then I'm kind of thinking of like okay these agents combining with wallets what does that look like Risk optimization ton of other stuff like that um on the gaming side I'm starting to see a lot of these agents being involved or AI rather being involved in terms of like creating autonomous entities that you can kind of interact with that evolve with your gameplay I'm starting to see adaptable environments so when you kind of move around you see these like autogenerated kind of Landscapes shift and a different social layer like people kind of interacting with these agents and I'm I kind of think of like is everyone going to interact with these agents on a chat gbt like interface or are they going to talk to each other on Twitter it's kind of up in the air now and I'm I'm curious to see how that evolves and then finally um I think a lot of these agents are going to do something which crypto has kind of been ambivalent towards but I think is going to happen more and more which is bridge the gap between web 2 to web3 I think these agents are going to help you do a bunch of crypto stuff but they're also going to help you do a bunch of non-crypto stuff and guys that's okay that's completely fine and I think it's going to be uh really conducive to Bringing more people within uh into this industry um and finally there's these things called Agent platforms and the reason why I bring this up is it's easy to talk about AI l1s like the bit tensors or the auras or whatever that might be but I think people need to focus on the consumer layer of crypto AI which is you know these different agent platforms are allowing you to build agents and then do different kinds of things so it's kind of like giving you the way to kind of build whatever you kind of want within an application layout which is through these agents right okay so where are we going I'm just going to kind of round this up I guess which which is um it's actually all going to crap and we're going to do Financial nihilism agents it's going to gamble everything away for you and we're going to have no utility from this obviously I'm doing this with a heavy dose of sarcasm I actually think we're going to go towards somewhere way more productive I've split this out into three stages in my opinion stage one is I think we need to focus on abstracting away all the complex crypto native stuff and focus on crypto agents and crypto AI stuff solving that kind of stuff so that could be compute or data on the resource side of things or helping you know abstract away a ton of the complex crypto onchain actions for you stage two will kind of be what I'm calling unsexy agents which is like agents that work for you in the back end that you don't really see that you don't really interact with but do a ton of important stuff like aggregate data or Bridge stuff for you from wallet to wallet that you don't really need to interface with but you know they're doing good work in the back end and then stage three which is consumer social agents which is like these Twitter agents that will get way more interactive and useful for you going forwards but I've spoken for way too long and I think it's time to to bring on the next ACT um I just want to give a quick brief primer of what's happening in this uh track today uh super exciting we're bringing on pretty leading uh industry panelists uh and talks but we're also going to be doing live demos of a ton of really interesting products we're actually going to start off with uh Alex atala who's going to be um demonstrating open router his product and a ton of other stuff so yeah I'm going to hand it over to him but thanks everyone thing okay okay I'm Alex atala uh the CEO and founder of open router and uh open router looks like this it's I believe the largest AI model router today um currently we have about a half a million developers and we process almost a trillion tokens a week which to put in another uh point of view is about 1.8 million tokens per second across all models a token is basically a word and before I get into the specifics and how and how I think this will play into the the AI agent sphere that e Jazz was talking about um I want to like talk a little bit about how this got started because it helps frame the reason for why you'd use a model router to begin with instead of going directly to a model it rewind the early part of 2020 3 it wasn't clear to anyone back then whether it was going to be a win or take all market for language models there were only maybe two usable ones uh Da Vinci 3 and coher command and there were a couple open source ones that weren't really usable and at the time you know there were a few experiments for how we could maybe let users choose which models they could bring to their page I started one called window AI that looks pretty similar to metamask you would type in what you want um and instead of the website choosing which model was going to be used you would be able to select your own and so here I selected an open- Source model and um I asked it to like generate directly in this web page and the website itself doesn't know what model I used but I maintain control of the model and control of the credentials so what what what turned out to happen happened open source models became really usable someone um fine-tuned llama for $600 and created alpaca and it was the first real serious usable model I ever tried out locally so uh it became kind of a a very interesting proposition whether there was going to be just hundreds thousands of these that were actively being used instead of a winner take all market and if it's not a winner take all Market you need a portal to discover these things and that's why we built open router so you can think of it as it's similar to Alchemy or infura it's one API that manages thousands of apis for you so that when it you know maybe you love this new model that came out um a new one will pop up in a week that you'll want to try out that gives you even stronger Intelligence on your tests and does better and so we see people switching models constantly on open router and that's what gives us uh data also on Which models are good at what so the rankings page here is kind of like common industry way of checking how much models are being used relative to one another and also which ones are used for different purposes so you can see you know role play and marketing use cases and uh and like science and Which models are actually most interesting for each category um you can see that today we've got Gemini flashes at the Top Model by a big shot but CLA the new claw 3.7 Sonet that jazz was talking about is is growing pretty dramatically and you can dive into individual model Pages check out their activity over time um get a sense of like how up they are which is actually really hard these are not it's not as simple as as running compute um almost all providers go down almost every week and so open router also provides a way to increase your up time now how does this fit in to to like the next 5 years it's very clear that AI is going to do way more code generation and way more coding in general than humans will and that might extend to writing um to tons of different types of content generation so what we have tried to build is a way of just making all of these things very accessible to AI agents um even fully autonomous ones you can go to credits purchase using crypto and uh you can also get an API for doing this programmatically using um you know just a simple API call and coinbase commerce and then your agent if you like can basically fund its own inference and have a virtuous cycle where it earns through some other flywheel and then directly purchases llm inference for any model that it wants and so we're seeing a couple a couple companies use this um you know an example of a of a crypto app on open route that's ranking on our leaderboard is is fractional uh fraction AI which uh is a way of monetizing large data sets and uh and you can kind of explore others over time down here as well so that's kind of an overview um happy to field questions afterwards um but I I want to draw attention to what the next speaker is going to talk about a little bit which is sort of another Frontier for open router having these models served by decentralized providers what does that mean well if you look at like Claude 3.7 Sonic you can't really do it with Claud because it's a closed Source model so only anthropic authorizes a white list of companies to be able to get the source code and serve the model and that's what you see down here but for open source models like let's take deep seek for example tons of companies can serve these and even individuals can host their own deep seek and it's just it's it's still difficult to list on open router we're making it easier but a very important point is that good inference is not necessarily going to be gatee kept anymore by just a company that has a good brand um we'll eventually move towards a future where you can have trusted execution on an untrusted server and if they're able to serve deep seek or any other model on open router with no privacy concerns then it's it's simple and easy and very high throughput to add them directly to the marketplace um we're not quite at that future yet but we have a pretty good start start with bit tensor and you can see bit tensor um on the Deep seek R1 free version that we have so it's totally free to use deep SE care um and you can see the the top providers uh shoots and targon uh which and you can see how much throughput they do this is the number of like tokens per second um their uptime over time and and you can directly use them on open router today so um a big question I think for these providers is going to be privacy right now it's it it it's not totally certain whether your prompt is going to be fully private so we ask people to opt in sharing their prompts before they can access to centralized providers I'm really excited for the future where we won't have to do that um and so if you're working on a you know te or like a trusted execution environment decentralized provider um would love to talk but for now we have a really good way of just exploring new models exploring new ways of serving them and getting users the best possible price best performance and best uptime um I think I'm close to time and so I might just invite John Durban from Shoot uh to come talk a little bit more about this yeah I think it was the end of 2023 and we were talking about fine-tuning models and you said you know open router I think there might be a product there I remember you saying that it's amazing the Juggernaut it's become but uh yeah so shoots I guess I won't go through the entire demo here but uh what is shoots we are a subnet on bit tensor so we're subnet 64 Sho or uh bit tensor has 70 subnets they all do different things ours is about being a sort of computational layer But A step above that so it's not bare compute it's the apiay layer the user layer the scaling layer it's all those things so it's basically like a giant kubernetes cluster um and anyone can build any image they want and deploy it it's all permissionless it's all headless payments um we do not just llm inference we have diffusion models we have image editing we have uh video uh generation you name it um so let's see here I'll go through some of these yeah so um if you think about agents I mean there's a ton of stuff you can do with just llms but there's also a lot of of things you can do that are totally outside of that space right so generating images like on on shoots we actually have a company that is building an RPG using shoots and they have two different luras that they run on top of one of the flux models and they use uh deep seek R1 to do the you know the actual character interaction so they've built an entire game on shoots just by deploying whatever Docker images that they wanted um on top of that there's text to speech transcription um if you want that voice interaction you need some way of actually doing so uh we have embedding models you name it so um how do we succeed well basically we just give miners some incentive the miners are the ones that actually do the work we do the routing and uh by incentivizing the right things we get the exact behavior that we need so um what we do is we incentivize total amount of compute you provide not just idle compute like your GPU is attached but that it's actually processing requests uh the number of invocations you do the number of unique apps you run so it's like your capacity your total bandwidth for providing service and then uh how quickly you deploy something that is a new app to the system so that's like a bounty and by doing that uh whenever you deploy an app you get a very very quick deploy all the miners are chasing after being the first ones to do it and then they'll scale with demand so um by basically incentivizing proper compute you automatically get all the scaling you need as a developer um our inventory I'm just going to open up the other page here you can go to this website anytime you want and see all of the gpus that are attached to our Network currently we have uh 825 h200 62 490s I me you can go through the list um and and one of the things that we do is provide llm inference we are a provider on open router and we're doing let's see yesterday we did 5.5 billion tokens uh between all the models that we host and interestingly if you look at the daily stats um for open router number four there is deeps R1 free anyone can use it to free and what's really cool about that is actually probably 99% of the usage is from two bit tents or subnets one is ours shoots and the other one is subnet four targon which is by the manifold team um so truly amazing you know that we can actually compete with some of these other proprietary you know Frontier Labs just from having an open source model that we can deploy um yeah if you have any questions feel free to come up and ask me [Applause] thank you so much okay hello everyone my name is Alonso I'm one of the founders at Crosman uh we are a developer platform that creates web CH like apis for web3 functionality so we work uh with some web3 companies but also focused on or web two ones that want to use for example embedded wallets to uh move some of their balance on chain to create loyalty programs using nfts and now we' focusing a lot on bringing AI agents on chain and some of the reasons have I mentioned earlier but our focus is on agents that want to become economic actors how do we bring them on chain with easy to use apis H okay in of Shilling I'm going to show you guys a demo of a product that we are launching right now it's in private beta but it's basically um totally onchain API that allows an AI agent to purchase any real world item entirely on chain okay so uh let me explain a little bit why first use cases uh so there's agents for example that are starting to become shopping assistants or personal assistants others that are doing gifting and giveaways uh others that are more internally operations within the company and may want to help for example fulfilling Red Bull in your fridge these AI agents need a way to pay we've talked about this before they need native interfaces and credit cards and things like that are not working well so uh for the last nine months or so we've been building a protocol and a set of apis that basically tokenize every real world inventory listed on chain as an nft and allow you to purchase those items directly as an onchain operation with any cryptocurrency across any chain evm Solana sui Aptos even cardano today I'm going to show you guys a demo um and uh show how simple this is to use this is all public code in mainnet uh you can clone it on GitHub and try it out later uh this is a repository it's a NEX application Crosman Amazon Shopper we call it it has the ability to buy any item you see on Amazon at the same price you see on Amazon with prime shipping um and how this works uh cannot dive into to all the details but uh we basically expose like this is a blockchain protocol we have rest apis but we also expose them as tools that you can Import in the primary agent Frameworks from web frame Frameworks like Eliza or virtuals to web two ones like L chain crei versel AI Etc and today's demo is going to be using versel AI for Simplicity on base paying with usdc to buy a random thing I found earlier okay I'm going to show you guys the code I think I need to make it bigger um and just get started try to do a quick walk through I don't know how technically the audience is but uh just so you see the steps um the idea is you import things most things as a library so there's just a couple things that you need to do one you need an agent with a wallet in this case just for testing purposes a private key that I've hardcoded and um I previously failed with usdc in basee uh two you need to cross me API key to get access to the API I've already gotten it before um and then sorry this is not correct you need to initialize a Veri agent this could be Eliza this could be any other agent that you like to choose sorry okay and oops it's hard to do this with one hand um and high level how this works is oh finally when you initialize the agent you need to pass it tools if if you're familiar with agent Frameworks many of these have a interface basically to extend the capability of the agent um we have a library called goat SDK that has 200 different tools available for agents and include this one crossman's headless checkout that allows you to buy real world items on chain okay um you do that and this is just setting up an agent Loop that I I cannot explain but it's like 100 lines of code um and with that I'm just going to show you guys the demo and let's pray to the demo gods that uh Wi-Fi and everything works well today okay so here we're talking with an agent interface this could be an autonomous agent whatever you'd like to be U and I'm going to say uh I have an Amazon I um I think here one second I found earlier this random uh really agous like bton of a goat and how this works you can specify many different ways what you want to buy but one of them is an Amazon URL directly so I'm just going to say buy me this with your wallet with usdc on base um and tell me first how much money you got you're going to see here some debug in instructions oh well here it already wants to buy H actually one second tell me first the wallet I want to show you guys that it's a real wallet this is debug instructions is making some tool calling and it will show you I have 175 bucks okay and now I going to buy it um and here I I need to specify some information so I'm going to ship this to the hotel so I don't get doxed um I have it here sorry okay and the recipient sorry yeah usdc Bas and if all goes well you're going to see here what the agent is doing it's getting its own address to figure out okay how am I going to pay uh and all of this is figuring out without any this there's no code rating for this this is just the agent figuring out what orders to do what what things to do he created an order it's paying for an order it's buying a token there's a transaction ID here I'll show you guys that is real okay 15 seconds ago is spending usdc see is getting an nft that proves the purchase okay and then um it should have sent me an email confirming the order all goes well okay well that will arrive at some point but I'll show you a previous order I got it sends you a a receipt and this gets shipped I'll put on Twitter when it arrives tomorrow it has one day shipping uh but it all works and reasons to do this completely reliable for an agent uh it's instant zero fees you pay directly in usdc and you have the entire inventory of Amazon Shopify print on demand websites available today and more coming every month if you want to check this out later uh I think the easiest thing is go DM Crosman or Alon GJ clone the repo or check out API dogs thank you [Applause] all right hey everyone I'm Kyle from Venice I'm going to talk to you about the Venice API and really how easy it is to use and some of the really cool things that you can do uh through Venice and through the Venice token so Venice if you don't know is a privacy focused AI company looking to give users access to excuse me um uncensored AI without ever sharing your private user data back with uh the provider whether that be open AI anthropic um whatever it may be so the cool thing about Venice is you can get access to the API if you have a simple Pro account you could also get access through through normal uscc payment rails and uh lastly you can get access by staking the Venice token and that's what I'm going to show you here so um if we go I just showed uh I already staked the token I didn't want to go through uh uh internet uh errors here so we we stake the token and by doing that you get a daily amount of what we call VCU and VCU are compute credits so what happens is um for every token that you stake on Venice you get X percentage of the overall Venice compute capacity and that resets per day so um it's a really sustainable way for your agent to have a long-term way to access uh llm inference and it's really important to think about the open- source aspect of going through some of the you know we we talked about some deep seek models some of the other amazing open source models that are out there that are uncensored that you can really trust that your agent is doing you know what you needed to do without any kind of manipulation um all right so here we got the we stake the vvv we got our VCU so we have some credits here and we're just going to create an API key so pretty much you see I have 17 credits okay uh we're going to just create an API key we'll name it test um so some of the settings here once you have an API key you actually have admin controls if you want so you can systematically create additional API keys for you know sub agents or other agents that are running different sorts of tools so you don't need to interact with the UI at all um you can really run this all programmatically which is really cool uh so we'll generate that API key um and then really the next step in in getting this running is going to either calling it systematically or going to our API reference page and uh just putting your ke you know for this uh I'm going to show you a cursor demo so we're going to grab a code model we'll run it um you'll see on the right here this is the list of all the different code focused models so we have deep seek we have quen uh we'll grab quen copy that and then uh you just pop over it's all the API is all based on open AI API format so really you can if you're using an open AI model if you're using an anthropic model you can just really uh drag and drop your API key in from Venice and uh and get it to run so uh this is a script that I probably shouldn't be showing but um so we're just going to test this uh so this is cursor pretty much it's a uh coding assistant program and IDE so I just uh if you do command L it'll bring up a chat window so I'll just say expl this code oh I'm sorry I'm going to go back and step I skipped a little so to configure that with um all right let's back out uh so it's not as easy as I just made it but uh so to configure that with Venice you're just going to go to the settings and cursor and you're going to go to the model section and scroll down and they have a nice open AI API key section uh there's a button to override the base URL to use um another source so you're just going to put in that uh Venice based URL which is api.

venice. and so on um you're going to paste in your API key here which we already did oops let's just go back and grab that we'll click verify um that runs a sample trans a sample API call just to make sure that everything's working properly and then you just add your model in and um and you're pretty much good to go so uh we have this chat that we're looking at explain this code I'm showing off cursor because I like it but um so what's going to happen is it's going to reach out and and you know we have our response there I think it's been like two or three minutes that I've been talking we just got an AI API Key by staking uh we connected it and and we're already running so uh I just want to show how simple it is and how beneficial this Venice mechanism of token staking for sustainable compute is when we talk about how we want agents to run over time like yeah we can use USD rails if we don't want to stake but having that ability where you don't have to worry about daily weekly monthly payments to keep that agent running is is really how we get to the fully autonomous layer of of you know sustainable agents so uh that's all I'm showing here today but happy to take any questions after sitting right over [Applause] there's supp but [Music] [Music] City rolling up chain [Music] she's a in Co for steel Wings smiiing latency like scissors through gel n centes chingow [Music] [Music] I'm super energized about AI um I am Matt Wright I am a CEO and co-founder of Gaia uh we do decentralized AI inference uh we've got some lovely friends here who are also building in the decentralized AI space uh we're going to go through what has been built in the decentralized AI ecosystem what needs to be built uh we'll go through some of the stack uh which a lot of us here are uh currently building have been building uh perhaps you know discuss what's been exciting out there and and you know how do we talk to developers who are uh you know building in the space and how do we guide them in the right direction uh to solve for some of the challenges and gaps we have in the ecosystem um I also have background in uh building accelerator programs and and investing and so I think all of us here have you know done very similar where we worked with builders on a on a daily basis and we're enter entering this really cool Paradigm where uh we now have to start to deal with uh you know machines and robots um as as customers and as Builders and so uh should be fun fun conversation um I think maybe we can start with uh some quick introductions um yeah sure thing so um hello so I'm uh Derek U I'm a software Builder been building software for a long time since the late 90s um um in web 3 since 2019 I'm known as the founder of a project called Moonbeam and uh uh also um also uh there we go also uh the CEO of a company called moonsong Labs uh so moonsong Labs we're a 50 person company um both building um building and building new companies uh specifically in the AI web 3 kind of intersection area so hey everyone this is Jasper I'm the CEO and co-founder of hyperbolic I did my math PhD at UC Berkeley I finish it in two years which make me the fastest person to finish a PhD in history of Berkeley uh before that I did my undergrad at P University the best university in China graduated with a highest owner and I wanton a few gadow in various mathematical competitions after graduation I work at C of Securities and Avalanche um by the end of 2022 I founded hyperb Labs with my co-founder and uh we have been building decentralized AI Cloud uh since 2022 and we raised 20 million so far backed by PR prestigious investors like poly chain variant life space faction Etc and uh we we have served more than 100K users in the world and we provide GPU Marketplace which allow anyone to contribute and access gpus and on top of the GPU Network that we built uh we have built AI INF service that serve uh tons of Open Source AI models uh we integrated with many uh mainstream uh applications uh including open router uh and uh the open a co-founder Andre kathi say that hyperbolic is his favorite place to try out base out models so if you are interested in building uh in desent ai go to our website and uh try out the apps awesome hey guys this is Ram I'm one of the core contributors at open Ledger I've been in the space from 2017 I got into it because I got paid in Bitcoin uh been building from then uh open Eder is about a year old uh we've been funded by polychain borderless and many more investors in the industry um we enable people to contribute data sets collaborate with these data sets in build specialized AI models uh think of us to be the web3 version of hugging face um we currently are in testnet right now going Manet soon we have about 1 million users who contributed data sets on top of us and more than 50 plus use cases that are being currently built and I'm happy to be here and talking about decentral as AI awesome thanks thanks guys um cool so I'll set the scene a little bit here uh for folks that are perhaps maybe a little new to decentralized AI um you know what we're doing at Gaia um just to give a bit of a a bias here um you know we're taking open source large language models right now there's 1.5 million available models on on hugging face uh you can take these uh take your own knowledge your own data natural language U and turn that into a living knowledge system uh where you can serve AI agents or AI apps uh inference and so for that to happen you know we we see the rise of Open Source llms open source AI tooling as a huge uh driver for why we're able to actually um you know help people um you know operate in a gentic uh you know uh uh scopes of work uh and kind of fight against hallucination in AI uh we think there's huge uh opportunities for verifying uh inference and and and you know validating uh you know agentic uh kind of workflows um I'd love to hear from from you guys here like um what do you see as the biggest gaps from getting us to uh this agentic future like you know I think a lot of us working uh inference and compute and um and in even in models um like what do you think you know know one at a time but what are we what are we missing sure uh you know one thing I would say that uh is there's a pretty big gap or difference between pretty good big gap or difference when you um compare you when you talk to web two companies regular web two companies and web 3 companies right so web three agent Builders you know tend towards the you know full autonomy sure let's go let's give it a you know give it a goal let's arm it with some money let's see what happens you know the the you know there's no Enterprise out there that isn't also looking at this but the conversation in my experience be quite a bit different right they're like okay well we we've automated these workflows but you know we're pretty concerned about giving autonomy to these agents you know we're not sure if the agent's going to go off the rails it's going to hallucinate is going to damage our brand is going to you know act in a way that's not conforming with our values and so that's I think there's a lot of concern along those lines that prevent the deployment of more autonomous agents in the Enterprise right now and to me that's kind of the more interesting uh segment right now because it's just so much larger than the web 3 agent audience right there isn't a business out there that isn't trying to figure this out right now you know across all Industries basically yeah I would say um web two and web 3 the main different is web 3 is all about exper experiments and try to like figure out what is the end goal of uh agents and then for web 2 it's all about fundamentals and product Market fit like whether this agent can actually deliver the good work and create value for the company for the customers and uh so because of difference uh you can see uh most agents that web two companies are building are actually utility based agent they help you to serve customers uh as a customer service agent uh they help you to write codes uh they help you to do analysis Etc but none of these uh make web three people very excited and that's the reason why in web fre space the most agents that you see are twitterbots uh you just just post twitters and people want to see how how good post they can make and then uh they try to let the agents to uh like have its own wallet and then uh do this different autonomous experiments uh but I would say currently the uh the risk the limit of having autonomous agents is uh the model itself like right now the model is not smart enough that uh it can it can't be autonomous so so this will need like at least two or three more years to have AGI level model and once we have that then it will be much easier to build autonomous agents and uh we we also need like uh if we really want autonomous agents then we need like crypto payments we need permissionless uh access to different infrastructure and that's one direction that have bic is pushing like we try to build an uh aut aous AI Cloud so that AI agents can just pay uh through crypto uh they can rent compute uh with some simple API calls uh through function calling and uh we we kind of it's too early to talk about autonomous agents at the moment but we're going to see more advancements in the next two years so I agree to most of the points that were discussed agents as you know is very Mish nowadays uh it I think it's it's very similar to what happened in crypto over the last um 8 to 9 years there's always some kind of speculation that's how IC was started and then there were projects which had utility same with D5 initially there was more of scams and then we had like actual use cases that came out um this has been the trend because in in crypto you get to go go to market right away right you go to you get to go public uh compared to traditional space where you have to like work your ass off to do that so uh I guess that's what kind of like brings down the utility in in web 3 we've also seen this happen in the traditional space if you take a look at it two years back when open AI launched angpt most of the agents that we launch where chat GPT rappers right they didn't have much use cases but then that kind of measured we saw products like perplexity use much better models models which were trained on a particular vertical like say you ask a question to publicity it gets the context of that using open aai and then if it's a math question it'll run through a fee model which is a specialized math model for you to give an answer right which is more precise than using like a generic model so the key behind this is to make sure these agents have the right data Fates and models so that they they can be much more precise and less hallucinating very similarly how humans have been over the last few years like you know I mean last all generations because we specialize in a particular field and we become an expert in that and that's how we are able to like provide specialized Solutions we not we don't have like a general knowledge right so um I guess that's what agents have to be they do not they don't need to use generalized model model generalized models are needed only just to understand context specialized models are the one which actually brings Precision um and that's what we want to Sol right we want to provide the infrastructure for people to collaborate with data and build models on top of us so that we have different use cases Sol like use cases around defi use cases around governance probably traditional use cases around Healthcare and so and so as well so um like the idea behind this is to enable people to contribute data sets so that we enrich AI with data like how we enrich web with content and make sure this agents on any kind of a Services have enough data sets enough models so that they can be smarter and uh that's that's what I believe is the infrastructure that is very much code needed and then obviously there's other providers like gpus and all that that is also needed as well amazing um thank you so I'm going to say something crazy here guys um in the next one to two years we're going to have more AI agents on the internet than humans on Earth 78 billion entities on the internet doing autonomous functions over time we'll get there um but what's what's insane is that we have you know the Advent of llms like deep seek that have now taken the the model and llm game to a zero sum game like it was once kind of this thought that Silicon Valley needed you know 10 billions of dollars to invest in in large language models to be sufficient we're now doing this at in the5 to $6 million range it's going to go all the way down even lower um sure performance and benchmarking uh is another conversation to have but um it's it's quite obvious that open source large language models at scale and agents um operating in autonomous capacities um is going to create this interesting world where you know we're interacting with agents as apps as the front end and it's agents all the way down um our agents will have agents um and so we're starting to build in this new paradigm where um you know we are talking to you know robots for you know machine to machine Commerce um and and software development so I think um one question I have for you guys is you know how do we you know uh give some advice to Builders out here who are looking to build for this new customer category um how do we approach it um you know they don't need some sexy ux they just need a API U like how do we how do we how do we approach that how do we give um Builders some advice I mean the first I just wanted to commented on one of your points that you know I too this kind of um deep seek some of these things we've seen recently you know that makes me very bullish in a way because you know the worst possible future in my mind is this world where there's you know just one two three companies that you know Mega companies Mega models they're controlling everything frankly there's not much room for web 3 in that picture either or or limited like picture right so to see you know small teams able to like innovate at the same like level as these large companies you know that to me speaks to a future where you know there's a lot of specialized players and that's like you know a place where web 3 can really like you know play a strong role right connecting facilitating like and coordinating between these different players and what do you think like in web 3 would be like those Primitives or intense that you know would be helpful yeah I mean you know this obviously goes back to the question we've grappled with a lot which is you know like what is web tree good for in this context and you know in the worst case it's it's not doing that much you're just slapping kind of a token on something but you know where I've come to is there's a few things that you know web 3 and blockchains are really good at one of them definitely is this to be act as a coordination mechanism so I you see that as the core idea behind the whole deep in category where basically you know the blockchain is acting as this coordination mechanism for a multi-sided Marketplace so I see that pattern as something that is highly likely to emerge with agents like you know being the the participants like in these networks and uh but you know the second point that I really lik that you made was around you know what to think about what does it mean when agents or your customer and agents are your suppliers I think that almost all the tools that we have today have been designed with humans right with an assumption that it's a human on either end of that so that to me is a huge opportunity to just you know all these things that exist today need to be rethought in the context of you know being agent first or age Ecentric so that's for new Builders now I think there's a huge amount of opportunity to just you know you ask yourself a question what does sales look like when you're selling when you're selling to an agent it probably looks a whole lot different than you know the sales motion that most companies go to a bar and doing BD yeah exactly so so that to me is like exciting and you know that's where I see a lot of opportunity for new for new projects thank you yes so um I think like we'll totally have a different ux uh when when we we have much more agents customers than human customers like way you said we no longer need UI what they just need is API and so we need to make sure that all the transactions that human all the inter interaction that human are current doing on the web page uh can be done through API or another possibility is like we know uh there are many different browser use agents and uh even the uh chbt opening a release like operator agent which can browse through internet uh through via a browser so that's another potential that you can see agents uh continue using browser as the interface to interact with the with the different applications but I would say uh preferably agents would prefer API so I think at least we need to Pro provide like apis as the way for agents to interact um that's what we are current doing like uh on hyperbolic we have this GP Marketplace so you need to uh first register your account you need to connect the wallet with the uh with the account and then also uh rent gpus through API and then access the gpus and so we almost finished all these different apis to allow agents to to access the compute and uh we start sing more and more different companies that just release like a markdown for the whole document documentation they have for their applications um I think that's the first step and we're going to see more and more uh design that's good for uh that's like very useful for the AI agents so the questions around how decentralization can enable more Renovations like what's what's the so if we're looking at a new customer category like agents um there's to um work with this new new kind of uh Persona yeah so um I believe the major point of decentralization is right I don't think we need to like really look at um targeting like centralized um projects which are like building like larger language model that's I think initially that's like Out Of Reach because there's so much resources and cost that goes into that uh it's like what happened with the internet right how like when internet came about like it is centralized but what we built on top of that like https and all that is what enabled the worldwide web to happen right so how can we use uh resources that is available in decentralization to enable much more bootstrapping to happen I think it's the best form of how we can contribute the you know the upcoming customer interface agents and stuff like that so if you can enable resources to be available and if you can enable incentiv session for people to contribute for these resources like for example data we incentivize people to contribute data sets so that we can build models I think that's one of the core aspects of DEC ination because we can use crypto rails to incentivize people across the globe and I don't have to worry about any kind of tips right so um this definitely happened with gpus if you take a look at uh the initial decentralized compute projects uh they enable gpus to get at much lower cost compared to the centralized providers you could get it at a fraction of cost that enable lot of builders in the space to actually rent these CPUs and build models and like experiment around that that's what we want to do as well we want to enable people to get access to data at much lower cost get access to models at lower cost so agents people are exploring with agents who want to experiment with this can have a plethora of models and data feeds that are available where you don't have to worry about paying like you know High inference cost or tokens to these larger language models you have much more incentivizes smaller models that are available which are precise compared to the larger language models there so I think bootstrapping and incentivization is a core of web three that we can use to innovate much more larger you know experiments that can be can that can be done in the space amazing we're um almost up for time but we'll ask one last question hopefully if that's okay um stop me if that's not okay um but you know looking at kind of the this machine to machine Commerce and uh you know these new uh customer categories uh or you know how do we enter like existing verticals like you know consumer gaming uh you know uh um uh defi um you know let's go maybe like real quick across like what verticals are interesting to you folks um for me personally uh very interested in agent F of course um not just agents man managing a portfolio but managing like a collateralized debt position and perhaps like you know offering a defi loan to me and and making sure I'm not liquidated um I'm also interested in uh future of work I'm working very closely with uh the mother Del folks with my co-founder James Young um on agents with jobs um so agents that can be virtual twins of community managers hackathon organizers U Bounty distribution agents Grant distribution agents um what what is interesting to you folks what verticals are um like a big opportunity yeah take maybe a little bit less of a a vertical but like I've mentioned the idea before about coordin blockchains as coordination systems that's one idea we're on on the other one that we've been exploring lately is using a blockchain as uh you know one of my colleagues has this uh expression he uses as a the network is the gym and what he means by that is using a network as kind of an open place where agents can contribute code other kinds of artifacts but then you're using a reinforcement learning mechanism to kind of evolve this like shared capability you know almost like an open source Cod project I think that's an interesting you know an interesting idea that leans into how you know AI works and like can produce like Superior results like we saw with deep seek using this reinforcement learning approach so yeah I'm interesting coding agent uh we're starting seeing like traditional web2 apps uh like cursor codium Etc you can now use agents to code and uh what I want to see is uh you actually have autonomous agents that can uh submit PR uh make edits of the code and even write codes from scratch uh that's something that I'm very excited about and that will push the limit of company like we're going to see humanless company or like just one single human manage 100 agents that build a 1 billion unicorn company uh in the very soon yeah I think uh def5 is definitely the space that I'm quite excited about we all know that D5 is quite complex we have like common consensus on that so uh definitely onboarding in D5 making it much more abstract agents can definitely help you out in that particular sector uh gaming is quite interesting as well um NPCs come can come to life with gaming especially like in web3 gaming there's always a problem of liquidity you don't have much players so if you can create like agents of your own versions on how you play and have those agents play across like games I think that can solve a lot of liquidity problem player liquidity problem in the def gaming space um I also think that governance is something which can be explored governance in web3 is pretty slow not much happens because you don't have people coming and actually voting for various proposals is there so you can have agents who understand who you are and go vote for governance uh that's an interesting uh you know sector that I would like to see world as well so these are some sectors amazing uh thank you thank you all thank you everyone for listening uh big round of applause for my panelist here