NEXUS Final Round - Blink AI - TOKEN2049 Singapore 2024

all right good morning everyone thanks for having me on the stage again so uh what I want to do today is give a quick introduction on what we did talked about yesterday and then uh give a preview of what we were building I think it's going to answer a lot of questions that came up yesterday as well as probably preempt some of the questions that are going to come up today uh so at a high level what blink Ki is trying to do is make anyone make it possible for anyone to build vet three experiences uh within our our Benchmark is less than a minute typically these are fungible non- fungible uh token use cases uh we want to make it super easy for anyone regardless of environment to be able to deploy that yesterday we saw a video of where we could essentially deploy a token uh in less than a minute everything will be intent based and the idea is that generative AI is putting the rest of the solution together so when it comes to setting up your own bonding curve to set up your liquidity and then uh if you wanted to go to a DEX you could either go straight to a de set up your own Bond income and the parameters for that as well so that's a token use case that's going to be ready in the next week or so for six uh evm environments and then Solana virtual machine and after that we're going to be releasing the nft uh the capability generate nfts within less than a minute as well and then post that it's going to be the prediction markets so this is a quick demo of what we showed yesterday is we have what's called as a turbo mode so this is for someone and the vast majority of the people that wanted to use it uh they wanted to go to the turmo mode and build something in less than 60 seconds here we customize a journey for different personas this is for someone who wants to quickly deploy something and then uh if you did not want to actually click on that you can voice prompt as well we support like 29 languages through 11 Labs uh this is what the product look like this is the demo we saw yesterday this is just depending on what type of wibe you want for your token uh it can create a landing page hook it up with a bonding curve with a smart contract the difference here is being all of this is being built by generative Ai and you actually get the smart contract this is your own it's not something that we hold the funds for uh it's deployed on any of the chains that you pick and choose you can change your mind and go back and build this as well uh you also get a um a chat window there to bring your community over discuss with them say your GMS and everything else likes that tokens there and short list them as well uh this is uh the bonding curve that we showed yesterday as well so this is where you actually design your own bonding curve so the AI will actually guide you through it so you can set up what market cap you want the token to come in at and how at what point when you want to go to a DEX or keep it there and what type of bonding curve you want the aggressiveness of the uh slope of the bonding curve as well that was for the turbo mode this is uh the more advanced mode that we didn't do yesterday I have a short demo of this which will explain how this would work and this would I know there were some other the questions around how generative AI is actually building this so this is essentially what it looks like when you go into this mode you'll see generative AI actually build a code and then you'll see the attribution score as well so this is something that a second model is check checking and it's actually trained on audited smart contracts and it can actually attribute where the code is coming from and then essentially explain the code as well so this is for developers or Advanced users who actually want to use it to build more advanced applications uh this is at the product architecture right now what we do is all the code that's generated by the um by the uh AI is compiled so any code that's coming out you know it's going to work uh we are also showing what's the code that's added by the generative AI versus what's actually coming from audited smart contracts and you have the option to put uh push the code to a GitHub uh repo as well from there what you don't see here are two additional elements that we're working on one is a open source llm uh model that I mentioned yesterday uh typically that's going to be a cloud model which is trained on audited smart contracts we use that to show your attribution and then the second model is the one that's going to be trained on vulnerabilities which can actually check the code and see where it's uh where it's not working so both of these both of these models actually checking the code generated by the core model which currently is the cloud which is the AI code generator in the future we're going to swap it out with a strawberry model which actually has like reasoning and uh which can understand context as well so the idea is you will have like a a model that's trained uh on uh and can actually do PhD level thinking building the code a second model which may not have have contact awareness but we've done what's called a CPT continuous pre-training that actually checks the code and then can point out vulnerabilities and then it all of that gets compiled so anytime you're building an advanced application uh it'll only take you down the path where the code will compile or it'll push you back to a point where you know your code actually will will compile and it's going to work so the idea is it'll almost work like a ads system or a lane keep assist with three models there's typically a close weight smart model that's building the smart contracts there's an open weight model that we've trained on audited smart contracts and another one on vulnerabilities and between the three of them they're actually keeping the the car on track it's almost like a team of developers working together to build some of more uh Advanced apps uh this is a road map you're going to have the turbo mode live um in Q3 along with the wi to prompt as well um next year early next year we'll have d as nfts and then thank you thank you um that's the end of the pitch so we've got questions coming up judges we're super excited about what you're building we're thank you for pitching again I think one of the things that got us excited about Blink was kind of the ability to do kind of Shopify Pages or whatever you like I think one of the concerns of the judges is around the code safety um obviously AI we're all excited about it but it's not perfect yet and we'll inevitably probably write some flaws in the code how do you plan to work with potential buggy code exactly so uh that's where uh the three model thing comes into play so so far what we are doing is we were compiling the code all the error that that comes through the main code generator we compile it but strawberry the models that came out last week is going to change that because actually has self-awareness and it can correct itself the second model that we actually building for the um for the attribution that can actually feed audited code back into the context window of the main model so next year we should see million token context window so it can actually feed things back in there and course correct and the third model which will will do continuous pre-training on on security vulnerabilities or you know if new security vulnerabilities we can put that into the third open model as well an open weight model typically a llama model between that you can any vulnerabilities that that model is detecting again you can push that back into the main model which is generating C so uh on the main uh you know the uh custom bepoke mode that you saw there's actually suggestions and it'll tell you what are the paths typically other people have gone down so when you're building more complex applications it can guide you through a path because the model is actually aware of you know been built before and if it's going in the wrong direction the other model which is on the security vulnerability side which is trained on that and continuously getting trained that can actually push you back it can push the main bottle back and point out the vulnerabilities with the existing models you actually you know couldn't do anything other than send it back with the self-aware models uh the ones that have uh reasoning capabilities can actually fix it and if it cannot fix it we'll just roll it back to a point where you know that the code is safe and Deployable and like I mentioned all of this code is compiled so you know it actually works yeah and is the plan always to stick to the kind of no code deployment of of tokens def5 Primitives and this kind of stuff or to move more to a co-pilot model later on yeah so we're looking at uh essentially we want to get Soup To Nuts right so as far as at least for the uh nfts and everything else it's mostly we can build the front end Pages like we showed yesterday and plug in all the rest of the elements I think in the future 3 to four years from now you won't need front ends AI most of the stuff is going to happen with AI agents anyways they're going to interact directly with the smart contracts so most of the focus could be just on building smart you know the best smart contracts and 95% of the work is going to be done by AI agents anyways so it's a direction that the SAS model is going in you know SAS typically the ux isn't great but it didn't matter because anyway it's used by you know large corporations it's not a consumer app but AI is literally eating SAS away because you don't need you know a nice UI or ux to actually work with a SAS model it can you know AI agents can do that so it's going to be we believe it's going to be the same with um you know with with uh the web three space as well as long as you're building smart contracts we're now plugging in the front end capabilities but if you're getting great smart contracts that's kind of the heart of this and then most of the rest of the stuff you know uh uh lending borrowing investing is all going to be done by agents um in the future do you think uh like let's say a project didn't use you guys out the gate but wants to launch some kind of module or like you know some sort of uh tangential thing yeah do you see a world where your technology can actually ingest what they've already built understand it and be able to integrate uh additional code and specific contracts yeah so we looked at it in two ways one is obviously the context awareness where you can bring that in so the larger context Windows the million plus context windows will help you do that but someone can actually build it the other thing that we're going to do is yesterday we show you the marketplace where people actually build stuff so we want to actually have a marketplace where people are building cool stuff and you can actually piggy back off of that right and similar to a Shopify model or something else it could be Revenue sharing all B in the smart contract that the Creator you don't need to start from scratch you already buil something and then you essentially take that and build other things with it so that's the way we look at it and essentially a lot of the frontier models are going to solve it the context windows are going to get so large that you don't need to do a lot of pre-training so even if you have a Clos weight model you can take the two open weight model like I said train them on specific things and the close weight model you can essentially you know train it to do the things that you want it to do and it can import contexts from existing templates or someone can bring cool stuff that they see open source if the models already don't know that and then we on top of that we're out of time thanks thanks very much all right thank you