AI on the Ledger: Transforming Blockchain with Artificial Intelligence - TOKEN2049 Singapore 2024

um and so one thing we in a sense wanted to start with or try and Infuse through the conversation was try to reflect on what has happened since a year ago in this particular area the kind of that intersection of uh Ai and and and blockchain should we call it that um and I wanted to start with um uh kind of my perspective which is that uh that in some senses there is a a heist going on um in the shape of the kind of both you know existing reorganization and impending reorganization of open AI as it seeks to go and raise 5 10 billion for the next phase of its kind of its roll out we're seeing something that was designed to be owned by the community um uh and to therefore be protected for from uh too much bias uh being made purely for profit because it's you know at least in the short term it's quite hard to raise that kind of money in order to build out uh these AI systems and so um given that and given that um we were discussing this this case just before now um there was a Stanford Google use case of a a simulated world called Smallville um where you get very they use used open AI chat Bots interacting in a quite complex agent to agent fashion and they end up um actually having a party for Valentine's Day and dating each other and exhibiting all sorts of quite complex behaviors that's just an example of how complex and sophisticated this is becoming and how fast um so we need a lot of money and we need to understand what's happening in these complex systems and so we're here to talk about why that could well be achieved by the intersection between these two worlds and to look at the question of how that might be achieved in terms of providing compute power CPU gpus you know xpu um that are can to be coming down the pike in in assembling the data that is used to pre-train train augment um these um these uh models in actually finding ways to to derive economic benefit for a broader set of people who could contribute to these models um and then also perhaps ways in which we can uh create um a low friction economy between these agents um as well so with recommendations reputation and and payments so those are kind of the topics so before we thrash through those extremely fast um I'd like to turn uh to the the panel to introduce themselves and I didn't introduce myself because I got interrupted Ed so many times but I'm uh managing partner of fabric I'm also on the councel near to declare my um affiliations Jamie do you want to go first uh yeah Jamie Burke founder and chairman of outly Ventures with the world's leading web3 accelerator we've celebrating our decade in web 3 this year there a lot of echo there you getting that too is what it is um and actually we've been investing at the intersection of um crypto and AI since about 2017 uh we published a paper called the convergence thesis back then we're about to publish something called the post web which is looking at the impact of the agentic web on the internet um and we've had a number of investments in this space probably the most well-known one is fetch.ai and its merger with ocean protocol and and Singularity net into artificial superintelligence great to be here I'm Jake Brookman I'm the founder and CEO of fund which is a crypto focused investing firm um based in New York and um our mission is to Champion the leaders of the new internet and we've been paying very close attention to the intersection of AI and web 3 made our first investment there uh in World coin in the early in early 2021 have a number of other companies in our portfolio including Jensen giz Bagel uh and others uh in the portfolio and following the space very closely so excited to talk a little bit more about it hi um my name is Terrence qua I'm the founder and CEO of humanity protocol uh we're building identity Focus blockchain where the very first thing that we're focused on is what we call proof of humanity so a little bit similar to uh to whirl coin in the sense that we're using Biometrics uh for proving that a person is really a person and we think that's going to be more important in uh as as as AI proliferates um but going beyond that what we want to do is actually create a token economy uh around verifiable credentials so that we can put different human attributes uh onto the blockchain so um yeah hello I'm Hugo co-founder of flair FES 1 evm blockchain with enshrined data protocols um where data Protocols are secured at the chain level thanks guys okay so sort of in some kind of Kumbaya philosophical sense we can agree that we want to maybe move things in an open Direction which is kind of what the intersection is about can you hear me um and um I think it'd be interesting to explore though how there can be technically Superior um Solutions offered by that intersection and even better than that hopefully things that deliver better economics short medium and long term to the participants uh in these open decentralized you know AI networks so who would like to get going um on that debate I think Jamie we can always rely upon uh in any of those four areas compute data models payment yeah well I think just to speak more generally um you know the reality is how we're all experiencing the kind of boom that is AI now is through these large language models that have effectively been trained on the open web and so a lot of that benefits kind of already been had um and so the real value is in providing context to both the llm and and other kind of models uh and often the context is in the longtail um it's with individual users it's with different Enterprise that have you know these data set proprietary data sets um and ultimately it requires collaboration and so you need incentive design to figure out how you're going to make that happen and so the kind of meta question is how do we get to Super intelligence is it going to be through some kind of Monolithic um organization that is coordinating Capital Financial capital and human capital um and kind of capturing or acquiring data and it's going to deliver this monolithic you know singular brilliant AI That's a specialist on everything um super specialist on everything or is it more likely that we're going to somehow find a way to coordinate the longtail of everybody else to open up share data train models um and then provide context to those models so they can be more relevant more personalized and not hallucinate um and so of course whilst it might end up being some kind of hybrid the reality is it's the collaborative AI piece that I think crypto speaks to really well because what respond really well to logic isn't actually people or organizations it's agents and so if we can look at agentic networks as both coordinating and and kind of playing these incentive games to develop train customize personalize the experience of super intelligence um then I think we get there through web 3 and so how do we coordinate trillions of dollars of resources which is what Sam AR wants to raise into a single company um or is it more likely that a Bitcoin like ethereum like set of protocols coordinate that in a trustless way and so I'm betting on the latter I think it's going to be the more likly way we get to Super intelligence but also it's the more likely way that we get the experience of super intelligence as an end user because I need a paradigm where I can trust the AI to share my innermost Secrets needs desires intents so I mean the contentious point when we're talking about the compute side of things is can we actually deliver something that is uh cheaper often is the word used than a the fully scaled AWS or Google Cloud kind of platform so jakay what's your thoughts on that so I I know that um in web 3 we often make that case especially with you know decentralized networks for things like compute and storage and I think actually the reality of um kind of expensive high-end computer is a little bit overblown because if you think about it um well what I'm saying is I I I I think compute is actually going to be a lot cheaper in a few years than it is today right because the market is responding to this bottleneck and demand for compute um and optimizations are happening in AI where uh you know we're starting to be able to if not do then at least like think about doing for example decentralized training on more commodity Hardware um and so I actually in summary don't think that having a cheaper compute network is the dimension upon which these networks will really succeed and compete I I very much think that that Dimension is openness and I think everyone basically in the world has an interest in having the process that produces AI models and AI pro products uh to be done in an open way right um like to the extent that you know we want to create intelligent models I think that can be done more effectively uh through open source we see that today we see a lot of Open Source contributions even in the current environment um and for you know like do we want AI to be created by you know a few hundred or a few thousand engineers in just three four or five big tech companies that are all located on the West Coast United States or do we want the world to be working on AI and then to the extent that AI is dangerous you know do we want a few open AI Engineers working on the safety problem or do we want like tens of thousands of Engineers globally working on the safety problem and to the extent that that problem cannot be solved uh that well um I actually think counter toly the answer is to have ai be even more open so the dangers of AI models can counter balance themselves so you know to me like openness is the thing that makes a lot of sense and um I think this move uh towards distributed training for different types of models and not just large ones and also small ones you've companies that are specializing in that as you said on kind of commodity Hardware um is then opens up the the possibility um for building these open systems or maybe it goes hand inand that in turn gener gener Ates the trust that um uh that Jamie is talking about and people are more prepared to give up their their data there is an argument that if you then do that you move you know the compute becomes a more kind of distributed compute environment you can tap into different data sources and that that those data sources um are they're going to give better trained models and better uh results and and Jamie on on your point like the open web has given us as maybe as much as it it can but put a number on it it's like orders of magnitude more data is actually held within social networks for for example let Al on the data that people might be encouraged to share if they they have a reason to or they trust the systens better so maybe I don't meaning to to skip you but I know Hugo is have some some views here um in that that future do you do you think that we going to be able to get the right data to the right place at the right time and get a better outcome I mean I think it's a it's a worthy ambition um to do decentralized training of these very large models you got a lot of barriers to getting there um you know privacy if you're sharing data who are you sharing it with how are they processing it um how are they processing it whilst they don't know what the actual data is because that would be the more desirable outcome um it's going to cost at some point certainly now uh it's not feasible to do decentralized training as cheaply as with a a data center um so to build this entire infrastructure out um take a lot of time and a lot of money just to get to the the through the research and build to be able to enable decentralized training um at which point where is open AI at that point um you know it's taken us what seven eight years to get like basic ZK working quite well um there's a lot more components that probably need to go into this kind of uh stuff so um here we like hard stuff though don't we like you know we want to be on the on the frontier well there's no point being on the frontier if someone's going to beat you to the P punch and you're never going to be able to get to really ship it cuz the money runs out um no it's a good point so I guess the question is can we is it not just about achieving it but doing something that is not possible for the folks doing the closed approach to to achieve accessing you may be able to get uh different way of man managing governance I you know you're not going to be able to get um you know the same kind of governance with a centralized player as you could with a a decentralized network um but whether that results in a better product in the long run is is as yet still an open question so actually it probably relates to kind of the uh you know validating humans a little bit um it is probably conventional wisdom that in Venture Capital that making bets on projects that rely on privacy as they cor value proposition is has not always been a a great one to make basically because of human nature which people care about privacy except only when it's too late so therefore they don't buy stuff um but at the same time we've started to see individuals and businesses when they're looking at the the prospects of giving up their very personal very very personal or proprietary information to open nii to to hesitate to hold on a second and the latest version I had an upgrade and you know it's now got a memory of of my different interactions and it's using it to compound its personalization for me and you're like do I want to do that so perhaps that is a core part of the the proposition for an open approach if people can be brought to to trust trustless we did that to ourselves um then maybe that would work Terren what do you think I think you know when you know I I I come from sort of a consumer Tech web background where I think ultimately when we sort of look at users of any sort of product or even when we're not talking about developers using AI but actually end consumers who are using Services uh I'm not sure they actually care that much whether it is centralized whether it's decentralized um you know I think with with much of blockchain Technology right now um there's a lot of infrastructure being built but there's not enough of a focus on building applications where people actually want to use or find it easy to use right so even when it comes to I think artificial intelligence you know I think Richard you mentioned people care about the Privacy people care about all their data being used again for for training etc etc um I'm I'm not actually sure people really care that much most of us have an Instagram account or Facebook account you know you're still using chat gbt because it's probably better than the other models out there so I I think and that that's that's for People Like Us whereas if you look at the general retail user consumer I I don't think they care they care even less they care they definitely care even less right and I think for for what we're trying to do um it's it's you know what we're trying to build is sort of self- sovereign identity so that whatever data they're they're they're they're providing they can revoke at any point in time um they can control it so it's it sort of plac to your point about you know potentially people would care but ultimately for us it's still focusing on creating an easy user experience right so GPU maybe will achieve it of the GPU level economics on on the data side um we better not rely on privacy so it better be uh better data uh what about in terms of you know models like Creations intellectual property so which includes the models and agents themselves does um do we see a role for web 3 in in that space and I think we can agree we're looking for a role that is triggered by a fundamental value proposition not a kind of philosophical desire to keep things decentralized well I mean I think to your point people just care about a superior produ and so this has been the disadvantage of web 3 I actually think even calling it web 3 on reflection is a bit of a misnomer because what we've done is we've in principle upgraded the functionality of the internet but it has no ux it's unusable for most developers it's unusable for most end users unless there's some compromise on its web freeness which kind of defeats the purpose and actually what's really interesting is in this decade that we built this infra internet infrastructure in parallel there's been advances in natural language and llms and I think when you combine that with a gentic networks and you imagine that the the direct user of web 3 infrastructure isn't a person it's an agent all the onchain stuff is done by agents the user interface like much of the web will become conversational natural language then actually all the complexity of web 3 is removed and its benefits can be revealed and so kind of coming back to that context Point um I believe that if you imagine a world where you don't need search engines anymore you know most people already replaced most search with chat GPT um you don't need as agents roll out you don't need an app for everything you don't need to be going into an app manually you don't even need a website for everything you don't need price comparison websites you don't need interruptive advertising you just need an agent to work with other agents to solve problems optimally and the user interface just needs to be as minimalistic as possible um and and so really for that to happen the user needs to provide as much context about themselves their needs the environment that they want to interface with with as possible um and so you know to do that to be able to do that effectively to have portability of context control and permissioning of context you need web 3 so I think the way you get the most Superior experience from AI will be from leveraging this web this kind of latent capability of web three infrastructure that makes a lot of sense so Sarah Sarah Tel over at Benchmark tweeted the other day that you know we're entering a new phase of kind of there's an app for that and it's going to be there's an agent for that um and I actually did count I've known Sarah for a while but I counted on on on Twitter X whatever um that I don't think it's there's an agent for that it's that there is an you know the user decides to go and seek it's there's an AI service that that identifies that there's a participant who might benefit from that service proactively and comes to to the participant effectively and that might be a participant that can contribute something supply side or will consume some of the service kind of both um and so exactly to your point then the how you interact with that is going to look quite different from an app in fact there's probably there is no app there is no app yeah I mean it's it it flips from an atttention economy to an intention economy so I think the next evolution of web 3 infrastructure is intent based architecture so I have an intent um and then the world organizes around me to solve that problem my UI will just be my preferred AI interface my preferred llm but the problem is solved by a collective of Agents so so if we take that further because sometimes it's useful to imagine what this sort of endpoint looks like a little bit to again come back to these sort of technical questions what's it going to look like on our our our phone so we're if there's no apps presumably there's no app store and um death of search death of the App Store death of inter advertising and then we're sort of talking or thinking at our phones what guys we still have landlines you know like you still have a landline I actually don't have a landline you still have what sry uh no I was just going to say um like technology does take time to turn over I don't I don't think like there won't be any apps you know a couple years from now I think we'll be using apps for a long time but it's a it's a really interesting question because even today right like there are uh a number of teams who are building products that look like this you have a chatbot which allows you to translate your natural language into as Jamie says intents for trading in crypto markets right so you could say something like I want to trade uh ethereum for salana and I want such and such an exchange rate and I want to bridge the result of that transaction to some other chain and so on and then the llm will translate that into blockchain transactions and execute that for you but what's interesting is you know is that really the consumer or retail interface for doing that like who is actually the customer of that and I'm actually quite skeptical that even though it's easier that normal people want to interact in that way so I think there's still a huge use case for you know just having a panel in your phone and trading the old way I mean history would tell us that normally it's not like everything gets replaced either quickly or totally so I I would you you know tend to agree for for what it's worth um but in in trying to Envision sort of where where it's going um maybe we can have that mixed vision in mind and and think about the next question which is um how are agents going to bank themselves how agents going to pay pay for things in a way that we're going to feel comfortable about or that will technically work who who wants to pipe up on that topic I'm happy to Twitter said well they can just use stripe um but uh there's a lot of crypto people or there's a lot of reasons why using crypto would be a lot cooler faster cheaper more efficient and I I I I actually think you know when you when you ask the payments and and sort of the banking question ultimately these agents have to be controlled by people right um do they I I I have a hard time believing that you know I say I want to buy a car and then you know I have money in in an account and then I just let the agent decide what's the best car for me I maybe maybe smaller transactions maybe smaller things potentially um but I I actually think people still want some sort of control and that's why I do think advertising will actually continue to exist um in fact you know one of the my the things that we're focusing on with with Humanity protocol is you know making sure that people you're advertising to people instead of AI agents because ultimately it should be people who are making these decisions yeah so let's see take a concrete example in the last 1 minute 20 seconds um so you may not want to buy the car or allow your agent to buy the car on your on your behalf especially if it's interacting with secondhand car salesman um but maybe it's allowed to make a reservation or make an inquiry or book or book of of viewing on your behalf in the run up to that that final purchase and certainly to do a a whole you know once you've set the your in um go and look at the type of car you want and keep track of that over time so and in that context uh I think we can see can't we sort of small payments being required um not just to say secure a booking but you know but also to um be allowed to use the underlying infrastructure including other agents themselves um so that's fair they might have to so they won't by the car but they're still going to have to do some kind of payment but maybe not through a stripe API why not through a stripe API can we be concrete I mean I I just personally think that like um people want to use newer Technologies even if it doesn't quite make sense right like like in in 1995 you know if you told someone like hey you know you're G we should order some pizza by using a computer like they thought that was really silly um but now we only order pizza through computers um it's a it's a social movement in a way and it's it will be the same with agents they will use crypto because it's the cool modern thing to do and eventually at scale it will become the more efficient thing to do yeah exactly so we're going to have to struggle to get out of current practices but I do think ultimately if you believe there's sort of possibly no phone but certainly no app store maybe no no apps um and we we're getting rid of some of those centralized intermediaries because perhaps we believe that a an intermesh set of networks that are trustless is something that we do trust more um that at that point in time you're not going to then want to impose stripe as a trusted intermediary on the payments framework for the agents it's just not going to be compatible with this new approach to doing and you know if we're trying to mesh the two we're going to struggle to get the full value out of the whole thing so that means it's more exciting but tougher to get there as as as you said I think the difference is the difference is between optimal outcomes and convenient outcomes so people seek convenient outcomes which is why we like using the same app or the same website or the same brand it's consistency um agents like Optimal outcomes and they don't mind complexity they can navigate complexity logically humans are logic so if you look at the consumer internet and the things that we do on it and you kind of apply maso's hierarchy of needs there are some things that we love doing like buying a car usually for most people is quite exciting you know they like looking at different ones and looking at the performance and stuff but um why do I need the Uber app like what function does it provide I just need a taxi hey you're running late do you want an Uber yeah book it for me it knows that I like premium or like whatever the parameters um and there are so many examples on consumer internet where you just don't need to do it um it's not pleasant it's not pleasurable and most often than not you might not even be getting the optimal outcome and so I think if you're looking at what gets disrupted you know versus um what stays and gets more immersive it's kind of on that Spectrum some things I want to do in a highly minimalist way some things I want to do in a immersive way anything that's minimalist will be unbundled by agents so it's going to make um keeping your calendar up to date even more important so you don't inadvertently walk out of the conference and automatically find yourself in a in Uber to the airport when actually you've moved your flight for for example or something even more confusing or embarrassing who knows um I think on that note um I must bring the conversation to a close any final thoughts happy too I mean I I I think it's a unbelievably interesting area to apply to blockchain we didn't really get through that much of like what can the use cases be what can blockchain do for AI rather than what AI can do for blockchain um but it's I think it can be very interesting thanks very much everybody thank you