The Convergence of AI and Web3 - TOKEN2049 Singapore 2023

hi everybody um what an intro it gets better and better um so very glad to be here and a little bit of the cool um uh AC very glad to be uh here at this event which seems to be a bigger and better than ever especially given the kind of the current part of the market cycle um and great attention to detail they're even checking whether I wanted my shirt tucked in or not backstage so congratulations to everybody who's been organizing it um so we're here to to show that um this whole kind of AI web3 intersection uh you know truly is the kind of the hottest topic um and that's um you know and it's not just some kind of IG or kind of lens or forecaster kind of meme uh some description so just to I'm going to ask everybody to on the panel to introduce themselves in a second um but for those who don't know me or kind of fabric you know we just strive to be um selected as kind of venture contributors this is the way we kind of look at it for um sort of daring if not maybe slightly crazy Founders in the kind of open web space and we've been going for a decade and working with um personally with folks like near co-investing with with Jake um and polka dot and ramp and immutable and so Brown all sorts of great folks like that um and uh um I can see lots of founders of crowding around the VCS on the side here already of the panel so um I what I'd like to do is to do these intros now something short and Punchy and but hopefully also kind of powerful um that uh also maybe explain some of the biases of your respective data sets as you train these these uh agents out here in the audience as to how to think about you um and then we'll kind of we'll go back into a little kind of layout for the conversation uh and then we'll kind of pass it back to the panel so uh why don't we start with Alex over the end here thank you Alex blania so uh I started working on Royal client three and a half years ago with Sam Altman who is the founder and CEO of openai and right before that I was in theoretical physics and applied AI systems to predict Quantum systems so I have like a little bit of a technical way less than Elia has but some technical background and a lot of thoughts on how this is going to go forward hi Amelia uh I prefer the famous rabbit hole of crypto I was for about 10 years working on machine learning and understanding natural language and knowledge and ending up working on Transformers which is the kind of core technology powering all of the cha GPT mid journey and all the current kind of innovation and then through a twists and turns end up starting your protocol which is a kind of layer one blockchain that's focused on really bringing a billion users on chain becoming an entry point level three thanks for having me my name is Jake Brookman I'm the founder and CEO of coin fund we're a crypto investment firm whose mission is to Champion the leaders of the new internet and we've been doing so since 2015 about a year ago we published our thesis on the intersection of web 3 and Ai and we've been making Investments actively in that area ever since so excited to have this conversation today Jamie Burke CEO and founder of outlier Ventures we've been around for about a decade now the world's leading web3 accelerator one of the most active sea stage investors in the space um and ai's really been at the heart of our thesis we wrote a paper back in I think 17 called the convergence thesis um and a significant part of our portfolio we've been exploring that convergence the intersection of um data and web3 Ai and blockchains um and excited to talk on the panel fantastic and so as we heard you know we sort of individually have been um interested in driving this intersection for the best part of half a decade we've got kind of you know more than half a decade of kind of crypto OG activity here as well as even you know llm Transformer activity so I think it's fantastic to have this kind of panel here so we can ask the question um you know what is the essence of of why this matters the intersection um and one way of thinking about web 3 is that it's about weighing up the the and balancing the interest of uh Sovereign individuals with free markets and with the the collective good and we all know that if you're kind of trying to weigh up and balance and manage good relationships trust is important if the trust is gone your relationship can be toast so a trust layer of some description is required something that can get a grip around the identity of uh individuals of humans but also the identity of data and Creations in some sense um so I think if we assume that's that's correct we also have the question of like well should that be centralized or is it absolutely essential that it is decentralized and not beholden to one individual God forbid or even you know one individual government so on that topic I want to pass over to Jamie who's he said he's been talking about and and thinking about this theory for a long uh period of time and and unpack um you know some of those questions yeah well I mean I think if you look pre-etherium or at least early ethereum a lot of the language if you go back in the you know time machine a lot the language was talking about truly smart contracts it was talking about truly autonomous Dows and so that's always been at the core of this essence of web3 and of course if we're really thinking about the next phase of the web you can't not think about AI being at the heart of that um and for us equally you couldn't look at blockchains and isolation which led us to uh come up with the convergence thesis but actually that was just informed by a lot of the innovations that we were seeing happen um actually a lot of my thinking was influenced by something called Trent McConaughey of ocean protocol investment that we share we were kind of friends and family around at the time who was the first person I met who was an AI person a legit AI person two exits under his belt who was coming into the world of web 3 even though we weren't calling it that at the time from the perspective of somebody who was thinking about designing in the future of AI and that opened up a lot of a lot of questions that led us to make several Investments Beyond ocean into fetch dot Ai and a number of others um were effectively um blockchains we're going to enable the commodification of data and of course uh what was going to be consuming that was was going to be AI that led us to begin to think about the direction of travel that that could take that would be different to the world of web 2 which was slightly kind of more monolithic and and centralized in terms of platforms so from our perspective we've always kind of thought of this leading towards agent-based systems and then the question comes well how are those agent-based systems own controlled how do they serve individuals enforce their sovereignty but also operate a collective level maybe or maybe not leveraging Dows and so I think that raises a lot of interesting questions about value creation value exchange and value ownership and offers a potential Paradigm where that's going to be more Equitable than the world of web 2 and I know that's probably something kind of we all share as a as a general framework any concrete examples from your portfolio that you want to cite yeah I mean I think fetch.i is is probably uh the better one or the more mature one um they have originally covered it from a kind of industrial context again a lot of the early thinking at least in Europe um from the kind of early ethereum Community was very much looking at the machine to machine economy so already thinking about blockchains as being a means through which machines can transact with one another make micro transactions obviously always on the assumption that there'll be kind of these scaling Solutions and so they do a lot of work with industrial organizations like Bosch for example on open Mobility systems where effectively if you're going to have autonomous vehicles navigating um navigating the world how do they uh pass information between one another transact and of course that gets even more complex when you think of the user's experience which is probably going to be multi-modal Transportation so you might start up in a car but you then might get on a train you might take a plane how do all these things organize around you as the user and their vision is that through this kind of distributed or decentralized Marketplace of Highly specialized agents yeah so so we're going to come back to that question of machine to machine interaction and even sort of just software agency software agent um and and also just to think about this whole Supply in this whole supply chain of you know the creativity that AI can now to generate how do we know that we can trust it there are going to be elements throughout the supply chain from the training through to the proof of persons and so forth that we need to address one area that's popped out of trying to address some of those challenges is zkml the combination of machine learning with um you know uh with zero knowledge proofs um those are sort of two pieces of Wizardry um and so I'm going to turn to Jake to talk about um what is the state of the art there what has he got excited actually you know happening and and what are the kind of shortcomings if you currently sees it absolutely um well zero knowledge proofs are a technology that among other things enable um verifiable computation and where this is finding the most product Market fit and actual use cases in today's market is actually within blockchain applications like historically blockchains and other decentralized networks have been fairly low power computers and when it comes to doing um you know intensive computations the tendency is we want to take that off chain but in doing so that computation becomes performed by a third party so how do we how do we trust this third party how do we know that they're performing the calculation that we're paying for and that they're doing it correctly there's been a bunch of solutions to this problem over time but zero knowledge proofs have turned out to be a really precise solution and so essentially what happens is this third party performs your computation and then provides a proof that it was done well um what we're seeing what we're seeing in the market is that this off-chain computation one of the most prominent use cases of it has been in Ai and there's a lot of people right now who are working on the problem of training models so in other words creating models in a decentralized context on decentralized networks and there are people who are working on the problem of CR of inference of models meaning uh you know using ZK to show the outputs of models and to be able to use those outputs within smart contracts so zkml is sort of the broad field of applications it's using zero knowledge proofs essentially bring AI on chain exactly so the different steps in the supply chain that ZK can be used to make sure that we can we can actually if we go out there and we want to get a result from AI we know that we can trust that the training was done on a certain data set um and that a certain model was then used to deliver us our results um and and we can trust the output to then do whatever we want in the kind of next stage of the of that sort of data supply chain if you will or inside supply chain um so I mean we're talking about kind of large language models um uh you know Ilia was instrumental an author of the paper back in 2017 tension is all you need about you know Transformers and um and so what I thought would be interesting to do is is to try and unpack this sort of dichotomy we've got between the incredible power of of these new large language models and and the challenge this creates in this um this you know data supply chain we're talking about um so maybe you want to dig into that earlier and again favorite examples of of of of where we're actually seeing code you know in operation in in this convergence yeah so the way to think about uh large language models is it's something that um kind of like deep learning in general is very expressive right and is able to capture a lot of knowledge but kind of historically we're struggling to train this at scale and struggling to get it to actually work in production so the reason why for example like Transformers came around is because we were actually not able to launch what was a technology stack before on google.com because it was just too slow it was uh would take you know like seconds to actually run inference even at smaller models that were running back then and Google just cannot do that and so the kind of what Transformers as a model allow us to do is just really align well with the current kind of silicon we have right versus parallel processing kind of Graphics units that specifically Nvidia producers but also now there's other companies now what fundamentally happened though through the work of open AI especially is this transition from you have a human in a loop right that needs to input information and then process output and interpret it and explain it to everybody else to a new way of interacting with Computing where the this models themselves are able to communicate with people even at scale right by for example posting on one of those you know social platforms and that kind of fundamentally shifted a lot of the things that you know before we would assume about Computing and that's why all those conversations started around well what does this mean for Humanity what does it mean for kind of media for content for all those things now the kind of From perspective of where those things starting to intersect right and mention about the identity interpretation is really important to understand how kind of this output came around right so if this is an output of a model and we want to prove it right that's important with ZK if this is uh an input data that goes into a model we want also want to know where it's coming from but at the end as a consumer what you want to know is that whatever you're looking at is actually associated with some reputation right and right now we actually kind of as a society moved to you know like we had non-secure Communications through HTTP and we kind of converge into https into secure encrypted communication with websites because we wanted to trust the content we see from the website is coming from this website and not being intercepted in the middle right and so similar thing needs to happen on the level of the uh kind of content not just website where as you see through on some book on some blog posts into you know a video YouTube video podcast you want to know that it was produced by the people it says it produced and that you know know kind of this reputation is has Association and so this is where the kind of blockchain technology is a very important tool because it enables us to use kind of a lot of the cryptography Primitives but then root them into some level of trust right can continue to aggregate reputation across multiple instances of you know content production such that when you look at some piece of content you can see what are the other pieces uh or what are the other interactions have happened and we see a lot of this happening across the centralized social media becoming kind of the you know the driver for this as well as you know some of the media companies are starting to think about how to how to adopt this so so good so we've um we've got uh software that can talk to software we've got we can then trust the output of that and then we can uh know when information has come from humans we can then uh trust that we understand where it's come from and how it can be connected to their reputation um some part of the link in that chain or value system or whatever has got to be the connection to the people themselves and proof of personhood and and I think um you know Alex and has worked with the world coin and tools for Humanity is in a very well positioned to talk about that so um ultimately when you think about all those elements together it's it's we can Envision looking back from 5 10 20 years in the future a new ways totally new ways in which we're going to organize ourselves and our and our lives and hopefully uh um um yeah we'll maybe we'll come back to the question of who's on top humans and machines and in a minute um and um in that context you know governance is extremely critical and how you possibly make use all of these elements to enhance ultimately democracy I mean one could argue that maybe that isn't a very high bar to make better democracy as desirable as it is better than it is today but love to hear you kind of unpack that and how that might the different techniques the different tools being used and anything else you want to cite as an example Go Alex sure so um I think as we are as Ilya just described um We Believe proof of personhood itself is going to be a fundamental primitive that we are going to need on the internet for a lot of different reasons one of them is just verifying online that we are actually humans and that that will be important for many contexts like I think Elon talks a lot about this on on X or Twitter about bots so these Bots what has changed is they will become much more sophisticated and much much smarter and of course we can rely on government systems for that but that will not work globally and we probably should not do that and I think crypto actually has a way to build reputation systems to build personal systems that work truly global and decoupled from government so this was one of the big motivations where we started working on it but then um going a little bit in the future which was obviously a big part by Sam started this company in the first place is uh Democratic governance over very powerful models and very powerful systems um because right now just taking openai as an example there is a small team of Engineers sitting at San Francisco making very consequential decisions and these decisions will be will become more and more consequential and that is just one cultural context this is just one small team and one of the main things that I think crypto can help with and likely will is how to democratize the access and the decisions over over these very important parameter changes and step changes that we're going to see in the coming years and then of course another one that is always a little bit controversial but we also believe will happen is the chance to implement some kind of well-free distribution Ubi because you will have a lot of singular entities that will create a lot of economic impact and power and distributing that with the world is going to be critical so these are some of the motivations uh that drive us thanks sorry thanks Alex I mean I think we maybe you can think about whether we can um start to do the math on uh how much fun could earn uh from Ubi given that I mean because that's can be challenging but if you want it we can talk about that but you know it's good to hear that one of the intersection points between web3 and AI is the fact that over the last decade uh necessarily there's a lot of work being done on decentralized governance because it's one of those core tenants within web3 and that now can be can come across into to AI uh nonetheless there are a lot of challenges in making things like dowels actually work and I know that Ilia you've been exploring opportunities to use you know AI agents to actually you know turbocharge the capabilities of Dao so maybe you want to drill into that um and then we'll move back this direction so jql primed yeah I think I mean can a blockchain's history has been a big experiment on governance right because on one side we have created technology that is designed to coordinate humans through protocols by defining you know instead of incentives and you know potential admission and getting people to be kind of contributing into that at the same time as a as that mission becomes less nebulous from produce more blocks it becomes harder and harder to do that and so we've you know the decentralized autonomous organizations Dallas have come come up as a tool to organize that but we've kind of influenced a few hype cycles of dials at this point and we've seen that they don't fully function uh because we still have people in a loop that are need to coordinate this and need to play the roles that are very similar to the roles in regular organizations but they're lacking all the tools that regular organizations have from like hiring people to firing people to being able to like do Performance Management and so kind of the reality is that like we have a lot of the kind of fundamental instruments that on blockchain are really good at allocating resources having transparency being able to kind of inspect reputation now of people who are participating kind of up from the perspective of this Mission but we don't have this coordination kind of execution and we relied on humans and and kind of people are way better at doing things than at coordinating large masses of people and so the idea here like can we start using some of this AI technology and specifically kind of the agent focused um AI that is now kind of serves as this execution slash like executive of the organization right they can be able to have all of the context of all the things that happen maybe able to onboard new people give them the context provide them the swim Lane where they should be contributing provide information and kind of reports to all the stakeholders and still be able kind of under the oversight of the governance of people that are kind of set up this Mission or you know participate in a token or whatever the way to coordinate that and so that idea I think is a how to revitalize dials but B I think generally is moving Us in the direction where um like AI is now serving the function that people are inherently bad at and as well they provide kind of this context and information and I think there's an interesting like uh philosophical question here which is you know people like oh AI will take over the world and this is bad versus my statement here is actually economically we will choose to work for AI is you know in some form of shape and it will be better for the whole kind of humanity because it's more effective to coordinate people that way well we still have the rage grid opportunities we choose to work for AIS and then we find out that we don't like that anymore you can reach quit into another AI certain options um look this is going to obviously prove to be absolutely critical debate closed versus open AI um you know you know there is this sort of gravitational force that you know within AI that tends to centralize the the data sets and if we don't have a way of kind of balancing that out it's it's going to be probably quite hard to to resist before we move on to the question of uh kind of uh improved economic coordination specifically uh driven by you know concrete use case in this space I know Jeremy wanted to build on this dowel Point yeah yeah well I mean I guess if you kind of pose the question um who are blockchains for I would argue it's probably not people in the day-to-day use of them right so who what a Dows for um if you imagine that actually blockchains and Dows are best used by agents they respond very well to logic unlike people um they respond very well to incentives people sometimes do people sometimes don't uh and actually the really interesting thing about web free if you think about its principles and why you would want those principles as part of the DNA of the future of AI is they have the potential to reconcile things that have been at odds with one another for the last you know 100 plus years which is free markets collectivism and the sovereignty of the individual and web3 has the potential to harmonize those three things in terms of collectivism a form of fluid collectivism but that respects the individual um but again if we're going to have tens of thousands hundreds of thousands billions of tokens incentives coordinating these networks these games I want my agent to play these games on my behalf I don't want to be having to make these transactions and swapping various tokens I just want to live my life and so that's kind of the the perspective that we come at when we think of AI and so we you know when we were kind of having a chat in preparation for this if you take that notion of a an agent that can understand your risk appetite and your resources and then make the economic decisions on your behalf to kind of optimize yield given a certain you know risk if you take that Atomic unit we could dream about actually doing a better job than the economic economists currently do at the absolute macro level and the global macro level so Jake I don't know whether you want to kind of you know build on what's happening today in that space and where we think it's going on chain well I'm not sure that if we yet have um startups that are doing global economic Central planning but I guess the the seed of the idea is that you know we we often put forward kind of capitalism right as the best um societal mechanism to create Commerce and uh products and uh Society really right um and but if we take a sort of dispassionate view of capitalism there's also a lot of waste I mean think about the typical Venture portfolio right where um you know 24 out of the 25 companies go to zero and one Returns the fund um and on the other side of the spectrum we have sort of the classical you know communist Central planner who we also know doesn't work so maybe the kind of the seed of the idea here is can we find a middle ground there where uh there's a planner but the planner is so intelligent and so um All-Seeing of the details of the economic activity that is happening that they're actually to your point are able to coordinate uh economically uh in a far superior way and and maybe this idea of Agents right um is kind of a starting point there of you know one of the really cool things we're seeing in the market right now as an application of AI web3 um at least I'm seeing this more than anything else is the application of AI to financial strategies and this is in a way you know kind of a financial planner that works on your behalf looks at your view of the market or of uh I don't know yield protocols and D5 whatever the case might be and says you know this strategy for you right now is going to deliver kind of an optimal return and and so there's a macro view there's a there's also this bottom-up view of planning as well so I'm definitely a born Optimist and I get excited by this this notion that we can do a better job of managing the economy for everyone and reach a kind of new a new Optimum um but it does interestingly if you have if you build it in this emergent fashion from all of these agents with our individual interests at heart and in a very narrow way first but then we combine it um it then speaks to the challenge of composition across this environment because we want to know that those narrow data sets that are having narrow objectives you know you know are actually based on the the correct data and and so forth and that the outcomes can be interpreted by other agents so there are still challenges as to your point to make to make this work effectively um so I actually forget whether I was told whether we were allowed to take any questions to the audience but I'm going to provoke you to come up with some anyway and maybe we maybe we can you can shout them and we can repeat them but in these last few minutes we're gonna kind of you know shift attack I guess a little bit um uh and talk about um the topic that sort of governments seem to care most about in this in this context um and I guess about how perhaps you know understandably with our own biases the web3 community thinks that we can contribute to what is termed as a kind of AI safety um thinking about the alignment with the objectives that we have um uh make sure human and machine objectives are aligned and there's a there's a big Summit coming up um in the UK in November run by the government a global one and I hope it's as inclusive as as possible um but so maybe everyone in the the panel while the organ the audience thinks of the questions wants to just talk about um how do we envisage this shift um where you know agents say might have to click to be clear on the the bias inherent in their data sets to to take responsibility for the outcomes that they suggest to other agents and to people and the liability that might go with that and then perhaps be involved autonomously in negotiating Insurance around that you know did you misdiagnose some health condition did you misdirect some some Financial strategy who wants to go first in some thoughts in this topic go earlier I can start I think sides of the problem here there's one side which is this is Agent or this is model that I use or represents me and I wanted to behave with like whatever my expectations are right and that is currently more what let's say open Ai and others are solving where they present it as a product right so they kind of have this as a something that represents their company and they need to make sure that that behaves in a way that you know aligned with their company values and at the same time doesn't you know doesn't do harm now if you look at from a global perspective though I think the the reality is it's not a AI safety problem it's actually the kind of human alignment question that we need to solve because the challenges of you know generative content fake news you know like they all span to you know beginning of humanity right in lab three we actually have Byzantine Soul tolerant as a core and because there was a misinformation you know in Byzantine Army and they were trying to figure out how to adjust for that right like so we literally the whole field is based on kind of misinformation and malicious actors and so the reality is it's a human problem of having malicious actors in AIS and like this language model and generative content is a new tool that allows to kind of do it at scale and so what we really need is actually a new way of kind of structuring society that is more akin to how we structure protocols in level three than how we do it right now where it's a lot of paperwork and you know language rules and kind of uh all the things that AI is really good at kind of manipulating and creating like lots and lots of demo demo demagogue content which you know can go and argue like I think the Goodwill Hunting movie is really good where he's in court and he's like citing some 1800s you know precedent of with a card to get things out out of jail this is what AI will be using right it's like some weird like edge cases in language to pretty much present this and so we need to start adjusting the kind of the framework under which the humanities operating to actually be able to have ai agents and humans working together in a coherent way and then you know we still need like what is a global optimal function right this goes to like if there is so much value created and just AI how are we going to be Distributing it how is this going to be operating like all of these questions kind of build on top of that as well but I think this is where like this Duality needs to be clear that there's a human kind of I call it Society operating system that needs to reboot and then there is a local like hey I'm using a model and I want to know it's correct too for my requirements right like how do we test that so maybe that's a good segue I don't want to rise to that challenge how much could an uh a citizen in this new um Society operating system earn and what would they have to do other than give up their Biometrics in order to to earn it Hazard at some numbers uh we don't have numbers but I think the very simple model is uh like whatever economic output increases by five percent every year and you have some non-linear function uh that that rhymes up I think very like within the next 10 to 20 years we will see at least for some groups of people actually have Ubi play out um so I think it's going to happen I think it's going to be really exciting and I think otherwise other than just the Ubi question I agree very much with Elia um that I think the the one main thing that crypto can't really do is that it can help us to to build societal tools to arrange ourselves with this new technology fantastic any comment here on these questions of alignment liability insurance well one thing that I'd like to add to the discussion I think is just kind of a general view of a lineman from a web 3 perspective and you know the web3 perspective is openness and actually in in the broader AI space we also see a very clear Trend that openness and open source development is creating a lot of innovation a lot of technological advancement and solving and optimizing a lot of the technology around AI And yet when we think about alignment you know most of the big companies that work on alignment think about it from a closed perspective they have a parental view that you know they will protect us from uh from AI in a way and what I think these companies are really saying is that you know we think that you know 50 maybe 100 people can solve this problem uh in a in a proprietary way and I think the web 3 view on that would be that we should open up the technology and we should have thousands of people working this promo or tens of thousands of people working this problem because that is actually the way um that it may be solved effectively and and from from all angles and so um you know this is why I am in an intersection of AI and web3 I think web3 Primitives uh could add meaningfully to that story I think it's a brilliant Point open source always wins yeah openness and transparency is not only the greatest disinfectant but um it's um the way of tapping into the greatest pool of talent that you can possibly have which we're going to need in this context I think that's a super important principle before we wrap up in anything more Jamie yeah I mean maybe just kind of close off on the point I think um you know when we're talking about protocols uh in the context of AI we're talking about and especially with web3 to what extent can we codify legal social protocols into actual technical protocols that can operate independently can be a coordinate coordinated in a distributed manner um and so you know questions around personhood um you know most people don't fully comprehend but there are already precedents in the US like a corporation is actually technically a person it has personhood in the U.S context um and you know there are already contracts offshore for example um where you can have ownerless corporations and so it's going to be interesting to see how a lot of these legal Concepts then get transferred into Ai and how they can be codified by leveraging um decentralized protocols brilliant so like from my perspective um we've always been excited at how web3 can allow us to scale the trust we used to have in the village up to a global scale and that perhaps now in combination with AI we can scale the intelligence that we can deal with the opportunities and challenges of operating that that Global Village in a way that is going to deliver better economics more Harmony lots of utopian Visions like that but thanks so much to the panel um and thanks the audience your attention sorry we didn't get any questions but if uh if you want to come and talk to us unfortunately we're out of time I'm being told here but if you want to come and talk to us afterwards I'm sure there will be a slot thanks so much