DAPPCON 2024: AI-Commerce: The Future Of AI Payments - Don Gossen

Transcript

first off thank you to nosis and the davcon organizers for inviting me to speak today I'm been asked to regail you about the topic of AI Commerce um those of you that are here are probably asking what the hell is that so uh from our point of view it's the future of AI payment uh my name is Don gwon I'm the co-founder and CEO of never mind we're building the decentralized payments protocol bit of background on myself I've been in the web 3 space for quite a while uh co-founded a project called ocean protocol um I've got a company called Kiko that helped build networks

like cell uh filecoin um bunch of other things and more recently have been focused on never mind um which is operating at the cross-section of artificial intelligence and web 3 and from our perspective um we see the world emerging in in a very uh particular way there's a ongoing debate at the moment whether or not um AI models and agents in particular are going to proliferate or whether they'll um consolidate uh under the control of one or a handful of organizations at never mind we believe

the former that we're going to see a proliferation of AI agents um following kind of the mistol mixol style um and that these agents will manifest in their trillions and change Commerce forever um and so we you want to support that Evolution um what's quite interesting is uh the way that this is evolving it's very reminiscent of the early days of the internet sort of early to mid 90s um what you had back then were a bunch of hobbyists that were adopting online tech building you know bolt and Boards

creating lists of websites that were interesting and what each one did um and all of that is kind of it's playing out again right um we're seeing these lists get created of what's a interesting AI um model or agent what it does um and you know it if we hearken back to that um to that sort of historical precedent um within that confine of these hobbyist Builders you had uh a subset that tried to commercialize what was going on at that point and what was required was a shift um that leveraged

the new technology that enabled uh e-commerce to actually take flight and begin to scale so Solutions like PayPal emerged that facilitated uh this Evolution from traditional non-digital Commerce into the digital realm as e-commerce um and we see the evolution of this into the AI side of things um manifesting as AI Commerce and we want to be the as never mind the PayPal of AI Commerce so you might be asking yourselves well why does you know why does AI need uh its own payment solution um the reality is AI agents and payments

for those agents and and how they transact is non-trivial um and we're going to we'll get into this a bit more but when you look at how an agent operates and start to decompose its individual pieces what you start to recognize is the dynamic nature of these agents and the variability of um the services that can be invoked in order to service those the the um Dynamic requests that are made um and these variable services that can be invoked have corresponding variable costs that need to be translated into a cost for the user um this is a non-trivial action and what we see after you know discussing and

interviewing um and over a hundred organizations and Builders is that those that want to start to uh monetize or commercialize the agents that they're building are having to effectively build their own bespoke payment system for their agent um and as a result these payments systems aren't interoperable um they're not modular by Design and therefore they're going to have a lot of difficulty in scaling just to look at what an agent is and kind of how we can think about the anatomy of an agent um it goes a little bit beyond you know maybe the pure GPT

component where you're interfacing directly with a a large language model or a diffusion model itself um within an agent there is a backend um that provides for a bunch of orchestration and rooting that takes place so a user will submit a prompt and that prompt will go to an agent and that agent may have its own discrete data sets um or embeddings um those embeddings might be retrievable from an index or a vector DB um and the individual prompt will get decomposed and there will be a determination by this back end of how sophisticated a response uh is

required um and root those requests to potentially different models that that agent uses um depending on the the type of response that's needed Andor the level of sophistication of that response um so then the question becomes okay how do I account for that variable nature I'm you know I can submit a a request or a prompt uh and at the same time another user can submit a entirely different request to that same agent my request might be very simple in nature while the other requesters um requests might be complicated with respect to the

technology that needs to be invoked it might be a multimodal style request that's anticipating text and images as a response um um so for instance I submit a request it's simple in nature and that gets rooted to let's say GPT 3.5 versus you know another request coming in that is multimodal in nature and requires gp4 just for sake of example now the cost on servicing those requests can be an order of magnitude and difference um and so accounting for that uh variable difference and then Translating that into to a settlement component um to the user is uh a not

uncomplicated proposition um and so what we're doing is helping to simplify that with a a system of unit accounting um as well as a system for enabling the payment and Redemption of of requests we do this um through a credits system um this is the prevailing type of solution that exists for these these uh particular style of of solutions at the moment um open AI is using a rendition of this if you've used uh chat GPT and you've paid the 20 bucks a month for credits um you're getting an allotment of tokens within a given range of of time if you go over that allotment

then you're throttled um we're doing the same sort of thing but our backend is using crypto rails for this credits issuance uh uh Redemption and actually um system of unit accounting that that uh works with the metering so to break it all down a a builder can issue a set of credits that are going to enable um the usage by a user of the the agent that the Builder developed um never mind enables that issuance and the the purchase by those users then when they make a request to that particular agent that those credits apply to we

work with the back end of that agent system and the metering uh of that solution to translate that uh usage cost into the Redemption of those credits um and so what you get is a relatively straightforward solution that accounts for the usage properly but also supports the uh Dynamic nature of the inputs and the variable responses that these agents can provide if we sort of conceptualize where this all fits within the greater AI stack um we see the payments piece being um useful throughout the entire

stack but our particular focus is at the agent level um in this diagram if that at the top um so working with you know the likes of uh the aonis team Ola on Olas uh polywrap um super agent Etc uh but the solution can be used at the the agent tooling um level which would be you know the the um agent Frameworks that are emerging um as well as at the AI tooling level which is where a lot of the discussion is going on right now around model development INF pring data storage weight storage etc etc I think the next thing to answer is

why would you want to build a decentralized AI payments protocol um and again we go back to this concept of and and belief that we're witnessing the rise of these trillions of agents that are going to change Commerce forever um there's obviously I think the very rudiment um somewhat flippant response to this of like AIS aren't going to have bank accounts what are they going to have they're going to have wallets on chain um you know that's I think applicable but doesn't tell the entire story one of the main drivers for decentralizing the payments piece we

think it's absolutely necessary in order for these AI agents to uh manifest and to scale and proliferate especially in a semi-autonomous and and more importantly an autonomous um uh way the issue is the payments choke point and what do I mean by that it's the fact that if the payments for these agents goes through a centralized singular Authority like an open AI A Google a Microsoft a Facebook what have you the ability for those organizations to

deplatform an agent is is massive right they hold they call the shots if Microsoft disagrees with an agent or thinks it's competing with a particular agent it can shut that agent's ability to transact off with the flick of a switch and the reality is it doesn't matter if the rest of that AI agent stack is decentralized if its ability to transact has been shut off it might as well not exist um so we think it's Paramount to address the payments choke point by decentralizing the payments rails for these agents to leverage um and so that's the goal to be

the engine behind AI Commerce uh become the PayPal of AI and enable this interoperability and interactivity um between these emerging agents as they manifest Where Do We Go From Here I think what needs to be done now um and we're seeing it start to manifest an increased focus on agent development um both in with respect to the Silicon based AI agents that are emerging but also how carbon based agents are going to interact I think the you know the um the the jury still out on what the interface is going to be for

us in in terms of of leveraging and coexisting with these agents um it's likely not going to be a text based one um so you know working on building out these agents and then integrating with them um so we're doing our part where the payments solution for Olas we're working with po wrap flock Napa embed Roo um we've actually crossed the aisle into the web 2 I space partnering with super agent um so yeah I think the focus just needs to be on a proliferation of of Agents um and then manifesting the benefit from that right uh if you think

about productive semi-autonomous autonomous AI agents out there uh creating value the reality is if our um anticipation and hypothesis is correct and we are going to see trillions of these agents emerge then the reality is likely that the GDP of humans pales in comparison to that of these agents um and therefore the the scale and economic impact alone is vast um what does the future hold again I think there is an argument here for the decentralization of the payments piece um if you look at you know what de

centralization means and how tokens are actually a proxy uh often times for for other assets you know rwas data sets weights of models models themselves Etc um I think there is a likelihood in the future that due to the nature of an agent its ability to self-replicate and establish effectively an unlimited number of connections that that system of proxying through tokens might disappear as well well and these agents just get to a stage where they're effectively trading on their service or their asset directly in what amounts to a bartering system um so I think it's a good uh place to start

in terms of enabling that potential future outcome there's other more crypto native things that I think we're going to see we're already starting to see uh discussion around initial model offerings um you know I think we'll also likely see hint hint um artificial intelligence credits or coin offerings in the future uh really taking advantage of what web 3 is good at and that is capital accumulation and um allocation and putting that to work in the creation of developing out these agents so with

that I will open the floor to questions um any questions if not uh yeah I I love the formulation of the IMO uh or I mean it's hilarious to think about like just aping into an an agent that I'm really excited about um but yeah my question is probably one that you get all the time and it's it's one that the crypto crowd hates the one of compliance like so let's say I spin up this agent that I I I want the agent to go on to the the Deep Web or dark web I wanted to go on

to Silk Road and I wanted to get the best possible drugs possible for me and um you know where can I access this agent I I want to pay for that agent first off so I mean uh obviously you don't have the ability to stop that agent right because it's decentralized but in terms of compliance what what happens for for NE never mind I mean so I've got two answers to this the first is just like the generic blanket web 3 response that's been toted around for you know the last 15 years we're enabling the decentralization of the payments but if illegal activity is

taking place it's still a legal activity right um you know so and it's been I think relatively accurately proven that like um you're you're not doing this autonomously or anonymously right so there's ways to resolve like maybe who built that AI for example I think in with respect to the broader like Financial Outlook um as it relates to existing um Financial systems and the regulation that applies there I this is a bit flippant but I look at who's or what is actually

transacting and if it's carbon based agents not if it's sorry if it's silicon based agents that are transacting with other silicon based agents and this doesn't affect you and I and therefore the regulatory system that has been established to protect you and I in in part because we can't establish our dun Bar's number is too low we don't know everything that's going on we can't establish that number of connections but an AI can regardless we're somewhat removed from that economic environment and the corresponding activity that takes place so you can make the argument that once this really starts to manifest in scale

that these are agents that are independent of that regulatory regime um whether or not that's an accurate interpretation I think you know we'll definitely see some perverted views on this but it I think that's sort of the starting point right um a dog isn't a human so therefore regulation doesn't apply to a dog is it the same for this autonomous thing over here yeah I a question of regulatory capture and also uh uh yeah like you have to prove intent in the first place in order to charge someone with something so you you'd have to be able to yeah charge I'm

also of the opinion that once this ecosystem emerges that it's going to operate on a plane that we're just largely ambivalent to right we we know it's there but it's it doesn't impact us and therefore like our regulatory constraints will be non applicable cuz we won't even know that they are applicable right um they the agents themselves will create their own regulatory regime all right great we're done