DAPPCON 2024: Sliding Doors For CryptoXAI Diffusion - Matthew Stephenson

Transcript

thanks everybody for being here um going to be talking about this idea of crypto as the sort of like picks and shovels for AI something you probably heard around in the narrative a little bit and I want to try to try to flesh it out and poke at it and discuss it a little bit uh if you don't know me I'm Matt Stevenson I'm the head of research at panta a little bit about my background um I'm I guess I have research partner here I don't know which is the more sounding title so I'm just trying to impress you uh give you both shots you know just in case um I'm an econ PhD at Columbia mostly micro kind of a kind of a game theory approach I've been in crypto since 2011 is like the cool story but I've actually been in research crypto only since 2017 uh eth research a

company called block science which is at least kind of familiar more generally in the city um you know and and and you might know of it um and then I've done some early writings that you know I think have been reasonably insightful about the way things would go right like so I wrote about blockchain's commitment to devices Back in the Day in 2017 dii rpgfan so like to brag on myself a little bit you know I I I think I generally will get to things early and say semi credible things about them so hopefully we're going to try to do that today uh and and that's my Twitter because I usually forget to give it um so I I'm mainly wri or coming to you today with a sort of extension of this 2023 paper with some some colleagues um with with uh barnabe and and Tomas and

Shen and Jonathan and David day uh that we got to present at the AI conference uh nups last year on Cooperative AI via decentralized commitment devices so the essential the essential angle here is that like there's a future in which they're autonomous agents and they transact on chain and they utilize commitment to solve coordination problems right uh and I think we give a pretty robust reason why that would be so the the AI folks bought it uh at least the ones we talk to or talk to us um and the tldr from a from a kind of game theory way is basically if you train AIS in this computationally efficient way which is called you know external regret minimization uh you wind up with an equilibrium that's like not quite non-cooperative it's not robust to strategic deviations is the idea what

does that look like it looks like me and so that connection is is pretty robust um and the idea that the intuition that it goes to meev is about right uh we think and so that that's the tldr of that paper what I'd like to do this talk is extend it and generalize it a little bit so in particular um you know that one is about a particular type of like multi-agent reinforcement learning um we did that as an example case but I want to talk about it in a more kind of realistic case right like what we basically have now is we have these two types of models we'll talk about this in a minute right we have Transformers and we have diffusion models which just kind of appeared on the scene and uh and those are those are interesting but they they weren't specifically addressed in

our paper um and then I'd like to sort of explore the way those markets might develop um and how those might map to different possible Futures so we're going to have to do a little tiny bit of something that feels a little like futurism which I'm pretty uncomfortable with like as an academic you know it's like the line is like it's always hard enough to predict the past or whatever um but I'm I'm going to do my best here okay and then uh and then lastly we're going to talk about this me concept a little more in detail as well right so we're going to talk about this specifically as being demand in compatibility from a multi-agent future so why am uncomfortable with futurism there is a tradition of VC futurism right where you basically take the set of things you invest in recently and

then you try to map them on to the union of the set of like the podcast you heard recently and then like you know the two big sci-fi novels you read or whatever right and so like that's usually the way you kind of do futurism here um GNA going to not do that here but uh but I'm mindful of that tradition so so so I'm going to try to not not straight from it instead let's try to like look at what we have and be a little theoretically grounded so I said right we have two models they just sort of appeared on the scene we have a Transformer model and we have a diffusion model or an architecture however you want say it right um to my not to my eyes they didn't appear exceptionally different or better than any other architecture that was proposed except they happen to work

don't take it from me the guy who taught me ml uh a little bit David blly he said the same thing he's like if I show you it on a whiteboard there's nothing about it that's going to be especially interesting to you uh right compared to any other model it just so happens that these things work in a really interesting way they show high returns to data they show high returns to compute and they seem to do these just useful things that seems obvious so that's Baseline but that gives us three sources of uncertainty and to be like good good futurists here let's try to walk through them with a little bit of logic right we have uncertainty about like I don't know what if a new architecture drops from the sky that's one possibility right like just happened to us last year or actually several

years ago right but uh we understood it last year and then there's more compute and more data right which is the sort of like you hang around open AI or whatever that's that's practically The Motto right more compute more data more compute more data they they show great returns to both compute and data so we have uncertainties around how those uh those production functions essentially wind up playing out so I'm going to when we talk about these Futures I'm going to give you like two broad bucket cases right uh one is just this kind of like AGI ASI like what does money even mean man sort of future and if we go there I'm not going to try to talk about it because like I don't think you can predict anything in that future I'm not going to try to predict what crypto AI is going to be right so whenever we go

there I don't know I'll see you guys in the paperclip maximizer or in our great Utopia or whatever it is uh but then otherwise I'm just basically going to say probably we wind up with crypto as picks and shovels now I keep promising that I'm going to try to deliver on it it just takes some stage setting right so the first path is new architectures I have no idea what that's going to be that just goes right to AI Utopia what does that money even mean man I I don't I don't know right can't really predict that so we're going to kind of skip that one and just say well it seems like we probably only got two that kind of do what we think is especially useful so I don't know what our probability or expectation is on this but I'm I'm skipping it but now let's talk about

return returns to compute right so let's say returns to compute are like stepwise crazy great and we have some new you know some new crazy thing that's in front of us once we have the new Nvidia chip or whatever well that's probably again you know our shogi Utopia whatever it is but in the more realistic case where just the returns maybe let's say continue to be high or even diminish I think you start to get some Market segmentation basically you get you know what we call in the in the industry here monopolistic competition right where some people are maybe competing over the high end right you're constantly buying the new chip you're you're you're trying to compete on The Cutting Edge of computation but for consumers and other cases they're happy with a little less right so the idea is this is called like satisficing you can think about it like

TVs or something right yes there's always newer and better and more highres TVs but the pace at which people consume them right they start to get kind of happy with what they have it's like things can only be so HD even if you could make the you know fewer and fewer pixels right um so in that case I think you get diffusion and again I'm sort of like pushing this diffusion to the crypto pick and shovels argument second we have returns to data right if it's the case that returns to data are really high um right then again we're in AI Utopia but if it follows some reasonable functional form right then we probably start to get Market segmentation too some companies we can which can capture data and really really improve their models maybe may win that becomes maybe a little unique and

different because it's kind of oligopolistic maybe it turns out that the prompts you you input are really important inputs to the next model which gives you this kind of like positive feedback loop that's that probably tends toward Monopoly or lopoly on some level on the model level but I think it's again we can probably imagine that there's still this satisficing element where some people are happy with the the cheaper or more differentiated or more specialized product in which case we at least have some kicking over to this consumer diffusion uh pick and shovels argument this is the whole map just I don't know proof of work here um and so I said now I've got to justify why blockchain what I really just gave you is why consumer diffusion right why a blockchain so I think if you hang around

crypto Twitter you'll kind of get some flavor of like crypto is like 247 always on you know Super Battle tested Financial rails it's programmable money I I'll disagree with that by the way and like I think that's that's not essentially an incorrect take right uh but it's kind it's kind of the 101 and and it's it's a little handwavy we might say right I think the programmable money part winds up being the especially interesting part of that in in my view because it just it isn't clear to me why you couldn't do banking 24/7 always on right like it doesn't seem absolutely impossible that you could never out compete right on that front it just happens to be the way banking is done today let's say so the more advanced one from my esteemed co-author jaava um he gave a talk last month he did a nice

sort of like here's why blockchains are an ideal substrate I'd encourage you to walk his watch his talk at uh Stanford last month at at Erica's uh open source AI Summit um but I'm going to do a sort of specialized version here I'm going to focus on a particular ular sort of area of commitments and demand right and the reason I do it is is I I uh my my kind of MOT is as you know as a researcher in the space is if you want to solve a problem you try to identify the small toy game that the problem essentially is and then you solve it based on that game so I was trained as a as a game theorist so that's essentially what I do um and what is the toy game where what what is the model I we're basically trying to

look at an agent based future right like specific specifically in the paper we're looking at this particular use case of you have an AI agent that's maybe let's say going to going to do your bding and so what is the what is the game what is the toy game remodeling well that's what we're going to try to do like what does it look like if you and I have two different competing agents right and what happens if we're we have some objective and they compete on that objective well I'm going to argue that uh the the demand version of it kind of looks like this it's going to be a coordination game you know that's pretty typical right so you have a bottle of wine you have two patrons with uh with incompatible demands right each one let's say wants the whole bottle right and then let's say you have two agents

tasked to I don't know go out to the store and get you one or send a message for it to be delivered or whatever it is right um we should talk a little bit about what it means for it to for the demands to be incompatible you each want the bottle of wine but let's say you would actually like you'd strictly prefer right half a bottle to nothing or whatever right you you'd like some wine at the very least and so what this is going to do is it's going to create create this kind of like Continuum of equilibria where what we' just like is the bottle of wine and so the the problem I'm show you the model in a second is what happens when our demands are incompatible when we both send the signal that we both want the whole bottle of wine right like those are just

incompatible demands the world can't meet those and that's just to just to cheat and and give you aead you know give you a little info ahead that's the kind of meish side if that sounds like me no surprise right this is the real world me piece so what is the game theory sort of toy model version of this look like um this is a this is this is a really discreet simplified version of the coordination game you know you could actually just sort of draw it as a line right with with multiple equilibria but it looks like this we're just going to say you have three things you'd like to get out of the wine right like you could do a low a low demand a medium demand and a high demand you say high is like I want most of the bottle mediums I want

half lows I want 1/3 right and then it just so happens when we both send that signal if they're compatible we each get our demand and if they're incompatible we're going to say we both get zero you can say whatever that is right like maybe we pay a penalty maybe we just have to we're frustrated by no wine because couldn't be our interest couldn't be served maybe we just have to send something again right depends on how we want to think about the the the sort of failure mode of this but we're just going to call that zero and the worst outcome and so if you just follow the logic right it's obvious if we both request half a bottle of wine we can do that right if I request high and you request low we can do that uh and so on right but you can also see that if we

both request low we can do that too we just leave a little bit on the table and so what I bolded here is like the parto Frontier right the idea that we're still we're we're each doing as well as we possibly could and getting as much wine as we possibly could and that's where we that's where we'd like to live so this part is is is almost silly because it's so straightforward but I think this is essentially the right way to talk about it which is if you allow one of these two parties in the wine to have a commitment device right which is basically basically to um to to commit upfront to their demand then you wind up with the equilibrium that they can sort of enforce right so uh they get to enforce the the equilibrium they like

right if I don't get this then I penalize myself in some way thus you know I'm really going to try to do it or if I don't get this then I'm going to penalize you in some way right like anything that that looks like a commitment a conditional commitment right such that if this doesn't go through then X bad thing will happen that changes these particular payoffs means that I get to enforce the equilibrium of my choosing right I always get High um I always get high probably shouldn't say it that way I always get the high payoff um so right and and of course this is crude and and and discreet right in actuality you'd get everything but Epsilon right because you just have to offer somebody else the minimum so that they agree so if this is the case then you just want to deviate

to Commitment right of course I'd love to have a commitment device and this is sometimes you'll read this as like commitment races or something like that right um well if that's the case if we're going to both if somebody's going to be better off if they go for a commitment and they're going to get everything but Epsilon you know I might like to have a commitment structure too so the idea is just I would like to go to a decentralized commitment device which is essentially a blockchain right you get to program these commitments um you get to program however you want to model the commitment game theoretically right whether it's whether it's self slashing whether it's it's other slashing whether it's something like that right you can you can credibly commit and you can send the message to somebody else and you could enforce this

equilibrium which means somebody else wants to do it too and when you oh I skipped the slide or I don't I have the slide hidden here but if you it turns out you can trust me on this that if you smooth the function slightly and you give both people access to Commitment they wind up pretty close to the 50/50 that's where it usually winds up going uh basically has to do with how bad it is when you get really really far away right when you really you know when you're um when you're a special sort of uncertainty that Happ happens when you're close to demanding at all right um so here's what we've sort of covered as the futurism part right we have one this idea that aite Tech is going to

inevitably diffuse decentralize and so on right uh we've decided or I've contended that decentralized AI agents are useful I didn't spend a lot of time defending that one hopefully it's in the air enough that that seems to be true you know open AI is doing a lot in this I can sort of feel to me that it would be very useful um so so I sort of took that as given and then we have this game theory part which basically says if you're if you're um if there's some competing resource that you're both after then you're better off deviating to Commitment right um and then whatever it becomes whether it becomes like a stackelberg game or whether you want to think of it as like a bargaining game or whatever you can wind up with different outcomes but essentially the model is going to be around the same right and so

that would have agents transacting on decentral decentralized commitment devices because it's just incentive compatible they both want to do it I don't think I've proven that I've just sort of like gestured toward it in what I think is a pretty plausible way and then again the specific outcome of how these things work out depends on your model and I honestly don't think there's an answer is my genuine opinion but uh but I think the answer would live on a blockchain um at any rate that is going to be my talk I'm going to finish a little bit early just so I can save room for questions if there are any if anybody wants to just you know bail and go to the party I understand that too I'm happy to talk outside so thank you very much you