Disrupting the Cloud with dePIN - Vukasin Vukoje | web3mine

hey everyone uh so today I'm going to talk about disrupting the cloud with dipin but essentially what it is is basically using crypto incentives to disrupt like infrastructure that U is backing basically any app or any software that we are using today now thein is pretty hot right now but talking to people I've realized that like very few people actually know what dpin is uh but there are like a bunch of reasons why de in is hot uh by the way this reason's been around for probably three years but it takes some time for this to compound and to actually get to a point where these are relevant enough for society to actually s them so everyone knows that a few large organizations on most of the information that we are producing as a society today uh we also know that um most of the infrastructure is basically not equally accessible so if you are uh based in a a country that is not inside like uh some Western influence like super hard for you to actually access any kind of computing and usually you even though you can you would be like ball necked by some quotas so if you want five gpus maybe it's fine but as soon as you start like asking for 100 gpus even though you have the financial means to actually pay for it uh you don't actually get this quotas super easily uh ignoring the fact that like most of this Computing infrastructure is actually in the Western Hemisphere uh usually in the US now also the cost of this Computing infrastructure is super high and the reason for this is not that we as a society were not able to invent more efficient Computing paradigms but it's rather because there is a artificial Monopoly that was built uh by probably five companies at this uh time that are just keeping prices super high in an artificial way and they are keeping margins like more than 50 60% and this is taking into consideration all the cost that they have which which is super high like most of the American companies are not really notorious for being efficient uh but rather like they are like growing fast earning a lot and like spending a lot uh and even with all of that craziness uh they are able to like capture 60% of margins now uh most importantly something that is becoming more and more relevant is the fact that usually these companies are based in a legal uh jurisdiction that is outside of our control so if you're not find with some apps using our data for whatever uh it's not like we have a say because we are not based in these jurisdictions so the only uh governance that would happen in that particular case is in the jurisdiction where these companies are based even though they have presence outside of their J jurisdictions most of the rules are actually like made uh in the main jurisdiction now and something that is super relevant today uh is the fact that like if you have resources and you want to be part of these networks they will not let you in so if you have a big data center in Serbia and you have super affordable electricity either because you have wind farms or because you have solar energy you can't join uh these networks and you can't benefit from the tooling that was built there and you can't benefit uh by the apps that are using this tooling that by the way we built by all of us it was not just these Cloud providers so what's happen happing today so like by the way all of this was actually the case for three four years at least since we started working on this three years ago like this was uh these were the assumptions that we made but what's becoming even more alone today is the fact that like now it's obvious that AI is not actually like a software problem but it's rather a problem on how are you coordinating Computing so the ball neck is on the Computing resource so like how do we as a society bring more resources to a point where we can coordinate to these resources and we can actually enable Society to build larger and larger models uh now it's not just about Computing it's also about how do we actually get enough data to fill this models with relevant data from everywhere what's happening today is that most of the Daya is also coming from a few places uh usually like on the Western Hemisphere so like all these models are really opinionated from like a lens of uh the Western Society uh and we have also physical constraints so like if you want to actually uh build a large enough of a model it's not going to be a few Central entities figuring out how to get to the data but what's happening today is that you would have someone like cop ai go to uh Reddit and a bunch of other folks and they would basically buy like for hundreds of millions of dollars the data that we are producing but at some point even that data is not enough and we are at that point now so it needs to be crowd sources it needs to be like a model that we can all contribute to and not just thousands of people in Silicon Valley now you probably have a GPU at home that is seeing idle and like a bunch of us have there is a bunch of Hardware that is seeing idle like many uh places on Earth uh but the issue is not in the hardware not being there the issue is in the hardware being utilizable so uh there are a bunch of gpus that Gamers have like idling all the time at home but they are not doing anything and the reason for for that is that we are not uh good at coordinating uh the compute and demand Supply at this point so how the question is how do I actually go and do jobs for this big supply of computing I can't because all of this is centralized in these computer providers uh and ultimately we are at the point where like we've been pushing for like uh growing like our infrastructure in a centralized way that we are really hitting some physical constraints so most of the data centers are not capped by the size of U the buildings right now they're basically capped by the energy that they can actually procure there and when you're moving energy around you have like some loss and usually what's going to happen is that you're going to build a large enough of a building that can actually utilize the energy source next to it uh but yeah going back let's just try to understand what dein is so for deepin to be a think you need to have a provider uh also known as the validat someone that actually does the job that the network uh uh actually facilitates then you need to have a token holder that's someone that has like some voting rights uh and is able to to govern the protocol and of course has ownership and has skin in the game for the protocol to succeed and then you have the economy which is basically aligning all these sectors in a way that like this ecosystem works and of course the tooling that facilitates that interaction so for example if you're a computer provider you want to be able to ex these jobs like that needs to be some tooling that facilitates me to interact with this economy of like Global jobs that are out there now so specifically what you need to focus as a DP network is basically the definition of the job so I really need to have like a good sense of what is the job to be done uh the validation of the job so the validation can either be formal validation based on some proofs or it can also be like optimistic validation where oops where you are basically doing the computation again and then you have like some incentive that prevents providers from cheating that they have actually done some computation uh and uh you need the tooling for the developers to interact with this so like if you're uh building a dpin network for AI what you need to do is you need to like integrate into the numis of this world pandas and like a bunch of other tooling that uh they scientists use rather than just building an abstraction and hoping that something is going to happen usually it doesn't uh although incentives uh are strong enough so that all these providers will be providing this capacity longer term so you have time before you build this tooling and this is usually what happened in the more mature uh ecosystems out there uh like filecoin and a few others and and ultimately you need the payment Trails so like someone needs to be able to get paid in order for this ecosystem to work so if you zoom out uh we have the token holder the token holder usually cares about being able to stake their tokens no one wants to get diluted by the inflation that is happening on the network and they want a long a token usually they want a long the token because someone is uh pitching that their network is going to solve a particular Computing problem either in the biospace or whatever some data prep or AI training then you have have the networks the networks need to make sure that the consensus is kept so that everyone is not cheating and that there is some basic infrastructure that people can build on and it needs to validate the results again the results could be validated either by actually getting approv and really like uh efficiently knowing that something was validated by looking at the proof or redoing the execution of the job if you're redoing the execution of the job uh you likely will have like another job that is generated as a challenge job which basically then is able to earn rewards by optimistically challenging the previous jobs that were done and there there are a bunch of details here that I want get into and of course the network needs to generate the rewards usually this is coming from some inflation uh because ultimately like the network in order to function it needs the providers so it needs this compute uh capacity and for for this like some rewards need to be there and then we have the workers which are just gain jobs executing jobs and getting paid for executing these jobs uh the motivation for each of them of course uh the uh the token holder just long uh and get heals and for the computer providers they would want to get like something that is as a plugin earn model where like I have these gpus or I have these CPUs or a bunch of storage and I basically just run one thing and I'm part of this Global Network of resources uh yeah but ultimately the goal is to gain that sense of okay there is this compute demand which we all know that exist there but how do I get uh to interact with this demand how do I actually execute these jobs and I'm able to earn on that uh some of the tring compute networks out there are Jensen Alo fluence lier Runner and filecoin uh and these networks been around for a while so so lier was there like before the previous bull market survived the bare Market somehow render as well filecoin was one of the largest icos and is still the largest DP Network today and fluence recently launched with their vasam executables uh these are super cool uh but the fact is that these are solving particular problems which uh some of them were already sold by the cloud so for example lambdas uh ml training or inference uh video encoding and rendering of course there are like particular cases like Calo where they're doing snark generation but most of them are focused on products that were there for a while and the thing is that there are many more products than what we can see on this screen uh and the fact is that uh we have very few DP networks uh today even though like the hype is super big and the reason for that is that uh building D in networks is super hard you need to worry about bootstrapping a network of providers uh this alone is a multi-year Endeavor so like you need to like imagine how do you get from getting like a few providers that are mining your think to getting tens of thousands of providers that are able to allocate resources for a relatively long period of time for you to get to a point where there is some utility that these providers are building this alone requires you hundreds of millions of dollars of incentives uh and I went through this like working out protocol abs and boost trapping filecoin it took us billions of dollars of incentives because it was first the first Network out there to try to actually do deepin and uh it's hard uh first because you'll probably be wrong so you don't know what the price of your token is going to be before you launch and when you launch all the parameters are basically based on that price you would think that you would have like some economical balance but usually that's not the case so what you get is either like it's too hard to participate in these networks or there rewards that providers are getting are too high which might seem good but ultimately you'll always get to a balance and like if the rewards are too high what's going to happen is that the to token is going to um just drop because you're paying someone like probably 50 times more or 100 times more than what that something is actually worth which is fine in the bootstrapping phase but it's super hard to survive this period especially in an environment that is not bullish and uh if you look at like all the l1s that we have today like probably had tens of thousands of l1s very few have actually gone through like gain to like at least 100 million uh commitment in the network which is also required for the security of the network and of course you need to know what you're doing so like you need to know like token economics you need experts in token economics because it's not enough to launch a thing and hope it's going to work you actually need like to define the parameter super well which is super super hard um cool but generally like there are another set of things that like you as a DP Network need to worry about which is defining the workload so defining the job that needs to be done improving the tooling uh increasing the utility of the network and what I mean by this is sure you have some incentives and people interact somehow between each other but the thing is that like most of the time these DP networks are kind of useless because like you're providing something but no one is using it you're still getting in incentives and you're still participating in these economies but very rarely you able to generate some value uh and this is fine in the bootstrapping phase but this is not okay long term because if that's the case long term we'll never be able to disrupt the cloud because ultimately we are building um infrastructure that is useless of course these crypto incentives are allowing us to do this for a relatively long period of time before the tech matures but if we keep doing that at some point it's just going to be irrelevant uh and yeah when you there is some utility you also want to increase the revenue that your providers can make so at some point you should have clients that are paying your providers directly and not the network subsidizing uh the bootstrapping phase and uh what I think the solution is and um uh is basically a network that is able to run other networks there is so much in common between uh the consensus layer and the research layer between all of these DP networks that it seems that at this point we are basically rebuilding 70% of the stack uh and for that we have built Rama uh which basically is a network that is able to run other networks uh today ramoo has more than 100 petabytes of storage capacity and it has like tens of thousands of CPUs and gpus uh with more than a thousand active providers in the network now this capacity is there to participate in multiple networks one of them is Falon and we have other networks coming along as well but the point of this network is to provide that Universal horizontal layer for R DP networks to build on top and when we do that then you don't not need to worry about like uh bootstrapping uh the first group of providers that would be running your workload you don't need to worry about incentives as much because these providers are not only waiting for your jobs to be executed but trer they are continuously doing other work and executing your job is basically just like a small addition of compute resources for them and most importantly you're able to interrup between the different uh uh DP networks so imagine that you have some data on Falcon and there is a bunch of storage there uh which is super affordable by the way storage on Falcon today is probably at least 10 times more affordable than anywhere else uh so like including the glaciers of this world uh and uh then you want to execute some ETL pipeline uh ETL is basically just transforming Thea in such a way that you can inject that into a a training of a model or like something else and then you have the training doing all of this basically requires you to have some local interaction in the data center so you can't imagine this data flying around the planet because that's how we thought that crypto networks should be built but rather you need to like have some awareness of some locality and that locality is going to be in a data center uh so what what you get by using rmo is you basically don't need to like uh do the bootstrapping phase at all so you're basically directly gang in the phase where you're defining the workload and actually focusing on the validation of the workload so all the crypto uh complexity like building a network and uh doing the consensus and everything else is basically not your problem anymore more uh you need minimum uh uh understanding of token economics of course like you're building a network you need to have some understanding of token economics but you do not need to like worry about like how much it's actually going to cost the provider to run your thing because these providers are going to be there anyways and they are going to figure out whether it makes sense for them to execute your job or not based on the price that you put for for your job which ultimately leads into like 100 of millions of dollars of potential saving for each of these networks because if you're trying to boostrap it by yourself you'll probably just throw a bunch of tokens at providers and these providers might be there tomorrow they might not but what's for certain is that some of these tokens are going to end up on the open market which could kill your economy and you could fail because of that but ultimately uh it's a big risk uh yeah so yeah I just want to wrap up with that uh thank you so much and uh hope we have some questions if that's a thing here cool thank you so much