EthCC 2: Gilles Fedak

so hi everyone my name is Jeff addict and I'm co-founder and CEO of exec and today the talk today is going to be about this insurance marketplace and the way we are doing that at I say your dignity centralized marketplace we are facing several issues and that's many challenges and I think that's what's important to understand actually you know what are the challenges of building this centralized marketplace on top of ethereal and I think that many of the solutions we are having could be applied to to order this earth this centralized marketplace so the agenda today is the following first I'm going to briefly introduce the centralized marketplace I would take that I exact example as an example at the eject what we're doing it's a marketplace for trading computing resources we will see you know what are the new usages and what are the new business use case that I load by the fact that the marketplace is decentralized we will also look at the implementation challenges there are many and some of our solutions and finally I will conclude and give some our perspectives so decentralize marketplace what's that there are certain number four characteristics first the first assumption is that you have transactions that are peer-to-peer without any central authority that take a fizzer maturity and so all the all the steps of verification for buyers and supply and suppliers are handled by the blockchain and smart contract which also you know automate the verification of the contribution and and proceed with payments so at the moment you have you look at you in the etherium ecosystems many kind of centralized marketplace that can be for photo exchanges to trade digital asset and cryptocurrencies it can be generic marketplace such as a phone sensor open Bazaar it could be also marketplace for trading not fungible tokens such as opportunity offers for trading computing resources such as storage or high exec so this is really hot topic Africa at the moment in particular with a decentralized exchange places one important feature is that this centralized marketplace are the apps so you expect the same features that exist for the apps for your decentralized marketplace okay so typically the fact that is such a chip resistant something really important just some hold the decentralized marketplace should not prevent someone to enter the marketplace that's something important should be autonomous transparent and stoppable secure resilient Exeter etc so there was a grid paper like I encourage you to read publish at the communications of ACM that you know compare and present the two models so the decentralized one and about traditional you know e marketplace platforms such as eBay you go Airbnb on and whatever I just want to focus on two important aspects that you know that I will explain more in the rest of the tour the first one is just so typically if you go in a centralized marketplace last Chas you know you BIR you basically trust the platform and then the platform has several mechanisms that evaluate the buyers and the sellers search you know as such that as a consumer you will be happy with the service provided by the different stakeholders of the marketplace of course in a decentralized way this is different you don't have this fault party so there should be some aggregate there should be some mechanism that you know automatically automatically and and transparently evaluate everyone and I will show you some example and I exact the way we are doing that for trading computing resources and so another important feature is privacy and security with centralized marketplace you basically give all your information to Airbnb let's say and and you don't have any power about what you want to disclose on that to disclose thanks to in the decentralized marketplace thanks to all those new technology that call button today we had several thoughts about identity zero knowledge proof etc you have a much more much more flexibility about the information you want to share and the information you don't want to share so find Stan you can have a decent sized marketplace where the terms and conditions are public for you know using a specific resource but still the price for instance for a given deal is kept private and same for identity you can have marketplace that you know forces you or you can you can add the marketplace to have a clear identity of the one you are interacting with like KY chief concerns for the exchanges or or you can have setups that are fully and anonymous until you have much more flexibility in this and same from privacy the thing that is really important to understand is that for this insurance marketplace the key word it really is just less the assumption is that there is no trust between the different stakeholders and so there must be some mechanism to overcome this limitation I will tell you I will give you some examples of that link so at exact we are doing blockchain based market approximate satellite computing so this is a marketplace for trading computing resources so by computing resources we are talking about servers so that machines we are talking about data sets and applications and we are using ethereal you know to hold all the payment the resource provisioning accountability and so forth and here of course providers interact in a peer-to-peer way and that's no central authority so why do we think this is important first because there are many many doubts that are coming and then in more computing resources than just you know what the blockchain can provide the new storage that the computing power they need to run any kind of application and so they need a DD and we think we can provide a dedicated in factor for doing that and in the future we think that by creating a marketplace for computing resources we can eventually have you know a new form of cloud that would be probably cheaper and less and greener that consume less energy that is more efficient in particular with you know all the new scenario that are coming with edge computing for computing 5 G's etc here we need to have you know computing resources that are closer to the data extracted right so this require changing during the paradigm the way we are doing distributed computing infrastructure so in the market place in the distant relative plate you need to have a talker and this token is important for several reasons the first one is that this is the token that allows to build incentives in the network you do you cannot have trust I mean cannot have a trust a football a truthful party but on the other hand what you have thanks to the tokens the ability to build a great that manipulate the token and so that create incentives to behave correctly in the marketplace and this is a very strong tool actually that's something really important so to understand very it is a big picture of why exactly is working so you have users that are going to use the apps those dubs need extra computing power or extra data set they go on a market place in this market place they can find and provision the computing resources they want at the price that corresponds to to the requirements once they select the the computing resources the worker pulls for instance what couples are set of machines that are ready to execute task the task is then sent to the worker pools and all those machines are going to execute the task once the tasks are executed there are results the results now are going to be evaluated for an algorithm that we call proof of contribution or poco and the goal of this algorithm is say basically ok you asked for the execution of a task with that level of trust now the poco will look at the at the execution and say okay now I decide that this result has been executed with the level of trust you asked for and if it's ok the protocol the the poco will validate the result of the task and payments will happen and the results would that will then be you know sent back to the - to the user so when defining a marketplace you must first think of what's going to be the transaction between the buyer and the seller zone okay that's very easy the transaction is basically the execution of a task so now you have to decide what's going to be you know what a deed is going to be so the deal it's going to read you first wheel so you in our case we have requesters so requester are the one who asked for the execution of a task so they every turn order another it's actually a structure that say the price what I want to do the definition of my task and then you have three different providers there is one provider that provides the application so similarly the provider emits an order saying okay this is my application this is the price to execute my application these are the characteristic of my application for data set this is similar so you have the ability to deploy some data and expose those data and give a price for using the data within a signal computation and for worker pools this very similar also so you have machines and worker pools can publish order saying okay I'm okay to execute the task I'm okay to execute tasks at that given price so a deal is done when all those orders match and when a deal is is sealed we register that deal in the blockchain and once the deal is registered at the blockchain then we can execute the task there is the Potomac annealed that validates that the deal has been correctly fulfilled and then payment happens so here one very important thing is that when you define this transaction in your marketplace that basically going to give you the pricing model so here the direct consequence is that the pricing model is what we call paper tasks you paper task execution so so that's something important because in traditional cloud computing you don't pay for task execution you provision a computing resources you use it for a certain amount of time and and then you pay for the usage but you pay for the time you use the machine here you pay for the execution of the task with a quality that the task has been executive with the level of Church you ask for so I talked about proof of contribution this is something really important it's a treacherous environment so you may execute a task on a machine you don't know I do not necessarily trust so you need a mekinese that will basically say if I want a very high level of trust for the execution of my that meaning that basically I would my wizard to be the correct one there should be a mechanism that that took at that and I provides you quality that it's what you asked for and what you paid for so the way we are doing that that I exact so it's a framework first the level of trust can be different we can have a level of trust of zero meaning that basically send me the result you got and I don't care we'll manage or you can have a very high level of chat and if you exactly to me your application and machine your own church this is this algorri so it is taking reputation and result certification so result certification the idea is that simply to duplicate your execution of a task and then you compare the results so when comparing the results we do a majority vote in and if some when you know see if some workers went against the the console's then they are punished and we take the stake so with this mechanism with this mechanism there is no incentive in behaving incorrectly basically you can you can always if you are worker you can fake a result but there is no chance that you earn money idea slightly actually very often there is I think there is a confusion between decentralization and peer-to-peer network so peer-to-peer network this is the idea that every node in the network play the same role so if you remember back in the years peer-to-peer it was you know opposed to a client-server and we are saying that a node is both a client and a server so this authorization is very different at I exact we have many different words that identified and some of the component some of the service are actually centralized in the you know the sense of client-server so for instance we have a product that we call the dab store so the dab store everyone in the I exact network can't deploy an application we cannot it's also trip resistant so we cannot do any about that but we maintain a centralized service that lists the application that we like so basically people do a simple github request and if you like the application we can leave this application and similarly fun stones will have the worker pool manager so that's actually at the moment that's very easy to have your own worker you just start a simple agent we connect to a worker pool manager so there are centralized services but anyone can run those services and that's where the centralization happens so you have different walls but no one can prevent you for having your app store or your worker pool manager and that's something really important that's something really important because first we expect people to do that we expect people to build verticals because they have a special knowledge in doing artificial intelligence much better than us and they will be able to create some we articles with data set providers are going providers do especially soft and soft floor cafe OpenCV whatever and and and find the correct computing resources that are efficient cheap to execute that fortunes in the coming weeks we are going to launch a public worker who so what we call public worker pool is is a worker pool where machines from the India internet I mean anyone can connect and offer its laptop to altitude to computation so we will do one but we expect people to create some other public worker pool exactly like you know binding files you have mining farms and you have public money pools like dark pools so the way a decentralized marketplace is being designed I think it's probably closer to what we call a market network so let me try to explain you so typically you have buyers and sellers who in ok these are the dubs and user of the Dubs the sellers are the different provider so typically these two layer layers added at the top of the figure you can find that in a centralized marketplace what's different is the true the toward earlier that our you know what were called core services on core protocol so core services these are the one that we developed at I exert and we manage but that anyone can replicate okay so I'd one is able to create its own broker etc what's really important is why I put that market network is that you can imagine to have many other external services that connect this network and that fulfills special special services such as for instance something that is really important is kuratas the fact that you can select the applications give a recommendation for the application you can imagine for instance autism you're like in an exchange you have what we call OTC so imagine you are you are Pixar you're going to Disney you're going to render your next movie for that you need five hundred father nerves or 50,000 CPU hours of course we not go in the public market way to go on OTC and someone will you know use this phone call to to to to to to reserve those machines at at a bad price and then the core protocols typically so at the moment we are developing those core protocols and score smart contract but typically this should be handled by the community in the future we we we are expecting to have some governance rules so that collectively we decide about how we make the you know those smart contracts how we we can modify for instance the way we reward people the way we punish people and and the different mechanism that are part of the proof of contribution so these typically should be part of human consciousness among the community of people using Excel so typically this is something you cannot do in a centralized marketplace if you go in find Sansa Huber drivers in Newberg they don't have the right or they have nothing to say about how much they will they will make when they do when they do when they drive someone the fact that we are doing a decentralize marketplace also create news ages for instance the way we trade computing power in Isaac is very different than the way it's done traditionally in cloud computing usually cloud computing you go on a specific cloud provider let's say Amazon Web service and then you have a list of instances each instance has its own price and usually the catalog is quite huge plus the price varies depending on the region so that's actually very difficult to take a decision about where to execute if your task is highly demanding the way we do the way we treat computing power in i exac which were computing powers a commodity so think more you know think it as like the oil market you don't care about you know when you fuel your car you don't care if the oil come from Venezuela Iran or Texas you just fill your car and go here's the same idea you don't exactly know where which version is going to execute but what you are interesting in its finding the machine that will execute your task at the best price so this is the interface for the marketplace and you see that it really looks like trading you are so you have the price to execute a task that varies during time you have another book with bid and ask so people say I'm ready to execute tasks at that price and some others say ok I am I want to be executed I won't be my task to be executed at this price so you had a restaurant reject a dragster so so typically this kind this is I mean this is new it really doesn't take this for traditional cloud computing and this exists because the bucket president arrived it's because anyone can join and say okay I'm you know I'm I'm ready to do that to to exactly that the sprite that you can you can change the paradigm of using the cloud another really important thing is privacy and ant wrestlers the fact the so because you are going to execute in a remote machine there could be the that something that there is a a new challenge in keeping your data private and confidential so to solve this issue we are working using a technology that is called unclaimed or trusted execution environment and especially one provided by Intel which is called Intel SGX so what Intel SJ does is that it's basically a feature in your market microprocessor that encrypts a part of of your memory and when you execute when you execute a process within this Enclave the memories always encrypted and so this has very strong consequence the first one is that it's impossible for someone to tamper the execution so if you deploy your application there is no way for someone to generate a fake result for instance and the second important feature is that the data will be what we call an to end encrypted so from you know the input to the output there is no the the data is never decrypted even during the execution so thanks to that we can also create a new user such as for instance the data wallet so at the moment there are direct rly marketplaces for data but they only sell their data they don't rent the data so using a decentralized marketplace with you know like exact is doing plus the technology such as SGX we can now run the data so run the data this is a use case so that you can understand so here the use case is a somatic image segmentation so the input is one image and the output is the image with a segment Iike some of the shots which a sponsor I recognize that this is a body this is a tree this is a road the grass you know this is used for self-driving cars or external this is very useful so typically to to to to to execute this kind of application there is one element which is highly difficult to obtain which is the trained neural network to train a neural network this extremely costly it consumes a lot of CPU times and creating this data set is quite difficult usually so when someone has trained a neural network certainly he wants to or she wants to to to monetize that and so typically I think you can do that so you can have the data set that is going to be monetized which is the training on network plus an application provided by that provider plus summer GPUs to execute the D planning application and so using this all of the different providers are going to take a reward and watching what is interesting is that using for instance so if it's wow it as Jake does not work with GPUs but if it's a deep learning framework or machine learning forever that worked with CPU you can even imagine having your train network being used on some remote machine with the quality that nobody with the golfy that the worker machine will not be able to copy this trade neural network so you can no rent data and that's something that is really really important while keeping full ownership so when implementing at the centralized marketplace there are many challenges in terms of implementation in particular with respect to scalability so the first version of AI exact we were implementing everything the order book management deals etc using you know under a filmer blockchain so we now have a new approach for that that is called open decentralized brokering and that address the scalability program of the marketplace so here the idea is that all the deals are emitted of Shane and there is a special agent that we call broker and the broker can match the different deals and record on the blockchain the fact that deals have been sealed so that's great because it's more efficient it's less costly for people and plus it allows for you know other book management of any kind of complexity the other issue is that running the proof of contribution protocol and the mention is also extremely costly because for every task execution you basically have to interact with the blockchain saying this is me you know this is the proof of my contribution these are some harsh that describe the result etc etc so of course this is extra because so the idea for that is to create sidechain what we call a domain-specific sidechain so domain specific mean that this is a session that is meant to be used only for Isaac and that will that we can optimize according to our own requirements so typically we want high throughput low latency because you know it's ID and computing so this morning we had talked from Francois what we are about that and I won't elaborate that that much of this part so what you see here is that at the end we end up with a solution which is quite complex we have the main head we have a side chain plus we have things that happen of that happen of shine so that was quite difficult to manage this is what we call the sort of a hybrid brush and design so typically the idea is that the maynot is going to store values and governance and that there is a section where there's no value so typically you don't need proof of work of these kind of things that is optimized for running the decentralize marketplace and so the question now is in terms of a governor so for instance you want to you want to answer a very very bad basic question such as who is able to run a node and the side chain with a lottery two together none of the side chain how do we want to upgrade the consciousness and so typically this really this really raises some government issue and you want that to be solved by all the stakeholders of the marketplace okay so as a conclusion maybe I can take your bet and say that this centralized marketplace can be the next future killer app is either cryptocurrency for public blockchain eclipse but there are still a lot of challenges to you know to be a to be addressed to be able to achieve a scalability reasonable cost overhead performance etc if you like this tote don't miss the exact workshop it's going to be the day after tomorrow and there will be two hours of high exact and show you will have a more technical explanation of all the aspects of high exact and you can play with the platform learn how to create your own dab how to deploy your application how to trade computing resources or to be the next word for Wall Street for the trading CPU times thank you [Applause]