Decentralized Memory for the World Computer: A New Paradigm for Web3 | Muriel Médard - Optimum

[Music] you by now should know what that sound means we've got another great speaker coming up onto the stage this next talk is going to be meta settlement sorry it's not that is decentralized memory for the world computer it's a new paradigm for web 3 it's been given by the co-founder and CTO CEO of optimum Muriel Mard please give it up for Muriel [Applause] thanks Daniel um yes uh so let me see great so it's a pleasure to be here today um I'm Muriel Medard I'm going to talk about decentralized memory for the world computer uh and I hope to convince you uh that this is the a correct way of going about things and also is going to be the most efficient uh and future proof way to go about things uh I'm just going to tell you a little bit who I am uh so I'm going leave to pursue this fulltime from MIT at M uh was holding the NEC chair for software science and engineering for the entire School of Engineering and full professor in electrical engineering computer science I'm an MIT product um have uh three bachelor's degrees and so on and you know things were going so well in Academia you know I was getting all all all the the nice kudos and it was really something where leaving and coming to do this full-time was a huge jump and I hope to convince you as part of this talk that I made the right jump uh because I really believe that so let me step back and talk a little bit about what I made that jump so this is technology that has been coming out of my lab I've been working on for almost two decades uh you know slowly building it up and so you know what were we thinking when we started building this technology well we were starting to think about things from really the ground up you know that's generally how we try to approach things and Building Things from the ground up means to say well what does it mean to and let's start with what does it mean to be a computer and in particular what does it mean to be a decentralized computer like web 3 so you know here's a web 3 stack that uh we know and love you can argue as to whether you might you know compact some of these layers add some layers but you know I think we can most of us agree on these types of layers and I'm going to group these layers in um sort of the the top layers and the bottom layers here shown in purple with the actual Hardware shown in white uh and I'm going to try to map this to what a computer computer is starting from scratch I'm going to try to map it to the vman architecture so you know it's after World War II people are really starting to think about what does it mean to have a computer you know we've I've heard of touring and you know breaking the Enigma machine and so on and so you know there had been a lot of thought about what a computer is and uh vo noyman basically said okay here's what is to computer well you have to have an input and an output right you have to have Compu comp or something uh and you need to then spit out the answer we in agreement and then what goes inside well you have this Central Processing Unit you know our our beloved CPU which has control arithmetic uh and then you have to have the memory and you also need to have a means of getting to and from the memory so when you compute you write to memory you read from memory back and forth and if we look at uh web tree and we say you know how does that look how does this look relative to what something that V nor would recognize it doesn't look too weird okay maybe locally weird but overall you're like yeah you know you have operating systems evm svm you can argue to what extent their applications or not and then people build on top of those you have applications so it looks fairly recognizable when you go down here it doesn't look look like that at all it looks really really different from the way your computer works um so if you think of memory as being a closet and random access memory as being a closet where you put and take things out and are able to find them easily uh we don't have things that look like that at all so for instance let's look at the bus we use gossip which is really slow and anybody who's bought a b who's bought a computer computer or had to build a computer and put in a bus you know that you know if your bus is too slow that's going to become your bottleneck uh if you look at what's happening in um the memory it doesn't look like a closet where you put in things and take things out it you have the meol which looks like a pile of laundry on the floor and you have to rummage through it to see what's in it uh you have full nodes which is a closet that has everything you've ever had you know starting with like your baby onesie and then it looks so scary in that closet that you don't even want to go in there and look so you pay somebody maybe da to go and tell you whether you know the jacket that you're trying to wear is there so that doesn't look at all down here like what we would have in a normal computer and our thesis is the following if you're going to do something really different from vo noyman sure go ahead and do it but you better give me a really really really good reason as to why and why is that well you know it would be one thing if we said well look I I realize I'm doing things a funny way but hey I'm happy and everything's cheap and great no problem but you know it's not cheap and great things are slow transactions are slow storage is slow and scaling becomes really really difficult and of course one of the core aspects of decentralization is that it should be what allow ows you to scale so if somehow scaling is becoming a challenge while decentralized then that's not what we were trying to do so the approach that we're suggesting is let's go ahead and have a decentralized computer that looks like a dedicated computer and if you think of other decentralized networks that have really scale and to uh uh you know sort of been able to take the T test of time stood the test of time that's how they work they're decentralized say you know the electric grid is decentralized but I have a socket you know and that's how I have something which looks it's an abstraction which happens to be like it looks like a battery uh and it works just fine if I go to web 2 I also have something that's decentralized and I have a socket the socket is a websocket and it makes it look like I have a point-to-point connection to somebody else so what am I going to do here in web 3 I need something that's decentralized and is going to work like dedicated bus dedicated memory but I need that to work in a decentralized way and I need to have the correct apis right a socket is just an API and so I need an API to a complete distributed system okay so how do we do that on the memory so I have things that I'm trying to propagate say I'm trying to propagate blocks and we know how gossip works you know you send these blocks around and you hope that people get it eventually and it's slow why is it slow the physical links aren't slow so something's slow it must be that the protocols slow so imagine that I'm trying to propagate these uh these pieces and I'm trying to reconstruct all the different pieces if I'm trying to propagate them one by one and do replicating them and sort of gumming up everything versus if I just give you equations which represent the data the numbers that you're trying to send then I can reconstruct it from those equations very very efficiently so that's basically what we do in our uh in our in our gossip in our Optimum P2P we use you know random linear Network coding which was uh developed in my lab of which I'm co- adventure and we use it in gossip and it's provably Optimum which is why we call this Optimum it's provably Optimum some of you may have seen some of the research papers that came out uh in uh on ethereum Foundation about this and the huge gains like 50 fold gains that they see so basically this is what we are building uh with uh with our technology uh licensed from MIT okay so basically that gets us to this point here this API to actually be able to prop propagate the data remember this is what we need our bus in order to get to and from our memory unit what next well we have the decentralized memory now that I can get data around I'm can do uh reads and writs what do these reads and writes do so in the 1980s IBM went ahead and said you know let's find out what are the axioms of what a computer memory should be and let's go back to the closet I talked to you about you first have to have atomicity if I'm pulling a dress out of the closet I don't end up with just one sleeve okay I pulled the whole dress either it's there or it's not consistency this is a shared memory if I put a dress in the closet and then you put in a jacket and somebody comes in afterwards they may see that neither of us put something in or only the dress or the dress and the jacket but they're never going to see just the jacket they're never going to see some something that was not true you might see old information you're never going to see incorrect information out of order information and lastly you need to have durability if I stick something in the closet it's going to stay there until I decide that it's not going to stay there anymore okay so I need to have something which is Dependable if I turn on my computer I expect the stuff that I put in memory is there and if I want to put it in Cold Storage I do that I decide how to do that so we are able to do this in terms of reads and writes multi-reader multiwriter and the real trick here is that you know there were a couple of algorithms out there um in the computer science literature um those of you in the area may know ABD Rambo you know and if you know f it this is a very Niche area but you know the reason they stayed very fairly Niche was because they were doing a lot of duplication not using coding not using equations so the way we're able to make this scalable fast reliable and economically attractive is by using coding so then what who benefits from Optimum um validators and then of course l1s and l2s and let me push particular for validators because this is our first go to market is on validators um particularly with our first offerings of course first you need to do the bus before you do the memory uh we're looking at speeding up block propagation and lowering operational costs and of course the feedback we've gotten the reaction we've gotten from validators has been great so uh if you're a validator if you work with validators we would absolutely love to chat with you down the road then of course when you have the dram you have something that really is going to work on helping you with your me be preventing sandwich attacks remember because in that closet everything is always appearing in an order things never get out of order you can't fake anything uh and of course for l1's and l2s the gossip is able to really increase the bandwidth and reduce latency um whether you were using something like say um you know gossip sub um particularly in ethereum or whether you know you say working over salana and working over turbine which has like a 1950s Style coding making of those um uh of those uh equations you know which um which of course we we we we can Apple for uh and then of course the optimum Dam ensures Global Order is really going to basically ensure correct operation without having to go back and fix things when they go wrong Etc so uh this this framework here of the atomicity consistency durability is often called the acid framework uh versus what we usually do in web 3 which is the base you know best effort eventual consistency and basically you know just trying and seeing what happens so what's I call to action uh we're launching oh it's March 1st this month uh a private test net and running PC's um it's a private test net if you're interested please come talk to us uh please follow follow us on X uh very exciting we announced our Angel round yesterday so uh follow us on X you can see it uh we have really an amazing set of backers very supportive uh really helpful um our research papers on decentralized propagation and memory are coming in the coming weeks so I'll look for them on archive uh feel free also to reach out directly if you'd like a copy of them and with that uh I really hope that you'll try to uh contact us if you're in the uh in in the business of uh of doing the um uh of of of running nodes and I look forward to hearing from a lot of you thanks a lot [Applause]