Hashing It Out
Hashing It Out

Episode 85 · 1 year ago

Hashing It Out #85- Gauntlet.Network Tarun Chitra

ABOUT THIS EPISODE

In this episode, we explore the incentives of both consensus systems and defi with Tarun Chitra is the Founder & CEO of Gauntlet. Tarun is applying his experience of AI and financial simulation to blockchains. Using ‘agent-based simulation’ he helps protocol designers to discover unexpected strategies to extract more than their share of value.

Links:Gauntlet.Network

The Bitcoin Podcast Network Social Media

Donate to Hashing It Out

Discuss

Ow intmywelcome to hashing it out,paast Fer, wetalkd to the TEC, intevator's, mind blocked inintrostructure and decentralize that once we dive into the weeds figur atWyan helpe build this technology, the propise they face along the way I'mlooking and learn from the best in the business she can joind er wrek welcome back to everybody to Hase. YouNote, I will be your host today, along with GME everyone e say: what's up every my den? What's upeverybody deed, I don't know if we're going to do that every time, but we'reGonto borrow that from wherein colling fro now so today we have with us Taromchitra ofGauntlen Nerwark, Anja Somtins, Hey enteron and tethundercotlet, and I have kind of been stending time tryingto make the economics and and kind of cipricuencyland a little more interpretable. But I've spent a long time working onsimuation base research and hoping to applyit to PR as people develop,crazier and crazier systems Craz Yor hand. CRAZER is the plan Um, so the I think the the tramemorkquestion we ask e being this is usually what's your back story? How did you get intocripto a bogt intothe space Um Stillmot that please yeah? Definitelyso ten years ago I was working at acompany called D research and Webuilt Asex for doing physics, research, so ASEX, H, ind, like two thousand and ten I'd, saythe majority of the people who are working in asex ether, building itbuilding them for Um Telecom usecases. So a lot of like customs FFT, likeHyerpt, F, ft devices, for you know different sorts of like encoding andencryption mechanisms as well. As you know, defense husecase. But therewasn't this current frenzy, where right now, there's everone and their momtrying to build an ASEC for Solt driving cars, machine learning forCrypto et Cetera, but we we basically worked on Um. You know building this crazy, Asicprodphysicas research with the idea of applying it to drug discovery. So wewere Weber, building the simulation machine that simulated sort of thePortein purteen physics and yeah. So it as kind of s like crazybillionaires, research, lab the thing and he's spending all his o his fortuneon building these a six- and you know two thousand and Tan and eleven. Whenwhen I was working there, there were a bunch of times where we would buy chip,space, Um and the way th the chip face. Selectionworks is, you know, there's a big there's, a big physical wafer andunless you are spending a billion dollars or buying a huge amount ofchips, basically TSNC and Samsang, and peoplelike that, will just kind of tell you to...

Wel, probably even less plightly thanthis fuck off. They will instead be like go talk tosomeone who can take a hundred different orders, put them togetheronto one wafer and like make it economical for them at like a billiondollars, 'cause like they, they really can't run their fabs economically untila a million dollars. So a middleman designer that itegrates all these likelittle people who aren't important enough for their time at goes to Tasand C yeah, so the Wai. The way this works isbasically you take the wayfer. They you cut it into blocks so h, let'ssay one millimeter square box and then you auction those blocks. So, like youmake a design in RTL you symthesize Yourrtl, you say like okay: This chipis going to be like ten hundred and ten ten years, ry, ten sen meters or thetotal surf area, and then you go and buy you tell the producer hey. We needthis much pace. The producer takes your design, add in a bunch of insulation and, like other, like physical hardware,logic, to make sure that your design doesn't leak into someone else's sothat when it's Sab so most fabs work by laser retching, and so you don't wantthe laser to accidentally someone else's design of a Gat to atch intoyours, Um, and so they hold the risk of making sure that your design Zo meat,and they also are often times whenever you hear U S. politicians complainabout like China, stealing or IP. It's usuallythis step. The word this happens often because, because the these uh these integrators, who put together abunch of orders, they actually have to have more detail about like what eachcircuit is doing, to make sure that there's like enough power in a certainplace, H R, like thereis enough insulation, so that there's not youknow tunneling event stuff like that. Maybe this is too deep on ont, no feeling mixed, because this isinteresting and crazy, but I like so you guys were making eght six ofeigh six essentially to bring down the cost of a single asic. Mobilsomeone wasdoing this for them, but I during I part yeah yeah a so like, like untilBitnan got to a certain size. They also had to BIE from the same person, and sowhat happened in two thousand n eleven was one time our supplier just who's,this person W integrace seasong. They we gaved them like twenty milliondollars and then they just like ghosted us and disappeared, and they disappeared for like two tothree months, and you know, when you give someone twenty million dollars,it's still an ESCRO. So, at least it wasn't like you gave them money andthey completely ran away. They they kindogo to us, didn't tell uswhat whas going to happen et CTEA and they they basically yea three months. AnderAK will give you a ten percent discount. Now, when you, you have a hundredpeople working on a chip, and you get a three month delay because someoneghosts to you, I don't think a ten percent discount isexactly what is going to cover that. So we we like we, I guess my my workthreatened to Su them and all the stuff, and in that time that's how we foundout about this bik oin miner, which at that time later became the people whoThay Avalon Um, which was one t a earlyde comminders. Well, I'm sorrysure I was like people are willing to pay twenty million dollars to buy upchipspace. I should go mind this thing 'cause. This seems crazy so that youjump like a CUNP. The line like the delay was because this provider waslike. Oh of these people ar like th this big oint miningproducer, was willing to outbid you so that they would bump you off and screwover just thek. You get a few extra like some more letim on Onthe Hash, woryeah exactly...

M, so you know like I said you know rightnow. If you go to th a similar supplier, it's extremely competitive and you'rewaiting for for months because of sell driving car lidar because of like TP NLchips like people, building special purpose and ederdships for doingmachine learning, plus Crypti like money devices. But back then there was no one. So weweren't used to this level of like aggressive competition, and so when, when we saw that I was just likeFloridan, it was like. I need to go mine. I bunch big point, because I needto understand what this thing is. If there's like, all these people in Chinaa going crazy about this um who we work with and yeah I mind abunch and then two thoutand and thirteen. There was like a big crashand I just sold everything, and it was like I'm never gon NA, never going tobe interested in th, the stuff again and then two thousand and fifteen. Thepaper by SAPA, Linsti and Zohar on ghostgreatest heavy heaviest, greedy heaviest, serve sumtry came out,and that was the first time I was like. Oh okay. This thing makes sense,theoretically 'cause having worked inteserted systems. I was like. Ididn't really understand why this works a lot of the proability theory that isuse as wrong, like INTOA proability, of like understanding how picondistributions work was wrong in like the early papers and foreign posts andstuff like that, and so I was like okay this. This isjust a bunch of crack cots and then that paper convinced me that there'sactually this safety and livest guarantees that are probalistic and andreal and, like the proofs, made a lot of sense and it seemed like it was- wasRigoros Ond. That was when I started really paying attention aciiertur fast forid, a little bit in twothousand and sixteen I went worked in high freconcy trading and we were doingsimulations of you know, trading strategies against Yo models of themarket and at that time prurfof steak was first coming out and that's when Iwas reading a lot of these like Algaran, not pursic was first coming out, ofcourse, theres pur point at Xtnsten, but um like Grimpo intems, again yeah, maybe because I was coming from like a moreacademic place as it. This research place, where, like Ou Kn W, I was likeH. Okay, like I needed to see the stamp of Appual of Sylvia an Matay I mean now.I feel the OPO kind of like you know, oftentimes he, the lustere of thefamous person Os not is a negative, but I I I saw the Algaran paper and I Ispent a lot of time reading it. The first version was way too long. Itwas like a hundred and thirty pages and then they finally cut it down to liketheir twenty page submission h, but that was when I was really like.Okay, this the cripography is amazing, but people are making derivatives andthey're, not pricing, that in into their fifty one percent model, Um and yeah. I just started working on writing a bunch of of code to to tovalue figure how to value the dirivatives that were implicitly beingcreated in Prov of state Um and the reason I say it's De deremitiveis: is you know in per stake? You bond an asset, you walk up an affet and youget payments probalistically IG. If you get slashd, you was abtainment. If youget a block, you win a payment. You also get transaction tos which woreconpesic, but then you also only get yourprincipal at the end if you can unbonde but there's a sense in which, ifeveryone unbonds at the same time and the network holds some liabilities, you might not actually get back all ofyour principale because of slashing of answer, dilution, events or things likethat, and so it looks a lot like a slop in Somtes, where you put principle upupfront, you get theuse recurrent payments and at the end, if, ifeverything matches up, you get your principle back and I just like hadn't I was looking atthis Os. I can prove of work. You don't...

...have that unless you have derivativeson your asix and my cashpar drod is and, and those are so liquid about it, itdoesn't factor and how people value thes thing. So you have this leg. Hibr lately sounds like a very uniquepigritive experiences that, and you come across these ideas ofbeing spoken about n one language you know and which is Tinisube comcuting.We also have this experience working in high frequency trading and finance, andso ye I mean, was: What did they talkabout drivers in this paper, or was this sorte emapping the model in yourmind, from H, frompo sake to nsertitutively to byas that's sort of financial mechanisms, yeah yeah, Sor Tha? I knowmyow Very Long, winding tendent, but ISM o doring. Really that's a goodquestion. I think it. I think it's like. I didn't totally realize t when I wasreading nespapers. I was reading them out of, like you know, I was readingthem on the subway on the way to work because, like it was just like ohthey're, interesting, diserted systems thing. I think when I first realizedthat there there was this derivative aspect with some former coworkers ofmine went and raised a lot of money to do astable coin, and I was I was like this as nuts and at that time they were likewe're going to make a lar one chain we're going to have this proof saking,and I I read their paper. They like sent it to a bunch of people before, and that was the first Tim of e day.They, I think, actually rally were the first people who I had seen havewriting about writing anything about these assets as being derervitive? U, but I think that was when it really hidme. It was like just the idea that the layer one chain would have a stablecoin attached to. It is like the only way that works is, is kind of addingindiseb synthetics and Yeah Jussoso. Where does that againlike? What is it that you? I think that Lik KINDOF clarifies where you'recoming from the BACX room? What are you doing now with it ywhat? Is I tatyou're using that in say, foerkinon these days yeah, so in agrathmic finance, you know a lot ofthe problems you work on are don't have sort of cloted form, analytic, mathsolutions, Um and you're you're. Really, you have toresort to newmerical methods, mainly simuation, to try to get an ideaof what is the expected outcome of the average outtom? What's kind of theninety five percent riscayout come how much risk do I hold stuff like that?You can't just say like: Oh I'm going to hold five dollars o risk at the endof the day you can say, however, I'm holding, I probably will hold five dollars arisk, but there is two percent proability that I oh someone, tendollars and those methods, I think, are veryimportant as a compliment to you know, Mart contrafact auditing and SecurityAundityn to find out if there a e are incentive issues and bugs, and these incontracts that come from how people use them rather than like, statisticallylike behavioraly. Then then, like one plus on equal side, and if I reenterthis particular function, call I can take Alfo and basically I I worked on those tools, ermore or less a decade in in different ways, and I thought that Oh people were buildingthe systems, but they're, not stress testing them against this stuff and soyeah. I we Gauntlewe we've kind of built a bunch of tools, fore it...

...building financial models directlyagainst these smart contracts. So we have a custom version of like an Herinclient that we made that's like optimized towards stress testing,different things, and we run. You know millions ofsimulations- I guess SOE protocals and say like here, is the difference. o Satthat outcomes hs proability of certain bad things happening good thingshappening, and if you tuned these parameters, here's how you change the probility ofCertin, O antimes coming out h. So so we really work on ONSTRAS sessing this.We initially we're working on P, a lot of proof, a stake type of stuff, but Ithink nowadays you know prifistic in practice seems tobe a dinosaur that osifies very quickly and no one is willing to change thenetwork. You know like everyone's, like: Oh we're, going to add all thesecomplicated governence things, but I guess what we're never going to vote onanything because we're too afraid to cha make any paremder changes once it'srunning versus, like you know s M as malined as it is, I think, like th, thedefi space is doing, inventing the exact same types of Assa except they're,much wilder they're, ch they're willing to change parameters. Do all thesecrazy things of them, but that experimentation is way more interestingand it's also the place where I think you need more safety guarantees. So Ijust I just want o synthesize so far grntly. What what it is. Myunderstanding is like. Basically, you take more contracts. Somebody might come with Setofmarcontracts, probably KPI stuff and they're, not looking for you to see ifthere'sbugs in it in the sense of like an auditor like myself, would look forum so much as are ther like strategies that exist in in this,like as a result of this contract existing that are nine eas tointicipate, but teed two undesirable otcomes said maybe to Haglike gothat th. That's agreat way have you: Have you seen the movie office space? Yes, of course, so an office pace. Youknow, there's H, theare these people, who figure out this bug in theiremployers system where they can like pay themselves, a sen, a penny moreevery day and it compounds until they take up all the companies riht. That's exactly the type of attack thatI think you can have hen a smar contract. I, where a certain type ofuser can figure out an advantage. That's not perfect, like it's a probalistic thinglike it's fifty one percent of the time they can take more of a blocker woardthan they're supposed to than their stake allows tem to, but they can they can. You know, eventhough there's a bunch of Variante in that strategy, like one out of fivetimes hat fails, they can still keep executing it until they slowly takemore than their fair share or more than kind of the rules that, like theintentions of the person who wrote the contract, where that Revenu wassupposed to get allocated in a certain way. But someone found a strategy inwhich they can slowly take more than they're supposed to and that's theoffice space, what I call office space attack and that's the type of thing wewe stress US 'cause, those are trading strances R, a good trading strategyfinds that loophole and slowly takes all of thet and that's the type ofthing you may not want in your desentrixchang. So I guess that's that's a gooddiffrentiator from you guys from like more traditional audiving companies,but are you familiar with the blocksizceteam and what they do? I, yes more or Ess WH. What would you saymakes you guys unique in comparison to them? Yeah sure I think you know wecome at these types of things from the more of like intrastructure and finance view of this,rather than kind of a more like high level game, theoreticalversion of the world where you try to...

...make this like really complicated modelof the system that doesn't interact with the real code andis, but you try to like solve it in a more fanty way. Um. I think we spend alot of time time on building these things as trading strategies and making sure that you get the executioncorrect in trading is eighty percent of your piano likereally understanding, for instance, the the distribution of of latencies ofpacket arrival times h is the difference between you losing a hundredmillion dollars and making a hundred million dollars yeah, and so we spend alot of time on on making sure we execute the code exactly as it's run onnodes making sure that the models themselvesare aware of low level. Boctene details and, like you know, things like things like howfast they're getting packets understanding when them poolunderstanding gas fitting, and so we focus more on the. If, likelow level details than we do, O'm like hey here's, a differential game thatmay or may not have a stable dash, and do you converge to it or not? You can argue that th, the methodologythat we choose can get stuck in local minimam more easily, but at the sametime I would say that it is more practical 'cause. It can provide youmore direct, see back, and you know it's completely morbled offhow people do this and trading where you K, ow. You have these reallydissingulation environments that are exactly the same as like the livetrading environment and you constantly running simuations as your Strai tredgees lerting live to say like predict whether you should turn thingsoff or whether you should like size up or stuff like that, and it's notperfect. It's not theoretically perfect, like I can. I can give you lot ofexamples where this tepe of methonology doesn't get you to the global optimum.But from an engineer standpoint, it's much more practical. Like you can knowwhen, like uncertain malformed transaction causes a catastrophic lossversus like. I know that in this high level model that doesn't ever interactwith the real botchin there is some potential bug that that could exist. SoI I think you know we just try to focus on being like really more anengineering side and like intrstructore side versus kind of here, 's adifferential game that yer trasolve. So I don't think you said like th. TheTERMI that I usually associated with Gonla is agent simulation, Oyo,actually GE, so maybe like 'cause, that's what'shappening right. How you're exporing B Mart contract systems, so dos liveOdeis, yeah for sure so, n agent base simulation yeah. Think that that's avery good point, I think agent BA stimulation is basically taking different models ofthe different types of users and you know if we're. If we're going to takethe simplest version of this, each user has two types of parameters:number one is sort of a profit or utility function which measures how much how they value certain thingsof the system. Like it Ma there might be one type of user who is superriskoverse and they value lending in compound much more than they valueborrowing because they don't like taking liverage. They, like theirutility, comes from earning one percent. Slowly and then ther re degend, gamblertype of people who all their utility comes from just like barrowing as muchas possible and like betding it on Phomo, thrty or something I don't know.Um and someoand gambles theyre, not venrational, but they might stumble across something Exat, crackd yeah. So so tthat's a good point. So th an idea is, you know you model things in in twoways: wonderstus utility or profit, and...

...one is the decision function and thedecision function in in in Classco econometrics, for instance, is thething that says: Hey, pull the slot machine lever or go, spend it on pokeror go. Invest it in a you know a bond it it's a thing that represents theactions that you actually take and in our case D, this is something again. Ithink we we differ quite a bit from blocks sine. So there's our actions arethe actual smart contract transactions and interacting with the real code,interacting with the blocchain itself versus like interacting with the stateconversion. That's like idealized, and you now have many of these agentsinteracting with the same block, chaingor the same smart contract andthey each have different utility functions and decision funtions. So oneof them is like Hey, I'm gonna, I'm going to be happiest. If I lend tocompound andother ones like hey, I'm going to be happiest if I borrow a lotand default in compout, because I I have this shipcoin and I would rather havedollars so I can deposit to compound- borrow stable, clains against it andsay goodbye chickcin without having to d Ki c Um, and you run many of these simulationsof like these different types of users and and try to get some statistics of.Oh, this type of user take most of the profit. Does this type of viewser havea lot of variance in their profit? But a really key thing to this is thatin distruruted systems and Crotography I sai the majority of ofpapers spend a a lot of timeworking on thismodel of people are eitherppperfectly, honest or their busating, and this getsto your point of what you're saying in terms of do Jun, Gabblers Pard, thedugen gamblers really doesn't an and in reality, they're, probably not becausemost of the assumption of Byzantine in in tcerinisystems and doesn't seemgeneral problem and stuff like that, really assumes that dizantine means anyaction that someone can take in a memory list way. Therre, not there's no,like there's no assumption that at each round of I say, like a vote iposencingsystem, someone doesn't like use how effective their previous strategy wasto influence their future strange. If you read these proofs, people alwaysprove these things, assuting kind of this memorylessness. They don'tactually assume that they're I mean now. People do care a little bit about these adaptivestrategies but t the traditional Canon, an jusuted systems, and I think this isone of the reasons why big coin a and crpicerencies took so long to beinvented, even though they're kind of simpler in some sense like s, because people like drilled it intotheir head that there's this model of purely honest and purely doesn't enthese degend gamblers. While they are bisitying, they still learn from theirprevious mistakes right, they're still trying to optimize their wealth somehowand they need to have memory of their previous actions, and so rationalagents unfortunately, is a much bigger space than honest, Indons, uh Yo. Youknow to a very particular theoretical sense:it's this infinite dimensional space, but it it's it's a large set of strategies.Then you can you only have a fine egd amount of time to stress test, thoughso you want to cover as many of those as possible, while also being able tobe computationally, feasible. Um- and that's that's! That's the art. I'd sayof these agent as models is really knowing how to cover the state spaceand perametrize. It effectively, even though you know it's this infanttimentional thing with shrarkpe complete problems everywhere. So how do you go about doing thesesimulations? What tools do you use? Have you guys written your own custom,stuff or yeah? So we we have a little. Webasically, you know for Geth andparity...

...and, like wrote, a bunch of customstuff for for for eath. Most of the reasons for for reallyhaving to build your own er fork. A lot of the stuff in the client is whenyou're running these H, clothed simulation environments. You want toremove a lot of stuff that you don't necessarily mean Um. Excuse me, for instance, like you actually do' on to Verifisignatures, why you're controlling the entire environment and you are you're,defining all the users, interacting with the system and you're going tospend most of your wallclock time, ncpu running your signature generation. Sowhy not just give people unique ideas and pretend it's trusted, because you'ryou're defining kind of all the users in the system. Another reason is that you may actuallywant to modify gas costs and see how under differentgas models or under something like e P, fifteen fifty nine like how how thatchanges, behaviorr in in gas auctions and stuff like that and so s a lot ofthe economics is actually endended quite low in t e in the clients, it'snot the submergent type of thing, and so you want it. You want to Cotrol thatother optumizations the way you have io and communication between the agents.They might sens sie, channel messages and stuff like that, and you want toAtryoathot, and the last thing is you need to model the external market, sothese agents are interacting with the block chain, but they also are tradingon Coin Daye, theyre, also interacting potentially with otherchains, but in our current environment. T S not like that. So we built our O NEnvironment. You can think of this like open ai GM or any of these like SR, reinforcement,learning playgrounds but customized around how youinteractwith, Blockchans and so or tooling, is is based on that sack. You know I when I because I guess I spent ten years N intrading an and the SASIC stuff. You know I wrote everything and C P. Pus f course: Mycofounderitwas, like let'srewrite everything in Rust, so most of our stuff now is is like kind of restand go, but we also have this sort of doing espicific language. That's in Tython, because datascientists only one es Pypon, I I I you're never going to find someonewho's like love to relbrate Rus as much as you o, rust fan or out there sayingthat, like you, can do all your newamerican rest sign me a Datoscientist who's not going to his base it, and so so we have python bindingsit that we compile for Bot the C ploplas rest goside, sothey datasign Tus, see if something that looks like Pie Torch and they can basically scripte thestrategies there and then from there. You know it runs on like te Custer, some normallike cobanity Syde of thing, and you can run you knowthousands of jobs and and a and get the results. So so the ideais like you y, the simulations are a lot of. It is centered aroundabstracting, certain things about the client away, optimizing the client andthen making a dsll for interacting e Tu Yeah. So O I gather you have prettymuch software, which is similar to blockscience cat cad, but, as youpreviously mentioned, the difference of what you guys are doing as you'retesting it with the actual block chain Yoe like not abstracting away thisentire environment. Is this stuff open source? Can I play Aron with this? If Iwant to it, it is not h. We we were working on open sourcing,some of it Whi. Just I think it's it's most of our. I I think we're at the current currentstate, we're just we've just been like really doing coaxing on boar doingthese so of audit. I getting Hev't, really untine yeah. I guess a lot ofthe time when you end up writing tooling. For yourself the coad isterrible and then you have to open source it and have to actually cleanshit up yeah exact.

I got that, but, but you guys areaiming at open sourcing. Some of these things, yeah yeah, definitely yeah. I I think H, it's it's also likewe're. We we've. We do want to support another virtualmachine at some point, so we're also trying to abstract some of the eat.parteat specific stuff out yeah, but yeah. It's not not. Quite. I thinkat that point. That's like a good point. Tho in source at well have to clean upa lot of stuff yeah in orderd. Just like one one really specific kinquestion it's like does it. You know so. You're,like looking at the application layer, insmar contracts, the Artyiu Ow, you takeinto account H,like frontrunning an and even like a chain, reord involling front running. Idon't know if you saw it it's I hard to explain an audiobe like this ral. OlDice, two win attack to haven a little while ago. Like do you think that thatyour agents would have stumbled across the man yeah so that that's a reallygood question, so we do model some aspects of consensus, so we separate B. Usually, the consensus model is not the exact cum for a bunch ofreasons. One is: is the walclock time of running the real cantensiscodasforever, because you're there's just tone of like Tine out thenetwork, the ways and like the actual way that these clients implement liketiappear networking, means that if you try to run the real contensivs code,you'll end up spanding, almost the real lalcork time and your goal andsimulation is always to make simulate the ratio of simulation time to lawclock time as high as possible, because you want to get the most samples aspossible. The more samples you have, the more stantistical confidence youhave and the more you trust these numbers Um and the the some of the Contensu compaents. We we do simulate, I wit thay, on eath. We don't usually tomake contenses, I think, for for protocols that have more of their incentives connectedto consensus. REDUC. So, like we've done a lot of simnations of Sello andSEO. Isnt is an East thork, but they have basically they've takengas and they've stripped out they consenseuss pieces of death and put ina proof of Sake, consensus implemented, THAs Mar contract. So it'ssimilar to move and Webro that way, and in that situation we actually do we. Wedo model kind of like the the forking behavior and also the theinteraction of the application layer with staking m. So I don't know if wewould have exactly got the DEX toin thing. We do model the Mempool and thenwe also have models for er. How, if I send like we, you know, I think we have a model for like the networkingus, if, if I have a hundred agents Um, what's the distribution of latenciesthat they have until they hit the first minor wer like in the per topeer gossipthat were commy hops today, do they go through M, so we have on Hihbom also that andthose do capture a little bit of of this porking behivior. But I don'tthink we can have these kind of gigantic reard uh situation of Ariiation, but one reasonI bring up the UH, the incentive alignment is, you know, Ithink we we wrote this paper last ye late last year that was presented at the StanfordWashing Conference. That shows that Um,...

...these applications can actuallyinterfere quite a bit with consensus and proof a stake in a way that's muchharder than they can improfin work where, if you have a a lending protocol,say like compound on on a per state network. If the yields on compound areway higher than the staking yields than people will be incentivized unbond andmove all their stake assets into the lending vehicle and the question is: Is it possible forthere to be more than fifty percent of sake in the lending vehicle so thatfifty one percent attacks become cheaper than attacking the lendingsmart contract, and so that paper kind shows that there are conditions underwhich that will happen where basically, depending on how you structure themonetary policy for certain monetary policies, it's actually you'reinevitably going to have this thing where, let's say it's a big point,Ontar pal gets superdeflationary eventually validators say: Look, I'mgetting no rewards, no newrewards, I'm just going to move all my money intolending and then the network becomes easy to Atha. So we studied those verycarefully because those are, I think, those are the cases that are extremelyimportant, and this is, I think, fundamentially whe're proof if theycanproe wor differ if a proov stick really first, first class financialobject, which means that it's competing with the things running on it, Um and so yeah. If that answers, oh e, wedo. We do care a lot about that when that happens, 'cause th that can be onvery tatastrophic in Lolet's talk about everyone'sfavorite discussion, again a bit of bufy. What's your current take on it, wetalked a bit briefly about it, but do you think the things which arehappening right now are being done responsibly enough? Do you think mostof the protocols that are out there should be doing agent base simulations with you guys?What's your faught on that? Yes, I think in defiit. Definitely mattersmore. As I said, pursa stake so far has beenthis weird exchange game of like how much stake can the single exchangeget, and you know it ostifies and sometimes like once, there's anexchange that has like thirty percent of stake.They basically never vote on any of the governence thins, because all of themare adverse to them. They're like Oh, this, you need to think more and likeyou don't get people, don't change them very much, so I think the modlingdoesn't matter if it's basically centralized. You know as much as people want to say.DEFITHE Sentris, it has way more users than any of these prive sik networks.andthere, it's not even close in some sense Um. I think the cool thing aboutDevi is that you have a lot more transparency into when everything is gowing up and so I'll give you an analogy in thefinancial crisis h. You know everyone favor villain is moreback security, but mortgage back security is one ofthe reasons that the blowup was even more tatastrophic than it should havebeen. Is that there's not much transparency into when mortgage backsecurities are mispriced? And the reason for this is let's say you own, ahome, your your paying your Pan, Yo, youryour mortgage and all of a sudden you're like you, use your job, and yousay: Okay, I'm not going to make my interest payment, I'm going to make t eprincipal payments, because I can you know the law less. You pay yourinterest later with extra fese and then you're still in a bad time and you'relike okay. I can't pay my princeal payment and then you say: okay, I I I del, then the bank that underwrotethat or the mortgage issuer has to say. Okay, this this mortgage defaulted andthen legally they have to tell all the...

...people they sold the security s Sheheythis more he's defaulted and then only then can you be like. Oh, the price ofthe security o should reflect the fact that the defolto that whole proceduretakes six months and you have no clue when the default actually intacts theprice and there's tons of principal agent problems that are not you can'tall purtisments cant observe DFI is basically still making morerepact Securitye butit's, not let's not kid ourselves, aou yeah, it's basicallythe same thing, but the price reflects the fact that the informationpropagates almost immediately- and I think that is a big source of of loss.That's an improvement. It may be a newance improvement, but Ithink tha, the transparent y actually does help. You ensure that the price,the price reflects the true risk Faster D, and so that to me, is afundamental innovation conpare to normal finands. Like you do in normalfinance, there's a lot of reason that people can try to hide information fromthe market and that doesn't get price and I think tha that that to me is likewh. What keeps me really interested of this stuff is like it's just it's sodifferent. I get that part part about the whole transparent and how thatmakes a fundamental difference to these products, but I ofsen feel like that iskind of an excuse by the protocol builders to build shitdier productsbecause they then say, ah it's transparent. If people like, if shipblows up, everyone can see, but it also assumes that the actors participatingare able to interpret the information correctly, which I think there's like abig gap between the amount of people using defi and the people who actuallyunderstand the protocols enough to be able to make like the nuanced decisionon whether something is priced in or not for sure. So that's where I think simulationssuperusal is you know you can you can give people who may not totallyunderstand all of these details an idea of what the Roi looks like and howdifferent strategies perform with this smart contract and under whatconditions it's profitable, not profitable, under what conditionsyou're holding on of risk- and I think simulation is- is the wayto bridge kind of the gap between. I wrote the contract and I know whattransaction that's omitting to hey, I'm a I'm, a user who's like I want someyield and I'm willing to take some risk, but I'm not willing to take thismutress and here's here's. My preferences encode my preferences intoan agent and run these simulations and say like that's what that' this is howill operate. I definitely think that's that's the direction that things willgo and that's the direction that normal finance hardin. So when you you ow, yougo to your four o one K in the: U S, you'll see these likemonnecarlooutcomes of like Oh. If you choose this mutual find, this is likethe average performance. This is like what you m the best case worst case. Anthey'll give you all, and people have already internalized all thesesimulations that hide a lot of detail of like how thearbitrage groups in ets work or how the funding rate ar works at the end of theday every day or what why their mutual fund. You know how the Nave rebalancesitselfdaily stuff, like that, and I think cricto feels like it's going togo that direction. It J S, isn't it's like the nineteen eighty rigt yeah. So do you think for like? If I couldget more mature, it would make sense for protocols to not only have allthese simulations done, but open source them and provide like step by stepprocesses for people using these products and how to interpret all theinformation that is taken into account. Yeah. Definitely, I think I thinkthey're starting to to do much more of that. I think like, for instance, balancers UX, a boulancer for a TeBackoni is sort of a uniswap H,...

Yautoman the marketmaker, but it has a different curve and it lets youhave portfolios in assets. So, instead of trading eath for dye, you can hold aportfolio kind of ik TF, a East dy maker, some other shipcorn. I don't know T Ican't think of enough names in my headbut, some Porfoilia Um, and they actually have a tiny little pianosimulator now. Is it a very accurate like detail thing that will make? No,but I it S it's the first time I've seen something that looks like these:The things that you see normal finance when you go to a four o one K ortobetterment or wealth front wor any of these INTEC products that you peopleInestin, where it gives you some prediction: some kind of high levelthing, if, like Oh, if you make this action, this is what outcomes you mighthave, and I think part of it is open sourcing. The sigation, the part of itis also U X. I I actually think you have to find a way to present thosethings see that people don't Rolter, don't glaze their eyes over and saylike. This is too much content, yeah Um, but but you're starting to see that inin ind- and I think it's just like you know, I I tend to really think kiptoereminds me of like how markets look like an th, the kind of nineties and andeighties, where, where basically, people especi, especially when, if y,especially in the sense of h you know in in the nineties, there wasthere are a lot of people running exchanges. So so, once the Internetexisted, people just started building it changes in their house. H You. It sounds crazy, but there theresthere was a guy in West Virginia, who had like thirty percent of equitisvolume in ND, nine Huneren and ninety five h from some you Kno ExchangeyeuBuil, sitting out sitting in his basement. It it's reminiscent of like twothousand and fifteen in Viklin, where, like people were running exchanges outof their basement in some sense right. There werethe quadriga was liter Li t abad example in some ways, but condriga wasn't really just a guy in as taeeright for a while, but as a Canadian, no one hurts and basically what happens was overtimelike the incumbents started, like compet a you. Have this thing where youhad thousands of exchanges and then that fragmented liquidity so much thatpeople started consolidating andine. They consolidated the contolidationleds, a kind of these Oligopili, which looks like what thecondase finance bit next by Ta. You know whatever, like the twenty twentypeople, whore, who have most Loum and then therewas kind of eventually areaction to that, and you have all these intect products that hide that, from the end user, likeRobin Hood and user, doesn't even know what exchange they're buying stock onanymore and this huge distributed systems promem, there's thousands ofexchanges. People need to figure out th. There's a law called Raganms Tiall, the? U S stars. I shoulds covy up that says that if someone wants to buy a share, an apple you asthe broker have to go to every exchange and say find the best price possibleand sell it to them at the best price. If you don't, you ere fined, PetersssiUm and this national best ben an offer guarantee basically destroyed all human brokers.It it it it. It basically said: Hey, there's hundreds of exchanges and ifyou incorrectly report the best price you get fined, so all it did was createthis kind of Pifor concetrating...

...industry to basically optimize pricediscovery and sending pricees Oror and that led to the Robin Hoods,because that basically killed all the barinks from from market making, and itmade it such that buying data was more important than havingrelationships and like Criptos kind of has a similarthing. EXEENTIV bying data. It's like you know, people, don't trust theirgovernment and they rad or you know, go down Thi this Ramhole, but I ha thisconsolidation and then kind of alternative, uh and user interface type of things. WhatI see the sefind Defi whereat's like there are a bunch of exchanges theyconsolidate, but a fe echanges, and then there's this network of likemarket makers and and Mar market participants who are inventing a userin therface that the anduser has no clue it doesn't. H, doesn't actuallyknow what they're interacting with, and I I see a lot of similarities inincrytar. I think there's there's going to be it's going to be so abstracted totheduser hat, there's, no difference ti them between buying a share of appleand buying bickr robin that's trying to do it, they're not really doing thatthey're acriptr probaby, think in e se o. So I guess you I mean one of the things IILAS indeeds firstquestion of DFI is tecsis kind of bushy show well bestUm. Do you see it professionalizing andlike sort of becoming more steeble, thelibel liste FO money? I think NTUT? I think. Maybe it's because I wasburned in the early days of Diclin and thousand eleven and thirteen andinteracted with really shitty exchanges and which I lost. I've lost at least well, and yet at that time. So this wasjust embarrassing. Thut, like I lost at least like ten to twenty big coin. Atthat time, to basically exchangs Tut just got struck down ore likedisappeared and stole my money, an exit scamp. I I think that hers is Coun Wast, Riat yejust, it'sgotthappen everybody at some point when they're a a in S, it'Syour ID yeah,yeah. Well, it was just tee. I was just painful but UH. I think it's like a little better thanthat tr at least like yes, its harder to reason about, but man, the scamartist and two thousand and eleven and thirteen like to thirteen were justlike unreal. Like everyone was brazen we just trying to take, take or youryour coins and like doing everything possible to to to take your keys away from you. I think, while dfaze Shicro, there areclearly things that seem to be animprovement and I think you' you'll seeconsolidation around those Um the question th th. I don't think I think,in the same way that there is like hundred of people saying, like I'mrunning a big qoin exchange to now. There's like twenty you're, going tosee the same thing where there's going to be thousands of there's going to betwo thousand and seventeen of dfit stuff. I don't know what shape thatcomes in whether it means there's a lot of capital that goes in or hether itans there's just like millions of projects, most of which are useless Um. But there's going to be something likethat and then whatever sticks from that is going to become the Coindaton biance,and I think it's going to end up just cosolidating, because people don't wantto trust their capital in too many places, and you know I think, that'swhy compounded maker kind of have the liquidity mot that people kind of trusttheir contracts for better or wors, even in spite of Bla Thursday peoplejust trust them a lot more than than the next second best option. It'sreally a like walk in effect and to some extent I think that that that's alittle bit of what happened is...

...happening for a theory Im. I remember Igot into this space being like or like full time working, an this being like.Oh there's like all these famous professors and there's all these likereally smart people working on building these competitorsand whatever, but it's pretty clear. I think to me that none of the other later ones reallyhave a serious chance. I I think they're, all going to end upbeing lar too, is for a derium at's, almost level. That's a whole huge other topic. I feellike to keep it. UNTRAC will alad. I wo goso you SA yeah. You say that there's going to be this another thosend andseventeen butjust for defy t to me. It feels like we're already there 'causein twenty seventeen, what we saw as a bunch of people, selling iseos sayingthat they're changing the world, whereas now we have dfi whereeveryone's like alwe're, giving all these financial products to people whomeet them. Whereas to me it looks like essentially just a bunch of white duneswho are traidors getting richer off of instruments that they're pumpingthemselves yeah. I think the the the difference was there was no youdidn't have to. You didn't have to prove anything intwo thousand and seventeen like the brock pierces of the worldwere able t to sell stuff without by eselling their lifestyle and notlike selling something t, a piece of code that was able to manage capital.They do tolay hraty ducks he's logiss. I I do distincly remember this party Iwas at in New York in two thousand and fifteen or sixteen, and he was like selling e s to a bunch ofpeople who were onpsychodolics and it like Holor. It was like. I don't seethat that sounds like Brok Pierce. It was definitely not the it'.Definitely not the type of thing. I think that, like the defied brows willbe able to do t, they don't have the charisma of the mighty ducks. Theydon't have the marketers of that two thousand and seventeen er it l annyingfor people to understand. I do agree with you that it's Kindofren peoplegetting richer, there's no there's nohes no getting yeah. I always laughwhenever people are like. Oh we're. Solving finance like people are GOINGTO,take a defilong to to buy a house and I'm like nother. Not. This is justtraitors trading. The people who take Defi Lonts to buytheir house already have enough cripto that is worth youer than their house toborrow again just Sol their crypto actually yeah. I I guess. Maybe it's definitely it's. Definitely. It'sdefinitely not like this fankly unbanked yeah. A lot of thetinktes aren't reallycapital efficient either. If that's the right turn to use for sure. I think that that's improving,though that's the thing where I'm I'm a little more hopeful that we will dobetter than the normal finance industry, because the lack of capital efficiencyis actually quite quite a bit of the source of inequality in that, if you, if it takes morecapital to extract a certain profit but but like only ten percent of the peoplehave enough capital to do it, then those tens at people will be takingthat blocker word. So to speak from that that instrument, and no one else,can conjoin Um. I I' like to say the THA harborage wasthe proof of work of defi right, like you, can only kind of like exact anddefic profit. If you can like actually prove to the protocol that you didn'tarbitrage and I think that's a fondamtntal differencethan I se of because Icos Arbitrage Wa, do you trusthooded man giving youpsychodelics that tells you to buy put...

...a million dollars into year. You knowwhat it's likeit's a little bit harder now.Will there be scams of course, but you know it it it's it's it's it's Somintr,I think the hardest Parttritfi, the there's, no way to do credit right,yeah, no identity, there's, no, there's no identity and there's nthere's no there's no way to do unsecured or undersecured at leastlending right, and until until someone Sawve H, identity, which was a muchthat is the banking thnbank acrypto needs more than more than like beingable to give people. Loans is like finding a good decuntrize idantity notion that still preserves privacy. SoMa will. Just put I token en though tit'll just be like. Oh here's, yourI's, your social credit tokens. Here's yourHighti corn, lateral yeah. Well, I mean, I guess, that's why odd of the PrivacyCom like somehow you have to find a way T to give. You know an anonymous identificationthat can generate zeroknowledge proofs of certain qualities of the ownerwithout having to h leak things overtime. The problem isjust like you can't really privacy coins have tobe built. Assuming that the transaction graph can be identified, the any digualidentification system has to make sure it. It doesn't leak information basedon the number of queries that it's made because like if I make a bunch ofqueries and someone can identify thattransaction Groun, though inscrew but transaction graph mixing, is probablythe most difficult technical problem. I think that exists in in Camtyou guysmigt disagree. I don't Don t know how you fee abou that I think that's thebiggest hurdle to ever doing these decent traes ID. That is also private D. do you thinkthat you can have the Um? You talk about the benefits of Dvi,other things, being transparency, so being publicly visible that you knowthings are being repriced. For example, can you get some kind of like that sortof public partet ability in some way in in a balance with some form of privacy,confidential transactions or or disintroducing, is introducing that through somethinglike an Aztec or like? I know that there's actually like more of aspectrum of primacy n conpidentiality with various technologies but isthere's,some mix of that that also allows you to have the gright level, O Audiability and public posibility into the actimies yeah. I think th again, all of thesethings boile own to somehow forcing thetransaction Gra like you can do all thes, eronolege group stuff that make asingle transaction on all. Like I'm going to say, perfect Yo Y, you you'vefigured out how to optimize Fri. You figured out how, to you know, compress your circuit reallyfast. You figured out how to do all the optimizations and bells and vhistels toMatur PR work. The problem is: That's just a single transacton and mixing thetransaction graph in such a way that you can still identify that somethingis true without leaking information overtime. I just don't know exactly how Youresalt o like there's, there's, there's some fundamental limit in nature to howthere's trade off between like number of interactions I make with the systemversus how private each of those neactions is and like the moreinteractions I make, even if they're private each time, I have to berevealing some information to the system, and I think that the only way to reallydo that is to somehow like throw away IDs a lot and and do thismixing type of thing on the transaction, graphhit, elt and but the problem with that is likethe Conpeationplese that grows way too...

...quickly. But I do think that there is there issome way of. There must be some way of doing a zerooledge proof on thetransaction graph that itself the whole transactor Gat for allpeople. That gives you that, like can do that, I just think I we're very fartechnically from the U Universe, e that happens and people Ifeel like people are really working toward that T. it's just like thatfeels to me like a twenty year and Gol Ove, like true transaction. Ga Blinding,I think the other thing is in the privacy coints based there's, just likea lot of claims that people make that are, I would say, under substantiatedright, like we've seen so many of these entirical attacks on Vicashio MinneraUm, that it is. It is very clear that, like it's really good to get single step,transaction privacy working but yeah, you somehow have to deal withthis. This temporal aspect. If you want to get all of the thingsyou ere, you were saying to work and yeah. I I just think there is somefundamental lower bound of like you can't have more than a certain numberof interactions before you start leaking someinformation there'se. Some information theoretic bound right like that says,and once we understand that, then we can design systems that try to likekeep yourself saturated as like. So that's as private as Sossibil ut I tolike, perhaps to guide you like your your personal assistant ever say: Hey,it's tend to use some new addresses or something like Broqede, your youraddresses, yeah, exactly like th. There needs to be some type of thing. T likeyou can yourself monitor that you have leaked x like some bits of information,and you say like uncomfortable weaking, this many dits and if I'm not, then Ihave to go, get Yo ite ID like. I think somehow that has to be bilgin to the UXthat have people use these systems. If you want them to be truly private yeah.That's that's a really fascinating question. Um Be interesting to see.whereer things go with that. So I think H. I think we're we're at time, and this is where we sort of start to wrapup and say. Thank you very much and maybe tellpeople where they can follow you and and len more go an and wash your workas evols po thanks yea. My twitteris is my name Tarenchitra and yeah. You can wewe write a lot ofpapers and Lodpost doldewirk and Yeh Justtan Esss. Whenever happyt HAP have tak any of the stuff utTN MAIT.

In-Stream Audio Search

NEW

Search across all episodes within this podcast

Episodes (109)