Hashing It Out
Hashing It Out

Episode 23 · 3 years ago

Hashing It Out #23: Avalanche - Emin Gün Sirer

ABOUT THIS EPISODE

Revolutionary! On this episode we have the brilliant Dr. Gün Sirer to whet our appetites for Avalanche, a new category of consensus protocol. He distinguishes Avalanche from Paxos/RAFT systems, blockchain, and traditional gossip networks to show how this new algorithm prevents double spends while allowing for high throughput (early tests supporting over 1200 tx/s). We learn about his new coin, Ava, which will test the limits of Avalanche's public consensus on a global scale! The future is exciting!

Now, Enterin Acinaat work welcome to hashing it out APO, gass orretalkd to the atech innovators behind blocked in intrestructure anddecentralized networks. We dive into the weed FOGIR OT. Why and how peoplebuild this technology n the problems they face along the way I'm listeningand learn from the best in the business. You can join our rank, evisod twenty three hashing it out. Asalways, I'm doctor Corry, petty and I'm here with Calin Cochet, say hello,Colin Helloco and today's guest is H. Somebody we've been tryingto, get ahold of for quite a while Dr Imon Gun Serir, and we brought you onto thebickcoin podcast m a long time ago and W you've been there on multiple times.But in the first time h you often you told us that no one calls you by Amonand everyone just calls you by Gunsr or Dor serer. Is that still true- or Ithink some of your celebrity over the past couple of years is definitely hadyo people call calling you in in for a long time, you've just gotten used toit. Now No yeah e I've got started to use it more and more often Um. So Irespond to anything. I I'll respond to hey yo, whatever else, but you know allmy friends call me good. That's my given name. The first name is aformality. You inherit from your grandfather, so uh, but it's okay andUH. The way I use it. Actually I I can use it to tell apart the people. Who'vemet me from the people who have not met me, so he people who haven't met mecalled me. M, N and people have met me in person. Call Me Good: U if any emailstart with dear emmn thees spam filter is much more. That's a good good trick. elleverybodyknows that Tso Puiso. We brought you on to talk abouta paper that just kindof came out of nowhere on IPFs regarding a anewcansensus mechanism, and it's been nown that you are your distributedsystems professor and you've been doing it for a very long time before. Pickingcan you can you start to, and I think a we'd like to do is try and get both abird's ey view of what this is and what its implications are, as well as a lotof some of the technical details and WH, how it differente itself from otherconsentius models, sure and, of course, you're referring to the avalanche paper.Matiscrirery exciting paper to to hit the airwaves just to provide the bigpicture here. So the the problem of consensus has been with us for a longtime. It's a long standing issue. It's et the core of almost every distributedsystem that people use right. So all of the Google services have exactly thesame issue that bidcoin and etherium and other Cyptic cursy's face, which isGoogle, has a bunch of servers. It would like to keep them in sink. Thisis a really big, fundamental problem and also from databases. You have abunch of database servers, they store some data, you'd like to keep them insink or in dis control. You have a bunch of machines, they'd like to takeconcerted action. They should all take the same action that once again is theconsensus problem, so the problem's been with us for at least forty yearsthat I can trace back and the early attempts were quite naive and M andwere were sort of ill founded and there were a series of papers that sort oflaid out the field and established the very first approach to this problemback in the late eighties. So the best known of those papers isLesley, lamports, Byzantine generals, paper and Um. It's a fat, fascinating paper and it'slaid out the field an IT started, this a style of solutions or family ofsolutions to the consensus problem that are sound, and I call that family, theLamport LISCAF family, the two people behind it are Lasley Lamport andBarbara Liscol, and they both have to ring awards for very, very good reasonsand they've done. Like huge huge service of the community byestablishing this area and by going into depth into that particular familyof solutions and all of those solutions rely on all of us figuring out whoshould keep the data? You know the set of notes that that e are going to be holding ourinformation and then coming up with a reed prothocall in a right provocol formodifying the P Data and reading the data. So Um h. So you know these protochols arestill in existence. They're used in...

...every permission network, sohyperlegure Horda, and you name it. You know, of course, googles lock, service.You know everything that you do ends up touching one of these H, theseclassical protocols to for consensus. So that's the one family. It starts outin the late eighties and has been hashed out to to death. If you ask me,there are Umpty anddifferent variants of these, but they all come down to onefundamental sort of mode of operation. We all have a quorum. Let's say just ifyou want a picture in your mind, it would be kind of like a parliamentright. You have. You know a parliament of, let's say a hundred senators, andthen you and I are supposed to to to to enact laws that everybody should agreeon. So then what should the prodocall be? While it turns out that there aresome that work well and can be proven and they all operate by US riding to asupermajority of the senators or h the notes and then reading from somesubsent of them to ensure that the set you I do and the set I read from alwayshas a correct node in its intersection or sometimes oplorality, a majority ofcorrect notes in its intersection all depending on the fault model, so so, okay, so thats, sort of that andand just from the mode of operation. You can see that this approach willfail. If you and I don't agree exactly Onwho, the senators are okay, whoexactly is in the room. So if you- and I don't know who those people are, thenwe can it's going to be an upheld battle. Ensuring that you and I haveintersections of the right Sizein in common and if we don't have that, thenall of the invariants break down the correctness breaks down and theprotocols stop working so they're, very fragile and by their mode of operation.They entail this sort of a an almost a permissioning step. We allhave to agree on who who is going to be in the set of agreeing nodes, O who's,going to be the senators or who's going to be the modes in the system andtherefore it's kind o difficult to build something like a cryptic currency,an open, permissionless system based on these sperticals. So that's why, whenthe second family hit the scenes, it was a major breakthrough and so thatmajor Bakthrough, of course, was done by one Satocina Camoto, whose identitywer still not very clear. On and Sado, she was brilliant ahe changed theproblem in multiple different ways and I can expand on those Um. If you guyswant to, and but the bottom line is he changed the way the prolocal works. So,instead of US having to know whois going to be in this system a priori, wenow have this lottery like system. We have these minors and the minor gets.You know self elected in s. In some sense he says: Hey. I solve this reallydifficult, cripto puzzle. You guys have to now bow down to me. I am the leaderfor the last ten minutes of operation and I will tell you exactly what shouldhave happened in the last ten minutes. That's approximately how the ACOMOTOconsensis potocal works, so that was two thousand and nine. So where are wenow? So we did? Nineteen and eighty nine was the first family and it l itlasted a long time. Two Thousand and nine was the second one with NacomotoConsensus, but NAKOMOTA has a big problem and the I has a multiple problems, but oneof the biggest ones, of course, is the energy expenditure. The miners have tobe c constantly hashing. There is no notion of efficient macomotots by bydesign it has to be inefficient. There is no downtime, the minors cannot beturned off so so there is that problem. So it's not green, it's not sustainable,and you know some of the people in the bitcon circles might say you screw theenvironment. Why do we care? You know let it burn or whatever Um? I don't subscribe to that, but somepeople do even if so, okay, so, let's say screwthe environment, let the let the Poles melt. You know it's going to be nicerin Etico New York, where I live. If that were to happen actually so m, eventhen we would still have a big problem because money leaks out of the system.It has to necessarily leak out of your closed store of value ecosystem,because you have to pay those people to burn that electricity and thatelectricity has a very real, very tangible cost in the multiple millionsof dollars per day. So so that's not a very efficient way to build a storevalue system and further. Of course, there are other problems with Nacamotosuch as the long latency for Bitcoin. It's ten minutes to the firstconfirmation, sixty minutes to h something you can actually take. YouKnow Sa to the bank, Um and, of course, the limited throughput, which thenleads to limited scale so that all of that has been hashed out in great depthand so the bitcoing community has. You know it used to be controversial to saythat these days, it's not because even the bitcoing community has realized hey.We have to go to layer, two solutions to handle the scale problem m. So sothere are problems with both families. The first family of the classicallanport liscoff style was great, except...

...it's not a good fit it. It requirespermissioning and Um the second family. The NACAMOTO style,is wonderful itself except you know it has it's not green. It's notsustainable. It's slow to first confirmation slow to finality and islimited in scale. So avalanche comes in two thousand and eighteen and it's onlythe third such event in history that can I can I canfind and an itintroduces a new family and it s tat. New Family combines the best of BOT, and how does this new familydifferentiate itself? I mean where, where is the value comer? How do how doyou come into agreement by not doing the things you did previously? It'sit's very difficult in my eyes to CP to come up with Suchal system where it'sopen and permissionless for joining and leaving. Yet you can be sure that youcome to an agreement h relatively quickly and efficiently. Exactly, Ithink, that's that's why this paper is so interesting. So the core mode ofoperation for for avalanche is fundamentally different from everythingthat came before it. And U for your audience, I think the easy way tosunrise it would be its gossip inspired. So what that means is the mode ofoperation for this brotocol Um has more in common with gossip style protocolsthan it does with anything else. The correctness does not stem fromintesections and so on, and so forth, Um and uh, and so so maybe the best wayfor me to describe it is to just sort of go into it and give people a senseof howit works for e oquip clarifiation is this: Is it is it? Is it flooding oris it gossip? It's Gossi e Wat'h, the differencewellka there as a slight difference between the two now flooding and gossipare very similar. It is not a flooding protocol. It's really inspired more bygossip than by anything else. So for for the audience again, floodingprotochols are those where you have something to say and you contact all ofyour neighbors and you tell them and then all of your neighbors contact allof their neighbors and theyt tell them Soas a way front of activity inside thenetwork as the sting disseminates through everybody and Um, so they areused, for example, they're used in bid coin to flood transactions. You wantyour transactions to get into the hands of the minors and the PR peer to peerOverlay, floods, that information but they're, used in all sorts of othersettings. Flooding portocolls are very, very useful, so routing protochols use,flooding right, so internet routing, for example, uses announcements thatsay things like I can get to CNN in two hops or whatever I can get to thisother destination in fivehops. All of that information is flooded, so incontrast in gossip protocols, it's there are different variants of them.But what you do is you know a bunch of people on in the network. You haveneighbors as before, but you don't contact all of them. You pick like asmall subset and then they pick a small subset and so on and so forth and H,and so so avalanche happens to be in the latter category. It doesn't requireyou to contact all of your neighbors, you, you contact only a small substedof them Oka and I guess Tha l the the decision making process of how thatmessage, at's related across Pper to peer as as it hops is, I guess, thedifferent chator between previous gossip. It's like gops gossip inspiredprodocols, so, okay, so in in gossip protocols. What you're trying to do isyou're trying to get information into the hands of other people. Okay, so,and that's a sensible thing to want and what you do, as I say, is you pick asubset of people that you you know and then you tell them, and then they telltheir friends and so forth, and if your network is, you know, O size n, thedissemination will take log and steps right. So why? Because, in effect, whatyou're doing is you're creating a random tree on the fly Um? So just forthe audience, you guys should know what they told me right before the show theysaid our audience is pretty technically sophisticated, so talk as you Wuld to anormal colleague. So that's what I'm going to do so. But let me let me slowme down if I, if I end up saying things or using terminals, yea absolutely um B,t what what you're really doing n e gossip, Protecos you're inducing a tree,a dis distribution. Three on the fly. I pick a couple of my neighbors. Theypick a couple of their neighbors, et CETERA, Etcetra and in Loggan steps.I'm GOINGTO reach everybody, because a tree with with height log end is goingto have two to e Logan. You know, and that equals an a number of nodes ofAC and reach. So that's all fine and Andy, but gossip protocols are notconsensus, protocals right. So in gossip, the goal is to disseminateinformation and in consensus. I want to know thatyou know that we all will, at the same time, in some atomic in inincontrovertible fashion, commit this...

...particular factoid or commit thisparticular stepinestate machine transition, so so that there's a hugestep in complexity between gossip, which is Kinda like a communicationprotocol and consensus, which is an agreement potocal, it's it's it's goingto be much much much more complicated to come to consensus than to justdisseminate something, and just to be clear. So the differentiator in that isis the ordering problem. Is that correct Um? So you can gossip all thesetransactions out there, but I mean unless they know what order and whichthey're executing t end, you don't really have a truth. Single stateacross the network is there anyway is that kind of? What's differentiatingbetween avalancian and traditional gossip phrotogal right, okay, yeah,that's a good way to view it Um, so Um. I think a simpler way to me. You'reright. That's one way to B view it so the simpler way to view it in my frommy perspective is what happens when there are conflicting messages, so Um in gossip. You will get to find outboth of the messages, but you won't know which one to do you you don't youknow, do we attack at dawn? Do we not Attack Aton? We have to decide rightand what did the leader actually say M and that's that's something that gossipprotocols cannot answer for you and in consensus protocals. At the end of theday, we will have chosen to attack or not attack and all the correct nodeswill either attack n mass or they will back off and not attack again, all inin unison. So that's the that's the big difference and you're right. If wecould, if, if everybody had ha, could see the same order, then in effect wewould have achieved some kind of decision between the two, but becausethese are independent machines operating at different speeds. Seeingthe the messages and different orders Um, then, in subject to differentvusitin behaviors by by participants in the network, they they might come tosee or perceive a different order if we just use a simple gossip protocol, soUm, so that's the big difference. Gossip will just give us the info, butwill not order. It will not determine the the true to chosen h decisions,whereas the concensus protocol will decide okay. This is what happened ofthe two conflicting things. You heard the one that we're going to pick isthis one? Okay, so we have consincis protocol isbasically B minimizing the difference betweenlocal state and global state, and that has to come through some type ofdecision tiging process on t on the local level. How do you? How do you dothat? Yeah? That that's that's exactly thequestion. So, as I mentioned, we used to know only two ways of of achievingconsensus. One was the classical style, the Lamport LESCO family of bothercoals,the other was Nacamoto and then comes avalanche and I'll. Tell you howavalanche works. It's very, very simple and ulike Nakomota, which takes thewhile to figure out and unlike LAMPORTLISCA, which also takes somereasoning and and and some math to figure out. Our avalanche is very, verystraightforward, so here I so it works. Imagine that you me- and you knoweverybody else ar in a giant crowded stadium, okay, we know roughly who's inthe stadium, but we don't know everybody who's in the stadium, and youmight actually know that there are you. Might you might know of five thousandextra people that I don't know about? I might know you know five thousand ite ayou don't know about the nuts that's going to be Oky for this protocol.That's one of the and magic awesome features of it. So suppose we're inthis really really large stadium, Suppose Marti Ibo Stadium in Brazilright this is the biggest it gets. I think it's a hundred thousand people,so there we are and and we'd like to come to an agreement and- and so theseagreement brotocols are difficult right. So there will be people in the stadiumwho might want to trip us they might. They might tell us one thing and totell our neighbors something else they might flip and flock. We will call themVysantine, Byzantine actors or, if you want to think of them astrolls or that's fine too so, but you want to come to a decision, and here ishow avalanche does this and suppose the decision is we want to pick of color?Okay, we want to pick between red verses, blue, okay, that's that's ourdecision. It's abstract m! Think of it as whatever there's think of it, thatsconfirming a transaction. If you will but buty nowtat's, not let's justactually go with the color of anallergy. So we're going to pick between red andblue and imagine that everybody has a red card and a bluke. In fact, one cardwith red on the front and blue on the back and everybody initially puts thecards on their forehead H, with with their preferred color standing outright. So, what's going to happen well, the following's going to happen in thevery first case on step was zero. Let's suppose that we have the the reds onthe Blues Distirre around our stadium equally, so we have a fifty fifty splitand there as like a mixed up. People-...

...and you know, half of them have redsfacing up and half of them have blues facing up. Our goal is to do some magicand at the end of that magic we want. We want people to pick one of them. Sohere is how avalancheis magic works. Okay, it works as follows. Every note asks just a very small number of othernotes what color they have on their forehead. That's it okay! So I contact,let's say five other people in this ginormous stadium. I just pick five andI say yea. What color do you have on your forehead and back come the answersand, and the assers are going to be something like red, red, blue, red,blue and Um, let's say, and so, if that were to happen th the majority looks tome like they're picking red. All. I do is as a correct note, based on theanswers I'm hearing from the audience, based on the fact that everybody seemsto be going towards red. I add my weight too red, so I flip my own card too red. That'sall that is as simple as the spurther coal gets in fact TAT's as simple asany consensus for the collgets, so Um. So what does this accomplish? While itcreates a metastable environment right? It creates an environment where theprotocol is highly unlikely to remain in this in this byvalance state. Okay,so that by valant state is the state where it could either pick red or bluethat fifty fifty division, that's by Valent, it could go either way and Um.So this this process of trying to everybody trying to put their weight onthe same side, creat open a sort of an inherently unbalanced situation. Whatwe call a METO, stable situation, so mede stable medestability is great.That's exactly what you want in M in a consensus protocol. You don't want thethe nodes waffling and flipping and flopping in the middle of the road youwant them, pushing their weight towards one decision or another, and so anavalanche. What's going to happen, okay! Well, so so we started. Fifty fiftyther were fifty thousand people with red on their foreheads and fiftythousand with blue and and everybody asked their five chosen friends whatcolor they had and nd at the end of the first round. Something interesting willhappen. What is that? But we're not going to achieve this decision in oneround that not everybody talks to everybody and and it's it's impossibleto to achieve a proper decision by the te but um at the end of the first round.Ok, what's going to happen is really interesting, we will have oversampledeither the reds or the blues. So in my example, I ended up over sampling, thereds and h. You know we have fifty w. We have a hundred thousand people doingthis intandem and uh at the end of the first round. You know, depending on howthe random dietosses come out or depending on who they witch and randomfriends they chose. We will end up with a situation. That's that's going to bemost likely. Fifty one percent read to forty nine percent blue, or vice versa.I don't know which way it's going to go, but the chances that it's going toremain and balance are Meniscu. So that's that's incredibly unlikely. Muchmore likely is th the transition towards one, the left edge or the rightedge, an oversampling and over representation of reds or blues andsuppose after the first round, Wewe veer towards t red okay. So then,what's going to happen after the second round, well, the chances that we'regoing to go towards the red is much much much higher. After a second roundof this exact same process, there were fifty one percent reds in the audience,so the chances of converting people to red are much higher. Fifty one percentthan converting somebody to blue, which is forty, nine percent, so so thatprocess is going to self amplify. So at the end of the second ground we wouldexpect you know making numbers up here. They're not know 'm rounding things up,but the chances are that we will find ourselves in a fifty three forty sevensplit. We will have gone towards the red even more and we will get fasterand faster towards that destination and then the third round. Of course, we gomore deeply into red territory. Until we reach a point that we call the pointof no return and past that point, everybody will comit it's safe tocommit. You know that the entire stadium will be the same color, so themath in the avalanche paper shows that for very low periods of of efor verylow numbers of rounds. So on the order of thirteen, or so you can achieveconsensus in very large networks. So, going back to our example, I think whatwas a Marcibo stadium hundred thousand people. I think the number is somethinglike fifteen or so so after fifteen rounds of this all you're doing isyou're sampling, five friends of yours, okay, that's all you're doing so. Atthe end of the Brotocol, you will not...

...have spoken to everybody in the stadium.To the contrary, you will have spoken to five times. Fifty, and what's thenumber, that's H, seventy five. I think so. You will have spoken to onlyseventy five people and yet you'll look around the stadium, and this is themagic everybody will be holding up the same color card against ther foreea,and that's why I think it's so cool that this thing operates in afundamentally different fashion. It's much more efficient Han, the way itworks and- and it gives you very, very strong guarantees with a fine tuningknob under the the hand of the under the thumb of the the system designeryou can change. How many rounds you go for. You can change how many people yousample to ensure that you know whatever level, of whatever probability ofreversion you're comfortable W th. This is equivalent. If you want a comparisonin Bitcoin, you have the number of confirmations, Um and th the higher.That number is. You know it's six verses. You know five versus six or soseven, the the longer you wait, the better it is in terms of whatlikelihood is ot getting undone. There are similar controls in avalanche aswell, so it'spotical that sounds really actually kindof reminds me on a grownd,eyed blue eye problem. It's a logic puzzle. That's actually the teacher Botexcat Xsenkity, whatever it is Um and it sounds a lot similar to that kind ofthat kind of problem s in that solution. An it is not. But my question is thatworks great with red versus Bullto, but what if there is eighteen, colors Umand you know m some of them qualify and some of Don? Uand like. If you have a Statin, full, very many different pieces ofconflicting information, the number of grounds it would take to come toconsensus. It seems would significantly increase, but also there's possibility,for maybe chaotic states, wor everything kind of half the audience,flitn half the IUDIENCE doesn't look or I guess eventually it would tynd tocome to consensus, but it would just takes significantly longer there'smultiplie pieces of complit information that correct not really so. The number ofoutstanding competing proposals is not really German to h to how long theprotocol runs. It doesn't matter S, imagine the same same situation. Okay,so imagine that we had three colored cards, red, blue, yellow and Um, andinstead of a fifty fifty split, I don't know what you want to have. Let's sayyou want to have one third split for each color. That's fine too, maybeactually just just take the following example. I imagine three differentdecisions: Red Blue, yellow. Let me illustrate with the t a very simple tosee. Example. Imagine that yellows are one percent. Ok in the one round, theywill be wiped out right, there'll, be no more yellows, so Um. So that'sthat's kind of interesting and if you start out in the one third one thirdone third scenario, which is the worst case for three colors while thenquickly, one of them will will dominate yeah right, the reds will will veerover the one third boundary and uh, and then they will again go into that samemetostable process and and Wullvier. You know in that that direction so m.In fact, the presence of more things to choose from makes it strictly easierfor the prolocol to decide, because the the other correct nodes get divided andM and therefore the self reinforcing stuff works, evenfaster for for the case of more more more colors but Youre Worrou that you'rite bout, one Callan, let me let me mention what you are right about. Thereis ther, have very insightful thing that that I don't want to lose, whichis it is possible for this Bolocol Soit is not possiblefor this protocol to have a safety violation. Okay, so that's one of thethings that's proven in the avalanche paper. What it says is if you pick redthe chances that I will pick blue are miniscules, so small that I don't needto worry about them. It's in fact, they are so small that they can be madesmaller than the chances of all of the oxygen atoms in this room. Justmigrating to the other side. Leaving me to suffocate just in Purni O on Thisidethat can happen it. It is possible rigt. We live in a probalistic universe.These atoms are colliding with each other. Some of them go left. Some ofthem go right and h. They could easily actually do this not easily, but theycould do this. This is a possibility, is just incredibly unlikely. We choosenot to worry about this and I, as far as I know, nobody has died of anightregen and random ballyt proabilities work yetohw,so so for that reason, Um, so, okay, so te safety violations inavalanche are incredibly unlikely, but but there is one senario which is:...

If you have a sizable Attackerokayso,you have an attacker, that's commanding, let's say twenty percent of the notesin the network. Then, unless you manage to break out of that initial, fiftyfifty split he can keep at end and he knows exactly what he's doing and he he he can see everything insidethe network. He knows exactly how people are proceeding, then he can keepyou in that zone before you start to fall off on the edges. So the protocol,this particular protocol Um, does not guarantee liveness for for decisions between conflictingtransactions. So that's a mouthful and I can expand what that means, but fornormal mortals, what it means is if somebody's trying to double spend, okayand and he launches an attack with his double spens at the same time. So sothat is, I issue two transactions, one paying Alice, one paying Bob. I also atthe network level using a lot of resources. I launch a a large attackthat requires me to see what's going on, then the paper very nicely says: Lookin that case. The attacker can keep these both of these decisions frommaking progress, so we neither pay alice nor we pay Bob and and then butthen the paper says look. This is of no consequence because of this otherconstruction that tha decisions of this kind that are stuck in limbo, limbohave absolutely no effect on other decisions, they're all independent. Soif I ever do this to you or to Allison boblet's Ay, then you guys paying eachother is completely unaffected. So so it's a very nice thing, so you know ata higher level tat we ne Ibet Impracticle, then just complete. Thisthought. Let me just complete. The stout is j St kind of a cute cute thingto to mention you know in every distributed system, you're going tohave some badness like there's some impossibilities that are always goingto have to be there, and these protocols, like both Nacamoto andclassical. You know, there's always like some badness that you need to dealwith like classical breaks down when you have bad actors over a certain threshold.HACOMOTA has other issues in this one: An avalanche all of the badness. All ofthe sort of the you can't get rid of this YOUV got to pay some price at somepoint, um its all collected into a tiny ball, and that ball is placed squarelyon the shoulders of anybody WHO's trying to cheat the system. So, ifyou're trying to cheat the system, then the system doesn't guarantee anythingfor you, and so that's good, because you aren't trying to cheat the systemin the first place. It's not going to go guarantee anything for you, but foreverybody else who wants to to assure that their transaction got committed.It provides a strong safety guarantey and provides a strong, lightne scarity.So can I? U Posit a situation where that might in reality become an issue?Maybe I'm wrong about this. Let's say: Let's say a coalition of the Matienstates. What to IMBARGO UM a smaller nation state they collect together and N, redirectall their hall theirparse Theris a thing of the PA. It's Erydiectoall thenever traffic and basically k the headlook. This is this is from now onwe're going to Um we're going to block any transaction tries to go through thenetwork from this particular set of Ip ranges and ower to do this on a state leperand owns what we're going to do. If Faces A, we control twenty percent ofthe Hash up the calculation POW. I don't know what you call it:transactual conf validation power on yet with decision. Let's call thatdecision, power, ecision power of the network, so we're going to constantlyflip from red to blue. U Every time they try to confirm, and so nobody onthe network can ever validate their information. It is that not a potential like networkLole threat for this particular Protogol. Yes, that that is a threat.So let's go through what can happen in response to that threat. So, just likethe fifty one percent attack is, is a potential attack on the Komoto Um. Ithink so. Okay, so, let's of course avalanche is going to have a number ofattacks possible against it. So let's go over it once again, so twentypercent is a rather small number. So eighty percent can easily pul pullahead of of of people who are who are trying todo soce, O, let's, let's just back up um: okay, let's back up anlet's starfrom scratch, so the Um, if you have a network where the themajority of the nodes are committed to the proper functioning of the coin?Okay, so just like just like Inbiton, youassume that fifty one percent is honest...

...in a similar context. Here we're goingto assume that the majority of the nodes Ar, in fact, thes assume a supermajority just to make it simple for the discussion here e seem a super majoritythat are committed to to seeing seeing transactions to it. If the person doingthe spending is not trying to cheat the system, these people have no chance.Okay, so that if, if it's me and I'm spending I'mgiving money only to Alice, then there is nothing they can do to stop it. So this is an interesting result. It'snot trivial, but but it is you know it's it's it's very.It's Ververy, very cool, so how? How is this possible very staiforn, somewhatstraightforward fashion, if I'm not issuing double spens, I'm doingvirtuous transactions, I'm actually just genuilly paying Alice, then thattransaction will disseminate to all. Let's say eighty percent good peopleand eighty percent will say well look I see no reason to reject thistransaction. Its signed by you know, whoever and and it should get to hisdestination. There is absolutely nothing that that twenty percent can doto counteract Um, a group as large, as as as the eighty percent of the networksaying, this should go, go forth so so that case is pretty straightforwardnow the second case is gets to be ASETARIC complicated scenarios, butagain they will come down to to the scenarios that have noconsequence for anyone except the attacker. So U, if you have a group, that's dedicatedto to sort of creating a liveless problem. They can only do that if thereis a com, thereis, a conflicting set of trasactions, if I'm paying Alice andBob at the same time, they can hold it in limbo, and that is not that I, let's actually contrast this withBitcoin in Bitcon. A minor will pick one of those two okay, a will getresolved Um, but in avalanche somebody could pick could could somebody couldlaunch an attack where they maintained the network, inperfect balance betweenAlice and Bob for a long time? Maybe okay lets I'm giving the attack or alot of power in this discussion, and I'm I'm going to get through that in asecond, because real attackers will not have this much power, but but the mainpoint here is okay, so for you know for an intellectually advance audience wehave to. We have to just be clear that you know this. PROTHOCOLE is different.If there are conflicting transactions, it doesn't guarantee liveness. So itdiffers from Nacamoto, but there should never be conflicting transactions. Aslong as the issuing person is honest, why the heck do I have the same utxospent both to Elis and to Bob? Why is it that I'm sending the same coins totwo different people? So I should never have done that and if I do do that,then the attacker um can hold us in limbo and if they do, then, then theydo nd. And what's the consequence of that, the consequence of that is feltonly by me. It doesn't hold up block production. There is no such notion ofa linear blockchain, an avalanche. It creates a directed acicgraph yeah. I was one Iwant t point that out 'cause it. It seems as though Um in the case that Iavealanche it's a typic onsens protocol that runs in parallel, like a director,a e Siograph Wood, and so you have the resolution of denial of service forsingle transactions which doesn't withhold the Denia doesn't have adetital service on the entire network, like you would, with a block producingblockchan right like if you you could deny a single double spend, it couldpotentially have problems or consequences that sprean across all ofthe transactions, then at work, but with when you have a resolution, T at's,so fine grained on the message level. Those who are trying to attack thesystem only have deies service issues or are, like you know, h issues with people coming Toconsin onthat transaction, well, not having an effect on anything else, so goodquality or good behaving transactions get through, but I have to care aboutwhat other people are doing to tonight service. That's exactly right! Thesystem will just rout around it in some sense right. It's creating this grapand H. Suddenly, the grap will have a node that splits into two and as longas the attacker isis launching at attack keeping it maintained, then thesystem might actually be in limbo and might be unable to decide. And thisagain just like you said Corey, the rest of the grap will grow around thosenotes and the system will proceed. Tat full capacity and the guys in limbowill just remain in limbo as long as the attack is in progress and UH. Ifthe attacker ever slows down for whatever reason, the system willactually pick one of the competing transactions and will actually resolveit. So that's not a problem and Um and of course there is another openissue which the paper doesn't actually advance enough, in my opinion, Um andUH. But you know everybody who's looked...

...into it, including my group in depth. H,says that there's lots of extra work to be done and lots of interesting thingst to examine um in the following scenario. What exactly is an attack orcapable of anyway, an avalanche Tsogh? The avalanche paper says: Let's, let'sassume an iniscient sort of anesa style adversary, an adversary that can seewhat everyone else is seeing an adversary that can l at whocomminicated with who who told what to whoever else and so forth. That's notthat's not actually realistic, even N S A is actually and is they can tap wires,but they they won't necessarily know what the communication s we're going touse and and Um encryption in in avalache, so s so that the paper isjust giving a lot of powers to to this crazy adversary. This this mythicaladversary, a realistic adversary, is not going to know exactly what's going on on each nowe t,it might have a statistical idea because it itself can poll, but itwon't know exactly what's happening so as a result, um we're in a veryinteresting situation. I think it it's you know, even with large large numbersof nodes adversaries might get overtaken so um. In any case, these areall aseteric academic sort of Um. You know, diversions from the mainpoint, which is this thing is, is is bulletproof in terms of its safetyguarantees and an its liveless guarantees ar alsowullet prove as long as nobody's misbehaving, and now we are we're kindolike we're going into depth on what happens when people misbehave Um withattackers that are trying to sort of do things at the same time. So it's a veryesetaric point. I didn't want to get the audience confused about this.Almost every avalized discussion kind o devolves into this because for adistributed systems geek this is these corner. Cases are the interesting oneof the main ones like the main usage of this network is going to be. I payAlice and Alice waits for my my transaction to pervade through thenetwork and she pulls a whole bunch of times and at the end of that, she isconvinced that everybody has seen this transaction and nobody will revert andthat's the end. So it's a fascinating, very, very quick process, so so thatthe phrotocol in practice ense of achieving finaality in in the matter ofa second to to a second and a half for really large networks and Um, and ithas throughputs on the order of tens of thousands of transactions per a second.So it's it's really an exciting area. Oh, it can do things that other peoplecannot do and- and my group is currently busy building a new coin ontop of the avalanche. The cort avalanche, consensus, protocol and andand ther are fascinating things. One can do on top that one could only dreamof. With other other points. I had A. I had a question about M. I feel like alot of this is in decent terms of the corner cases. We talk about withadversarial settings. Is the choice and how you randomly sample your ownfriends to ask questions that you have to assume that that sampling has donewell 'cause. You have large enough, like coalition. That can basically lieto everyone else, but if you can sample appropriately randomly, then I feellike that. It mitigates a good a lot of the power that they that I guessadversar would have Right. You should use a good random number generator Um.The the interesting thing, though, is h. You don't you know it's. It's therandom number of generations up to you and uh Um. You know it's. I don't know it'sit', it's fairly straight forward in in in how it would work right. You justpick any sort of any any good random number generator, the one that comeswith your Os. These days is good enough and, and then you sample according tothat and that'll, get you the the diversity that you need, so it just hasto be unpredictable to the attacker. AVAMACI, I think, is a reallyappropriate dame for how how how fast this gross 'cause Li e the decisionmaking process is very exponential right exactly exactly you start goingslowly away from the midpoint and then suddenly, you e just avalanche downlike in a giant mass, that's exactlyate, so I I'm still stuck on educatons and Ijust got one more to get out and sure I believe something Ardy addressd, but umthere's a backdoor in Sisco, routers Um and they you know, someody exploitsthat and winds up being able to anipulate the network so thattransactions that are posted aren't the transactions that were intended with Bashinan. There's, like an Audi,trailing kind of you know, sign things and have intent as part of the network. It doesn't seem like that's the case.You can't protast actual intense Y, actually proving that you intended todo disone, and only this on thing Um. Is that actually an issue, or am I justcoming up with something that's kind of too wacky for for reality?...

It sounds a little confusing to me. Sohow would you? How would you signal intent in bidcoin like if yous show atransaction Um? You know that that's a transaction in bitcoin right there isno. There is no notion of Woil I'm doing this payment, but I didn't reallyintend it in that coin. So so why should there be a similar you know w?Why should Vavalanche do something that the quinly doesn't do ESR, okay, so cool? So what is the progress you're making onactually building out this qoint? Ah, so that's a great question. Soconcensus protocols are fantastic building blocks, but they are not bythemselves a coin right. The coin system actually requires a Gazillianother decisions and in a lot of exciting new architecture. On top. Soat the moment I'm building a new coin. It's called Ava and UH. We created acompany to sort of take over the the development of this and M and we weresort of proceeding quite fast down the path of building a a fantastic newplatiform. I'm really excited about this, and let me try to give you thehigh level bits so, the Auvo coin, that we're building is based on avalancheand it takes advantage of the avalogical senses, protocols, protocoland m. It then, as a result, inherits all of the goodfeatures of evel. So what does that mean? Our time to finality is less thantwo seconds. So that's pretty good starting point s within two seconds.You have as much assurance as one would after I think the math. What did wee do?I think it's only the order of fifty confirmations in bitcoin or somethinglike that. It's the chances of it being reverted after two seconds. Are Ourminiscule Um. So that's! That's our starting point. Thats finality ourthroughboot is on the thousands to tens of thousands of trasactions per second,depending on the complexity of the transaction and Um. That's the throput, it's WerGreen! So if there's nothing to do, then we we can just sit there withoutconsuming any energy. There is no notion of constant energy expenditure,we're not going to consume the entire energy input of Denmark Um. Just so.People have some idea, so they can visualize it Um cripto currency miningtakes approximately two nuclear power plants, endtire net outputs, so I'msure everybody has driven at some point next to a giant nuclear power of plant.These are really really big, very impressive constructions and two ofthem, you know, imagine all the cooling towers, oil and and all that stuff twoof them are being used worldwide at the moment for confirming cryptocurrencies.That's way too much way way too much, and it's that energy could be used foranything else. If you can't do anything smelt aluminum, you will have at leastaccomplish something. So so we can achieve the same level ofsecurity as existing groupt occurrencies are in fact higher thanthat and and not require a tiny fraction of the electricity that'sgoing into CPTUCURRENCIS. So that's those are the cool futures Um that weinherit from avalanche itself, but there are other things we're buildingon top that are really exciting and I want to just get get through them very,very quickly. So number one avalanche is not going to be a single coinplatform. Yes, there will be a single coin ova that will fueel the wholesystem. That's true, but in in conjunction there will be the abilityto create any number of coins. On top that's going to be pretty exciting, somost people are somehow in this one coin on chain, mentality and uh, and sowe thought prertically about it. There's absolutely no reason why wecan't support multiple different coins, so you want to start a new coin thatrepresents real estate transactions that have unique properties. You knowyou can merge them if they're next to each other and so forth, then H, then H, then you you can do that Um or you knowyou want to represent gold on the chain on the grafh. In this case, absolutelyO and Gold has other unique um properties like if it's in the samevault and same bar. Maybe you could combine it and a so all of those kindsof special features we could we could support. In addition, we will be ableto support and we are able to support even now, multiple scripting languages,so there's going to be a base, scripting language that is simple andsafe and secure that anybody can count on El. You could introduce a corycoinwith its own unique scripting features or you, like you know. Let's say youlike: Moneros anonymity features and you like the the Bitcoin. You knowwhatever the Btcoin scripting language, you could combine them absolutely easyto do in in our language, so Um. So those are nice. Let's see what else canyou do? There are a lot a lot of other...

...things. One can do on top Um, the coinitself. The parameters of the coins itself. The coins themselves are subject togovernance. So one interesting thing you can do with Alva that you can'twith other coins, is use the the underlying consensus mechanism to pullfor social consensus, and so what does that mean? So you know in the design ofevery coin, there will be certain numbers, certain parameters that areabsolutely crucial and that are hard to guess for the system designer. So,let's take, for example, Setosi Na Comoto he ended up. No one of theparameters was how many coins shall I have outstanding twenty one million.Okay, that's easy, and that's that's! Okay, like once, you plick any number,it's it's as good as any other as long as it stays constant, that's fine, and then he picked the minting schedule.So every four years there's a halvening right and Um, and so initially everyblock gave you fifty fifty BTC. Then it was the case that it gave you twentyfive. These days you make twelve and a half pretty soon we're going to go downby another factor of too. So. How did he pick that right? So was it a goodchoice that that particular curve- and at times it was, you know it ran behind demand right,which was actually good for the coinholders, the the corn kind of youknow it Kindof. Did this mooning thing where you know, there's high demand notenough coins. The price goes to the moon, that's okay, but at times it'stoo much like for Atheorem right now, T hearias minting way more than it shouldit's. It's rewarding those minors who are not doing all that much work foryou know with too many coins, and so so that's a problem. They had the problemin the theory and now they're trying to make a whole bunch of changes to try tochange this m will an avalanche. You can do the following. You can say Iwant to push a special transaction onto the network and this is a transactionthat says, let's change the minting rate from you know. Let's say onepercent a year to compensate for loss to a zero percent, and if enough peopleagree with you, that is absolutely the kind of decision that you can make thatthat that decision will be confirmed through the same avalanchin technology,the same avalanching method and, and that will be the new law of the landand H, you know or c converse is possible too. It could well be that ifyou have just zero reward, then you won't have enough participants in thesystem. Then you might want to increase it and then once again that willrequire a special transaction, and you know- maybe you put it up to twopercent and and try to get more notes, t o Stak and more nose to jump in the ein the fray and so Um. So it's it's kind of interesting and UH. In fact, more than kind of interestingit, it frees us from having to make many of the decisions that other peopleend up making that are arbitrary and are then proven wrong by just whathappens in the real world. E Os is suffering from a bunch of decisionslike this. The Ram pricing is all kind of messed up. You know it's subject togaming and so on and so forth. Um so or IERIUM has a bunch of other issues withit. For example, the cost for perinstruction I etherium was pickedessentially on on one laptop. I gather right so people just measured how longit took to perform difference operations and then they they they put different numbers. You knowdifferent numbers and based on how long it top on a letup. Well, we canactually have it subject to governance so that people can have a meaningfulpick here. So so these are pretty interesting. There are many otherthings we're planning to do. We, of course, want to support smart contractsand we have a very interesting technique. We're doing so and at theend of the day, Alva will be a platform will be a payment's rail and acontracting rail that is much faster than anything known, much much muchmore decentralized than anything known, U and N way faster and more green thanthan everything else in existence. So I foresee a time not very long from nowwhere we end up seeing these coins consolidate we'll end up seeing a wholelot of coins just through in the towel, and these technologes are very hard tokill. They probably won't die, but a lot of communities will either pulltheir money out. WER will change their underlying technology because, what'sthe point of trying to maintain you know, fubar coins a aking rikt, it's it's much moresensible to move onto a better platform and M, and there will be techniques foractually migrating. Your utoset or your accounts balances to Ava and by doingso, you'll get o a huge, huge boose. So theytri me CAS ere picone consensusworks because it takes advantage of sanuti characteristics of theunderlying data structure or the block chain itself. There is no suchrequirement for avalanche. It's...

...literally, you have this many choices.Everyone signals what they want. The correct choice, the one that has themost an tha a subsent, should avalanche to the correct answer. Despite otherpeople trying to stop it M, so it doesn't seem it's beholden to any givendatastructure outside of a Dacg, so you can have that paralyzation of decisionmaking is that it is there something I'mmissing? Oh, that that's correct. What you said is correct: The DAG is not theDAG for avalanche. Is a performance optimization, it's very different from the dags sceneand other currencies just by the way like the way the Dag is used in Oyotaand the way the DAG is used in in avalanche are completely different,severe subtle difference, but it's it's a crucial difference, Um, so an and instead of the DAG, an avalanche.We could have used any other data structure that gives us concurrency, sowe could have used a multiset, for example. T would have been just as fineexcept multi sets are a little hard to reason about people. Don't reallyunderstand them, and you know the moment you say multi satgets confusing,whereas dags are kind of Nice to visualize. I don't know so so. Yeah we're goingwith Dagand Al Ava, Um and Um, and so and and dags are very, very nicebecause they give us that concurrency, you and I could be working on the let'ssay, north side of the DAG h. While Cory is working on the south side, andyou know, and other people are working on the westside or whatetver, so thethe DAG itself Um etpecially creates innate sharting Mi.You can just ignore some portion of it and grow some other portion of it. Sochartes work really well an embarassing pbarrassing, parallelsituation, where one worker doesn't have to care about what the otherworkers around it are doing. In regards to the decision that they're making.How do you h? How does this work in a situation where that's not the case andyou need to know information? That's dispersed across the network. Wer Youhave lik, you have raced conditions on those types of decisions. Yeah, of course, Weve thought quite abit about this. So one interesting thing is Um, let's see so in typical blockchingpothecals, you have to know the entire history of the chain- tat Ha Belacos inthis universe. In the in the Avalanche Universe, you only care about what wecall the live edge. Okay, anything that's in the past that has been spent.You could actually prune if you wanted to so so we we have the option in ournodes of pruning state and and ignoring the so t of the path o the GenesisVertex. So if you are strepped for space, you could just ignore it and youhave all the same guarantees all the same security for the live edge, evenwithout having to check it all the way back to the the genesis FORTEX and forsharding. If you decide that you wan to keep only half the Grafh, then you areperfectly capable of doing so and and and so in. In fact, if you decide youknow to be a lightweight client, that cares only about its own notes. That'sperfectly reasonable as well. So how do you think that tts of thingthat kind o throw me off about this is how you join an Unyou k? I hading jointhis Networkin, I'm ano O on you. Don't Ye know the entire history ofeverything, but have you know the Li titch? How do you know what edges areimportant? How do you know what's dependent and how do you know theinformation you're pulling from the network is correct, so you're set up ot,an initial state that is r Rel Collin, so there the idea is h verysimilar. The attack that you're mentioning is very similar, infact sidentical to what's known as an eclipse attack right, so I come into bidcoin.How do I know that the tale I got is the actual tale? How do I know thatsome idiot did not intercept my connection to the network and feed meTi's a different chain t an than the one that d the rest of the world hasfed me and in fact, in Bitcoin there are the way they get around this isthey have about nine or ten seed nodes, and so you connect to one of the seedsand the seeds themselves. Give you give you the different tales that they knowof, and hopefully none of them lie to you and then you find the hardestlongest tail end, and that then defines the the block chain forbidcoin, andit's going to be, of course, a very difficult, difficult process to try toattack this forbid going because you k now you'd have to go to all of thoseseeds and somehow intercept what they do, but but let's not fool ourselves.It's just nine or ten seeds that they have right that that's what it comesdown to in in Ava. It's somewhat similar reasoning, butwith much larger numbers all you have to do to join ha network Tis find onenode. That knows of other correct notes. You got to find just one correct note:The joint process is simply of going to a bunch of seeds and saying hey. Youknow. I know you are part of the...

...network who do you know whose part ofthe network and then walking down that path? And if you walk down that path, NYoure going to find all the other notes that speak the same, that the samelanguage and you will find all of the other nodes that have the appropriatestaking amounts that have participate in the network and youill be able to vetthe current sent that you're talking to okay, but that nos informing you ofother nodes you're, also depending on that node as being an honest actor atsome point, craxide like Um. ' I misunderstanding that so a Il find one nod and likehegivmegimme other notes right, Yeh Ladno could feed you any number ofother bad notes. They could eclipse you right. They C L, racreate, put Tencanvillage for you and Youd. Think Oh look. There are like millions of nones outhere but except theywere, all the same guy with the two horns and the tridentin his hand right. So so that's that's the situation you want to avoid andit's easy to detect those situations, because, as long as as long as you canfind one good guy who could connect, you connect you to the rest of thenetwork and once you connect to the rest of the network, you'll notice thatthere is a dag and as part of that Dag there exists a whole set of correctnose that have been vetted and have staked and so forth. And you know whatthere is another Dag on the side that has nothing in common O R, very littleuncommon, presu. Well, nothing at the live age Tatat's in common with thisparticular set of nodes and- and it turns out that all of the guys, withthe horns and the tridents and so forth, the red guys, the devils Um, thosedevilish people are all on some side, Dag and and that's not a difficultprocess to to resolve so yeah. That's something else. A Kutolike things like Aota, I was like you could have part of the network in towards baseen it's own value,proposition, meaning that there they diverged at some point it didn't seemlike there was anyway to really you reunite in a solid way. Um H, you know, l, you gave the a example earlier. Youknow you could only focus on one portion of the network as opposed tothe other portion of the network. Noyeuga of the north side and UNCARE.Carrying about the satisside now double spend would, of course, requare that Icommunicate with the other side. But U? What? If the rules of the PROTOGOLElfer getting divirgent? Is that something that we have to worry about m? That's your t! You touchim upon twodifferent things there. Let me let me try to pick them apart. The first onehas to do with Ayota. I'm I'm a little afraid of saying anything about Yota,because because one of the designers of that system misunderstands everythingand then than gets up at the and goes crazy and and sends avid, is good dumb as good, but h comefrom beyond a misunderstands things on occasion and and then and then, when hemisunderstands things, then the entire of entire Aota army comes after me andh. You know Um, so so, let's be clear about AOTAS situation. So what I saidso far about Ayota is I like Ayota, the the Dag idea. You know is a good ideaand Um, and so there there are many othersystems that use dags, Um and H. it's a good idea in AOTAS case. What I thinkyou're referring to is the problems they've had with the side tangle. Theyhave. They have this this thing on the side that some nodes seem to believe inand seem to issue transactions out of, and they are convinced that this issome kind of a long, sustained attack and- and I'm not I'm, I'm wondering-and I've always been wondering this for like in a months on end now I don'tknow how to evaluate it. I suspect that I oto might just have a buggy clientsomewhere. I don't. I can't imagine that somebody's spending all that timeand energy creating a side tangle so m. So I think it's it's not really a M.It's not really a sort of fundamental whatever it is. I think that's sort ofa different thing than what they think it is Um and Um. N AOTAS is aninteresting case. It's its correctness is still onon unknown in my book in thebook of many other researchers so m, so I think I'm going to leave Yota atthat t. The way we use the DAG isfundamentally different and Um, and I think I kinda would need to sit downwith C FD for ten minutes to describe how it's different to him, so that heunderstands that. Oh yeah, these guys are doing something iava that Eres notat all AOTA, like I alto, makes a very simple. It just uses the DAG structureto to encode the votes, whereas Ava does not, and that thatmakes a big difference in in in the final outcome Um. So, but the secondthing you ask about was what happens when there's a divergence in the setsof rules chosen by the participants...

...right? That's a big question: We'veseen that question come up in other contexts, Um. So, for example, I don't know, what's a Goodwell, youknow in bit going case right. What happens when, when your notions of whatshould be valid are different from mine, so those are cases when forks developright and in fragile systems, systems that are subject to lots and lots ofprichos and parameters. Every single parameter is a divergence point. Everysingle parameter is a fork inducing changepoint. So you want to change theblocksize Bam. You have this enormous debate, a bike shetding up the Wazopeople who used to be best friends, are suddenly blocking each other. Peopleare going on brigades, people are hiring troll squads, people areorganizing dragon, dens or whatever, and it's just a big mess, and, andfundamentally that was because your protocol really relied on a huge topoeand it turned out in this case the very first item in that Tuple, the Blocksiyecreated a diverts. There are many other things to d disagree upon and at thisrate, bitcoin cash is going to figure and figure those out for us. So theyseem to disagree on blocks eyes. They seem to disagree on features and so andso forth. Um One of the Nice things we want to do an Avalancheis to reduce thenumber of such divergence points and technthear technological tricks weredoing this, where people pay the costs associated with a change Um only if they are willing to to go alongwith it, and that's perfectly okay for me to support things that you don'tsupport and yet to be able to come to consensus on on decisions pertaining tothe to the same. So that's something that we can do Um. We believe in a large large majority ofthe cases, but there will of course be cases where you know people. Somepeople want to go this way and th some other people want to go and thenmutually exclusive direction, and so in those cases there is no choice Bot tocreate a new network. I Ao there's nothing. Anyone can do about that. Sothat's just I was just the fact of life that one has to cope with so you're,going to let market forces decide Huh, of course, as with anything in fact,but we're Gon. We want to let the market forces decide in the network. Soif people want to want to have really heavy weight, very very heavy, youeight smart contracts and not everybody wants to carry them, then at the momentyou're out of Luck Right Y. u you look at the theorium and you say I want todeploy this thing and then they they'll just say no right. There isn't enoughgas. You could pay for this contract and you're done bit. That's the end foryou. Theorum is not your platform anymore. So that's not the same. Foravaan at is not going to be the same same mechanism. So that's that's theinteresting part. We we won't have to turn people away, so I know we'regetting towards towards O very w. You know we're getting towards the end ofthe show, but Um. U Did mention spark contract support as being one of yourgoals. One of the ways Ho thetherium maintains consistencieis that everybodyon the network executes every every transaction. Call? U, in thisparticular case that that isn't necessarily going to be true, correct,not everybody on the network is going to be executing everybody else's codeor scripts. I guess m. So my question is how you ensure that things likestate management from of VI like Ebm or assum, are correctacross the network. DISIT adhere to the same principles as simple transactionsand value exchange, or is there something Um, some other protectionsthat you can implement, such as boutes, like Trubit, would offer that ould comebe inherit to the PROTOGOL. What a lovely question called, and thatwas that was incredibly in cightful. And U and I do want to answer it,except I'm afraid it's going to take another five to ten minutes and andalso it will give away a lot of the cool stuff that we have in the works.So with your permission, I just want to say this: We we have a paper coming upin the next two months, or so we call it the coin dynamics paper, the OFAcoin, dynamics paper and Um, and that's going to be the place where, where thisparticular issue is addressed, it's a very insightful question. thit sexactly the right question, which is how do you agree or must younecessarily have everybody, execute everything, intandem, ethereum styleand uh? And if you don't, then how not write? How do you not ensure? How doyou not require that? And so so let me let me take that off line if you wouldand put in a forward reference there'll be a coin dynamics paper. I'm reallyreally proud of that paper. It's you know everything I've done in my career.I've. I've typically brought this sort of a you know like this fresh lookright so um. I don't know people get sot of vostified in how they thinkabout coins. You know one coin, one scripting language, one chain, you knowand h. You know this is how they should...

...look. You should have your own vm. Itshould look like this and that and Um- and I think my approach has been tosort of question even those basics, and so we did that withote coin and it'sreally really fresh. It does almost everything differently and, and yet it gives you a veryfamiliar environment at the end of the day, so Um. If we do everything right, which youknow, we I believe we have Thi scin dynamics paper- will end updescribing a universe where Youd can take your solidity applications thingsyou you've done an etherium and port them directly to avalanche bite by bite.Just take the same code, run it on avalanche and H, and it magically works in this newsetting, but with far better guarantees, and- and you know what you can alsohave more than one CRYPTOKITI's application at the same time, which isa fascinating teen to be able to support there as a performance. So soI'm really excited, and so, with your permission, let me not answer thequestion and just say in in two months, or so there ill be a COI dynamics paperthat does exactly go into the great gory details of the technicals requiredto to handle that that question Ahi actually like to have you back on theshow at some point to discuss some of your progress, so yeah that' be that'd, be great thing tokind of bring up at that time too. I have another Ih've. Another questionhere that ywe've been we've touched on a few times ancelarily, but have itgone into the details of it and that is m all all open, permissioner splotchange that have some type of value associated with them. Um have some type of incentive mechanismbuilt in through stake with with ou no bekoin. It's upfront staking of realworld energy at the chance of solving the correct Ou, know, chryctographicpuzzles an then submita block, but proop o stake consenssus mechanisms youhave like on Chan asset staking. So then Lo inscentives antisincentives.What type of thing are you using in terms of like the incentive mechanismsfor people to do the right thing within Ovicoin? A great question also, so youknow it takes a little while to get into to get the right foundation andthen all the good questions come out at the very, very and Um so, okay, so inin Ava we're going to be using steak with minting. So the way this is goingto work is so to participate on the network. Peoplehave to have some skin in the game. I don't want everybody w every Joe Schmowith his you know, fake soct puppets try to participateand and potentially interfere with the network. We want the participants tohave some some economic incentive. Anybody can join the network to listenbut to to have decision power. You have to show that you have committed someresources to the network so that that is going to happen in the form of stake.If you wanted to have a similar, similar sort of proof of skin in thegame and Bitcoin, you would have bought minors. In this case, you tie down someresources. There's a staking period, so steak is going to be the staking amount.Is People ask me how much is it going to take to te stake at the moment wehave. We have a name for it. It's called capital Delta Oky. So no nothinghas been set in stone. Yet it's just a Greek letter at the moment. U And a ifyou can show capital, Dena, you're part of the network and Um. So with that youyou get to be consulted on every decision on every decision, notnecessarily every every decision that you know everything you hear of you canbe pulled for so, and you play a role in determininghow it Goesso. So That's Nice, U and, and the tie down period for yourcapital Delta coins is some other thing that people ask me about it's going tobe tow o. So it's another Greek letter just so Ithink I know approximately how long tow is going to be. I think I wanted to bearound about about a week, so you're not going to tie it down for fourmonths. This is not a fragile protocol that that has to have stability andcommitment an you know for four four months or whatever it's just going tobe one week so o around one week. So so you commit for one week some aount ofresources, and at the end of that that period you are guaranteed to get yourresources back. You could have you know you could have attacked the system. Youcould have done all sorts of things, but Um um, but we are not going to takeaway your coins for any reason, so a systems proof of stake systems thatpeople are are using with classical concensus prothecals necessarily end uphaving to punish misbehaviors. So if you look at Casper it has this notionof slashin flashing conditions and if you ever what we call Ikubok, that is,you say something to one person say...

...it's opposite to the other person. Ifyou ever do that. Well, that's really bad behavior in those systems and theyhave to make sure that you don't do that and therefore they have to punishyou by taking your coins away, which is really a problem. And why is theproblem? Because you can have software bugs nd? We were guaranteed to havethem. There are infinitely many bugs out there right, so you can have sulfurbugs that caus you to suddenly lose a bunch of your coins and- and that isreally so disconcerting. I would not want to build a new financial system onsuch a fragile or or some no. I I like technologies, where I can go to sleepat night, knowing that nothing bad is going to happen to me or or or thingsthat I set aside. So the coins should be safe from tampering or from protocal.You know user USORPING, Bite Protocal, so you'll get your coins back at theend of the towe period, plus a slight reward, and that's how minting is goingto be done. The amount of that reward is an annualized direct of return knownas R and Orow, I think, is what it is in anycase. U, so all these great letters are are unbound. At the moment we are goingto determine what their default values ought to be when we initially roll thecoin out. They'll be frozen for some time and then after that, they'll besubject to governance. When I say subject to governance, I don't meanarbitrary changes are possible, so in fact I actually don't want or don'tlike systems where the system can jump from any different Om one configurationto a completely different configuration all at the same time that Os not. Thatis not quiescent in my mind, that is not that's chaototiss yeah, it'definity,yea financial systems. You want them to to be somewhat. I don't know what theopposite of chaotic is Um. I don't know if I ant use a word orderly, but inthis context you want them to to to move with some inertia so and- and youalso want to understand the limits so every one of these these limits, theseparameters- capital Delta, tow et CETERA. They can, they are subject tochange, but they should be subject to only so much change, por a time period,so that people who don't like the new decision have time to to exit in in anorderly fashion. So it has to be a nice system to use, I suspect, the future orcombinations we have. I is just just right to to sort of giverise to a pretty nice system, very, very predictible, very safe from sortof somebody who's got money invested in it F perspective harmadized in a world of like like surfaceautomization, say smooth and an well behaved right. Smooth smooth is a goodtorm. Yes, we want. We want to avoid thiscontinuous behaviors when it comesto to to critical prameters like this soanyway, so yeah, that's how meanthing is going to be done. If you stake, youget some money at the end of every every towe period and you get torestake if you like or or take the money out or whatever you want to dowith it all right H. I think Lhet's start to rap on that. One that has beenthat's been a very ncychmal discussion on how ov it will work and now avalancheworks as it's Underlankano suspeconism there any type of question or commentthat you'd like to like Du, like to put forth to our audience that we didn'tget to ask you about. No, I think this is great. Were it's living through a very excitingperiod in incypal currency evolution Um? I think we ended up having whateighteen hundred different coins most of these are worthless. Weird things.Some of them are securities which are different theyre. There they'resomething else H, but ththe, sort of the new infrastructures that people areproposing all, but one of them all but ava areare coming from. You know either the the classical setting or the the lacomotive setting and and suddenlywe have a new option and people who sort of pushed forth the nokomoto idea.They ended up, pushing forth a lot and narratives around it about how thenetwork decides, how the network routes around this and that and and in manycases what they said, wasn't exactly right h. So we've seen this time andtime again. Bidcoin does not achieve social consenss, it doesn't find socialconsensus, it does not it's not geared for it. It's it's in fact can easilybreak apart. If there is a fracture at that level. U- and there are many otherthings that that are sort of pushed forth in sort of narrative form, butaren't really there in technical form. So the funny thing is for Ava Theyrethere like they ava, will find the social consensus. If there is one to befound, it will technically do so and h and there are a bunch of other thingsthat that it can do that a're in sor. The same vein of youknow this got said about Pitcoin, but BITCON didn't really do it, and here issomething that that might and so so I'm...

...excited it's a new era. I think in theevolution of these guarrencies well see a a year of consolidation next year,we'll also see other people, roll out proof of state coins and- and I think in that time, famewe're going to hopefully have ava come out and show the world whatit's like to have finality. That's on the order of a second or two and whatit's like to have a really really really high performance on the chainwet on the DAG. I should say, as well as a sing of fuatures, that that really,you know usurp everything else. So that's really sort of supersedeeverything else. That's been out there. So I'm really excited me too. There'sone other thing, O guy, an hats home. It heone other thing, okay. So this isimportant. I should really I should really nail this so so we have thisthing right, Oma coin, and it's going to be. It has the potential to beincredibly dicentralized right, and what do I mean by that? If I look atBidcon, I see nineteen miners and mining pools. Okay, that's the level ofthese centralization. You know sure we can get into details of how many youknow like those mining pools are really actually you know more people, BlahBlah Blah, and you know- and that becomes a necessary discussion of itskind, which is you know, yeah. They could really break apart and reform etCetera, but maybe not you know they Um. It's not clear that they would actuallybe able to detect malphesans by the pool operator. We've seen this time andtime again, the bitcoin gold attacks just happened where people attacked thecurrency and none of the participants. None of the miners even knew that theywere being used as part of an attack. They are just not geared for it, but inany case, let's see that to them. You know- maybe it's more, but it's notthat much more okay. So maybe you know there are a hundred entities out there.Maybe there are a thousand, but you ow, that's that's what we'r wetop out at Um. If I look at Ethereum, I see eleven miners and mining pools umdefining the entire chain. If I look at Eos, for you know all of the O W peoplemake fun of you for a bunch of reasons, but at the end of the day they havetwenty one entities that are producing blocks, so they in many ways have moredecentralization than others so m. suddenly we have ava and AVA has thepotitential to have to give an equal voice to tens of thousands to millionsof participants, so anybody can stake and anybody can then participate andand anybody can have as much voice as Gihan Wo. So you know, Gehan is, isgreat at building hardware. Heis wonder, is a brilliant guy have met him, he's afantastic, very well spoken guy and Um, and a and he's great at building theseminers that go and compute these hashes. But when it comes to cerializingtransactions, he is just like everybody else: oky He puts one transactionbefore another. That's all he does is a minor. All of that hashing is where hespends all his energy, but at the end of the day his contribution to thechain is as simple as what I just said. He just cerealizes transactions likeeveryone else, and I too could do that, and you too could do that, and in factI believe you too should do that and if you did, that would create a muchbetter world where everyone is participating in defining what happened,not leaving it up to a specialized minor class. That specialized minorclass creates a lot of issues for coins. I mentioned the Value Lea from fromstores of value, but theire other problems as well e. Why? Why are there?Why is there even a different class of entity in my network? Why aren't we allalined so so, leaving all that aside? AVA has the potential to give a voiceto anybody and to have people participate in a system ofunprecedented dicentralization levels. So you can have millions ofparticipants in this thing easily and that's a fascinating world that I hopewill be able to explore intuitively. It feels like the avalanche, wrose,inefficiency, th, the more you have hat lovelucky sentialization, coming toconsensus from the disparate distributed. Local state people whohave their own opinion on what's going on, is better the further defective the more you have that initial setright. The larger number in is the quicker you get to the final consensus.If there seems to be, I intuitively speaking, especially with like the moreoptions you have, you should also come to. You should also avalanche thecorrect decision quicker M notwell. Technically, the the thereis a parameter, the number of rounds and it does grow as you have. You havemore nose, but it doesn't grow linearly. In fact, it grows way way: SomewhenTeres Gros as log. So so it's it's going to you know. If you have tenthousand people, it might be. You know liveen rounds and if you I forget thenumbers now, if you have a hundred thousand, it's like fifteen ish. If I'mnot mistaken, no that's too much. Maybe...

...it's thirteen! I don't know I forget,but you know around seventeen rounds is the largest number that I've seen. Ifyou, I think, that's for tens of millions, if I'm not mistaken, so, yes,it grows very slowly and there is great value in having more people participatein the system. It's they're not necessarily adding to the speed. Idon't think they make it faster, but they do make it much more resilient, ao more of US yeah exactly I it they pushed back against anybody who mightacquire wealth and try to attack the system. It does sound like, though,that increases network traffic quite a bit no th t the mode of of anindividual network traffic, maybe put the load of inindividual through rounds,doesn't Gow with the size of the network. It does not that's right,soits it you end up. You end up sampling. Some number of people likefive to ten are some number of rounds. So as the network grows, you knowsuppose the number of number of people you sample is going to be the same, and if you go from eleven rounds tothirteen rounds, you know that's only whatever it is t it's going to be sorry, there's going to be twenty extrapackets for you. So that's not much at all. As long as you don't have clothesloops with in the network, it should Comu to consensus. That's the thing foryou: Athink Lonical old scale, the number of the mounto network traffickind of increases- it's Io Bet Bein Isu. It's probably notthat big deaanm, just kind of like kind, O curious, no nolet's cat triceit,because it's important M so colin its. U So for classical consensus protocols.Okay, so if you were to use Lamport viscof style protocols like what's usedin hyperlegure Um, I'm going to try to make a decision inthat in those protocols and in every case people I speak to could be lying.So then what I have to do is ask everybody, and then they have to askeverybody themselves. So if everybody has has to ask everybody else, you getyou get stuck with an en squared protocol so and square. It seems likeokay. Well, it's just a square wile, a Os Polo. U Squares grow really far HAV.If you have as few as a hundred a hundred validators, then suddenly youhave what is it a hundred times hundred ten thousand packets F flying out there?And it's just a you know it's just a lot of communication in avalanche. Ithink we just did the math a second ago if you have Um. If you have, you knowabout a hundred thousand participants now you ended up neading about seventyfive messages per decision. That's it's, and if you have more kids going to growa little bit but h, but that growth is is, is proportional to the log of thesize of the network, so inaggregate, so the number that's nsquared for classical is oofn for for avalanche, so that some total number ofpackegs that the network must bear is just all ten and the energy required.For let's say a million messages is, is a tiny fraction of of what you know say:Bitcoin or theorum consume per second Yeo, it's a very efficient protocole.Yet and if there were a a morse, lafare network growth- and I don't know ifthere is, I would guarantee that' s lower than that said. Yes, of course,yeah, and there is yes. Yes, there is a mores law equivalent for network speeds,so yeah all right. That's that's before we did HAV in oer tangent, which wecould easily do, let's Sa hats, Rapp Tirupo. Thank you, Dr Greenserer, forcoming on the show and helping US figure out how this works and, what'scoming into the future for our audience. If you enjoyd this, please cler lihtsubscribe. Tell your friends share it on Titer, you can find Uson allplatform spotify, the picorn pocast tcom, et Cetera, et Cetera, etcetera,so decrat thank before having Ne. This is a lot of him.

In-Stream Audio Search

NEW

Search across all episodes within this podcast

Episodes (108)