NEAR Whiteboard Series with NEAR | Technical Deep Dive Sessions, 9 Jan 2019

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Whiteboard Series with NEAR | Ep: 2 Anatoly Yakovenko from Solana

Description

In the second episode of our Whiteboard Series with NEAR, we had Anatoly Yakovenko (https://twitter.com/aeyakovenko) from Solana speaking with our Founders Alex Skidinov (https://twitter.com/AlexSkidanov) and Illia Polosukhin (https://twitter.com/ilblackdragon)

Learn more about Solona here: https://solana.com/

Download the video here: http://near.ai/solana10211

Follow the latest from NEAR Protocol on,

Discord: https://discord.gg/gBtUFKR
Medium: https://medium.com/nearprotocol
Twitter: https://twitter.com/nearprotocol
GitHub: https://github.com/nearprotocol

#FutureIsNEAR #BlockchainProtocols #TechTalk

A

Hey everybody: this is Alex Amelia from near and we here is Anatole from salon Houston do we have super deep dive and disgusts? We need to have details about their protocol. I tell you when they give a short intro, but yourself yeah,.

B

Sure I'm Anatole I'm, the founder, CEO solana Solanas a very high-performance watching, but the snow sharding opposite of near I'd love. To only give these guys about technology. All.

A

Right, let's start was just kind of you describe us a high level and then we'll go really deep.

B

Sure so, a very high level solana is based on this idea that you can use a clock, a global block to speed up a distributed system, and the city has been around for like a really long time. Since you know the 40s in our vlog is permissionless. So it's interesting about our clock is that it actually works before consensus like before. You actually have any notion an agreement in the system.

B

You still have a notion of time and that's why we can actually leverage that to speed up the network, so that was kind of like the the very shallow deep dive bench. So how the? How does our clock work so I? Think if you guys have been following the space and the last year, there's been kind of this explosion of ideas around verifiable delay, functions or VPS. I actually didn't didn't, know anything about PDFs when they came up with this I just had too much coffee and I was up until 4:00 in the morning an.

C

Hour clock is a maybe the simplest way you can implement the VDS. Imagine you.

B

Have a shot 256 any cryptographic, image, assistant, hash functions, work and this thing just loops over itself right, so it keeps going and going and going and everybody you know X amount of iterations. You just take a sample, so when I mean loops over itself, its output is the next input just beaker solely just runs, because it's peanut resistant, you can't any of the values that are going to occur. You know like a hundred million or eight iterations from that. You actually have to run this thing in a single core single process.

B

Without stopping so you know, let's say this is like you know: 1 million iterations and the value is like zero, be a whatever right. This is 2 million at least likes your OCD, and it just keeps going forever. So these samples represent a data structure that tells you that somebody somewhere spent real time building.

B

So what's interesting about this data structure, is you can encode messages into it? So I can simply take a message or a big blob of data and hash it and then kind of stick it in here before we do this iteration. So.

C

Now I have some.

B

Value here, I had some some bytes. It's that I've inserted here so now. At this point, all the subsequent fascists have been modified in an unpredictable way because it's premature system, so it also is interesting as if this thing itself also contained a reference to some other value that it observed. So now. This guarantees that this message was created.

B

After this point and before this point so now we have the status structure right, but in codes time, but also an order of events and the relative order between them, as well as the the up relative amount of time between them right. So we call this thing proof of history, because it gives you a proof of this historical record.

B

So now imagine you have this data structure but I send to you guys. You know, with a like a thumb, drive attached to a carrier, pigeon right.

C

B

You're looking at this data structure right- and this is like the head pointer- this is the last last thing you see and somewhere here you see a network message somewhere here you see a network message somewhere here. You see network messages right. So, given this last thing that you've observed you can compute whatever algorithm, you choose the actus that settles a network.

D

What's the sense of the network.

B

Whatever algorithm you choose, you could say that somebody had sent like a message once over the last two weeks. Right yeah make sense whatever it doesn't matter. What it is. You pick some algorithm right, but the point is that at this point this point in the in the lecture right at this point in history, which is identified by the count right, you know, 1 billion and some value, 0 X blah blah blah you've decided that this is the state of the network. So.

C

What's interesting is that everyone else observing this data structure using.

B

The same algorithm will decide the exact same state of the network, and if we optimistically assume that everybody's looking at this thing, then this is how we keep the two generals problem right. If everybody optimistically assumes that they're looking at this thing, they construct a message that references this value, but it's only valid if it appears some time later right when we say we attack right everybody attacks and if we're lucky and everyone did actually observe the same thing.

B

You'll see that as evidence here by a bunch of messages to indicate the attack which is concluded from the exact same information, everybody everywhere, sir, so that's fundamentally how our system works is that, based on this historical record, which is not cryptographic, it's not just. You know time stamps today, ashed in the you know in this right and made up, we can actually come to agreement about the state of the network and therefore make a decision, and that decision is consistent, along with everyone else optimistically. Right.

D

And so you say we cheating to generous program yeah you, so you saying that Solana would survive. Do you want me to have more than one third dimension actors.

B

So you can't like there Solutions is the generals problem right.

D

Do you think that this it would be saved? Even it is like, let's say, 40% of well-wishes.

B

Matters doesn't matter how many malicious actors are, because the way you win the two generals problem is they don't say.

D

Sorry I'm thinking over it, the young generals, probably doesn't matter the.

B

Way the way we actually solve it is when we look at this point here, we say that we solve it here right.

C

B

Know whatever one hour later, we decide oh yeah, we did actually solve it and here's the proof right, because the system then has evidence that it has come to agreement here and it has multiple confirmations and our system just like brew footwork the deeper this event occurs, the harder just a role yeah, so I can kind of like dive into that. So you may have like a bunch of forks, but could.

A

You only like one hour later, yeah.

D

No I'm not or not, yeah.

A

D

So, let's say hypothetically the person who runs the roof of history is malicious and he has to force let's say yeah and.

C

So he has a second.

B

D

B

He computers the same thing with he meets his modernization.

D

C

B

Fine, so we can kind of get into the consensus priciest one so.

D

This part, the image, so these are talking about the conscious it's.

A

Just a way for us to have.

D

Evidence that the message Adam yeah.

B

Because, once you look once you observe this structure, you don't really care how it was generated or whether there was multiple leaders. Two leaders, one, doesn't matter right kind of the date. So everything about the network is derived from the data itself. So there's no weak subjectivity and.

D

So III as a person consume this thumbdrive. How do I verify that all the hashes working good? Okay, so.

B

That's where you take your GPU part, my.

C

B

Look at each slice in between here right and then you use one core and every slice and about 4000 slices. You can verify it in this envelope, so one second can be verified in about a quarter millisecond on.

D

You so with 1gb will be very fine for time four thousand times faster. Yes, before.

B

Before GPUs are 16,000 times faster kind of like, if you look at your your like entry level of theory of mining running because I told you, he is a little gate and.

D

Then, for this position, let's say I'm planning to run a some sort of exchange and I need to be absolutely certain. You.

C

D

There's nothing weird happening: I need to verify everything from the Java support.

B

Depends depends, that's a.

D

Complicated question: if it's been four years since the Genesis we'll have a very fight in one third of a day, right.

B

Depends on your.

D

B

For parallelizable right.

D

That sounds, you seem to theme when you eat.

B

But in one hour you get spin up, you know, 100 GPUs, make sense good.

A

So for verifying, you still need to have this checkpoints relative frequency. So how big is a data structure.

B

We are generating, maybe at most two thousand per second, and that's simply because so we're kind of target is one gigabit EDB packets, there's 64 kilobytes. You can generate two thousand of them per second mm.

D

Yeah yeah Jake wise. It's the.

B

Same thing, because every time you encode the message, which is like a data packet, you need to record the state and the counter yeah II haven't seen before get paper set o first three million. If you merely have this yeah, you might be more now, there's a we haven't, measured it, but like AMD's thread, Ripper has chapter 56, specific instructions and.

D

So justin drink is very into details these days right yeah, so they want to build this easing that will be completing some other PDF. Not this one yeah.

B

So so our vdf is extremely simple: it doesn't do making fancy what they're doing is something that given like, let's say, 10 seconds worth of data. The verification time is may be constant or fast, like polynomial yeah, like so the the.

C

Difference, there's.

B

Ours grows linearly. You know, verification time is linear to the CPU time necessary to verify, but CPUs double cores every two years, so I think by the time they should be v-0 their hardware. We would be like an 8000 course right for GPU card.

D

Well, Sam I think unless the meetings unless I remember it from Justin mentioned, can wonderful because that's Alana wants to break you suggest now yeah. What's your interest, if you using different media, why.

B

When we swap it for something, that's not algorithmically better is weird we're not like. We don't care.

D

No sense yeah, but what you would be, but the.

C

Properties you want from you gave a very different from what they want like you, want it to.

D

Be you know you may want to begin through producing values, 2,000 half percent well,.

B

I'm, the inlet problem with the stuff is still in research is like they're using snart's and starts, and so the complicated things there is set up. So their approach maybe use a large MPC group, but it's like effectively like a snart's, tell setup.

A

Yeah either need.

B

A trusted group, so you need to make it huge, or you need some other mechanism to create like Kennedy initial seed. Such you know that this wasn't be computed before I know for us, because we're using sha-256 very, very dump, it is to somebody that's been working on operating systems, it's very transparent. What it does their setup requires this, like kind of like you, know, magic ceremony right, not.

A

Just they use a snark syrup, yeah yeah actually do.

C

D

A

Now you can immediately.

D

They just they just computing sound.

B

Refining operations, their operation as a celebrate, though,.

D

The semi prime right yeah yeah, so name is Simon crank.

A

Which is well yeah they're using a prime version of like the hash function. They use is like well I'm speculating here, but if you're assuming written commands and hash function- and you computed the only times you can confuse snark of that very easily, that means computation is being taxed sure.

A

B

The one part that's complicated, the other part is what we're doing is biasing the input. Every time.

C

B

Have a message in the network: we stick it into the into this loop right. The frequency of that we can. We can make it work with whatever function they do, but our main concern here is a security concern is, if you can bias this input, you can potentially generate a hash function if that, if this is not the image secure right, if this is not pretty much resistant, when you do this message, recording we want this part to be guaranteed to be secure, especially you can't like generate a loop, yeah I think.

B

Well, if done, I've talked to Justin, you feels confident that we can get our stuff working with theirs. I trust his opinion I'm happy to switch yeah. That was not a problem. Cool.

D

B

So, let's see, let's.

D

Say we have the wallet we're.

B

Gonna, do we do men's okay, so the main part about why things can be fast with a clock right is because when you do this active sub, computation, you're doing this without messaging. Anyone right, you got this data structure from a carrier. Pigeon drive. You don't have to talk to me with anyone else in the network.

B

You actually do this computation locally and that's the speed-up right, because now there's no there's no interactive, but you mean you have to do with anyone else, since they optimistically guessed that I've downloaded the right structure right and I'm gonna send it. So why is the best ship Pennings tell you why it's fast right sort of so the second part of why we can leverage this is uses that points, because in our setup we have a single leader at a time and a bunch of validators that are receiving this data.

B

So validator just don't care how to get this data right because they're deriving everything from the data directly, so this leader can broadcast. You know one over and validators orth of data to each one and then they exchange it. So that is effectively BitTorrent and you can scale this out to a log. You know tree structure, yeah right, so the fan.

C

Out of whatever.

B

Your fan own, is we verified this summer about 200? That means that in the second layer of about 40,000 notes, so with only two hops, so our finality grows with the log of the network size, and this fan can be pretty hugest apologies like stabbed Clark. It changes all the time. Yeah, there's no like how you get placed a state based it's a base in your state but effectively.

B

You know we're trying to get this group to be all the top state nodes and then everyone else and the protocol is designed such that you receive the state of partially from the network and it spattered of the race you're coding. So if X percentage of the notes- or you know malicious or send you bad data or just don't do nothing if that percentage is below the rest, your coding, you should be able to recover the full data set.

B

So I know now that this looks like a wireless network which is like it's all. Here's a welcome like actually helps here.

A

So there's the leaders so I'm assuming leader, gets rotated as well. So here's.

B

The fun part now things get interesting cell rotation right, and this is your goes into the consensus portion. So imagine we take this timeline and we divided up into slots, and these are. These are just basically counters right. You know proof in street counter. So this is you know one millions is 2 million whatever at every point and every slot there's a single leader that can belt think and transmit.

C

B

So this looks like time: division, multiple access, TDMA one of the earliest wireless protocols we've effectively taken. You know the rest of the time and split it up into slots where any single leader can be the only one that transmits.

B

So what that means is me, as a validator by either receive a transmission from this leader right or potentially, if I failed to reassemble the packets, a generator virtual tick or a virtual history yeah, because I can derive the history from the last point that I've observed by simply appending perfect, like hashes, with no data to it right. So now we have like kind of this protocol where the network either receives the data or fails to receive it, but it doesn't stop so.

D

Let's, let's say for perspective or validating to yeah sorry.

C

For those Peaks having.

D

Rather frequently right yeah, so if, if I wait first well like if I wait until I'm, confident that you know you simulated before it start computing hashes now.

B

You're always computer.

D

There's no everybody that comes in theaters hashes, but then we we're burning cheese again right. No just.

B

One chord for you for no I see yes,.

D

That's true yeah.

D

Now we have all the interesting things.

C

D

As why would they not always be candid, they never receive the message. Yeah.

B

Yeah, so so this is where, like things get actually like, what do you see? Branching is the second leader right is either observed a virtual lecture or a real one when I get to transmit- and this is either from here from here right.

C

B

Is the third one that spans this like now this the sterno right, but every time like every time, there's a new leader, a new slot? They have a number of options to pick from from the previous thing that they've observed, so the branching you know kind of starts blowing up here and how we reduce it is I, can kind of show you a different tree.

B

Does that make sense, yeah that makes sense, I'm wondering every leader, every leader either received the previous leader transmission and says: okay I received the data validated it now, I'm gonna transmit or they fail to, and they have created a virtual ledger. An.

C

B

Made from that one big group.

A

So there's options where they received one before one like one, one before etcetera, right right.

B

It gives it any kind how lighting they are great, so at any point right kind of one. So this one sends data l2 gets to decide well, did I, receive the data or did I receive the virtual. You.

C

B

So this one goes T or D right, C or D. No.3 is over here.

B

So the path here right can come from as if I'll two failed, l2 failed to observe l1 right, so they chose feet and I'm free. She actually saw, like you know, also fields of zero, one LTO and just T as well, and the only way this is chosen is if both of them failed to observe along, so that make sense, because, if l2 transmits data and this this node observed a failure for a-1- and this data depends on this one- it cannot verify. But this is correct. Okay,.

C

B

So it actually skips it and says I'm gonna branch, I'm gonna, prove that I've actually generated enough hashes up to now. Yeah.

D

Do you transmit yeah.

C

So do does not trust.

B

D

And it just needs its own portion, it doesn't just me the portion from a lot yeah in theory might have received l2 without receiving wonder and still choose to both.

B

Of them yeah, because if you not verify them thanks eyes, I mean there's repair happening and stuff like that. But the idea is that, like this, snow may actually fail to receive both and well because it failed to receive both. Its only option is to see both have failed right and what each note does is whenever they observe you know transmission data such that they vote on it. They reset their Poh to start running from there.

C

B

At any point, the last thing you voted on, that's your kind of starting point and you start producing caches and you can overrun. Basically everybody else up until it's your turn for transmit and that could be ten slots, it could be five slots, it could be zero and then you transmit that Plus hold the virtual ticks you you've generated and then you just MIT's up to the network. So the now we're playing receiving your data can actually derive it back to write.

C

B

Came from here and here's a previous tree that shows that you've waited X amount of time and what.

D

Does the for chisel, so let's say I see L 3, so 3 is sending me skills keeps me yeah, I prove yes, it's what I'm seeing here. Yes, I also see L 2. That sends me effectively transmitted via transmit yeah right, but that obviously my like right.

B

To the door in.

D

The past, which one do issues well like.

B

C

I, don't like the fun part or selection problem.

B

So imagine your validator right and we actually, let's go back a little bit to have consensus, works to kind of give a bigger, broader picture. So you have this proof. Mystery, ledger and imagine is just working kind of like at a very high level. Okay, it's working all the time and have some votes here right.

B

Every time I vote each vote is for a particular type of the ledger, and this well starts with a lock-up. Let's say and say: vote again now this voters a lockout from this photos to lock out of for right, so I vote again. This becomes 4. This becomes 8. This becomes 2. It is that every time I vote, my votes exponentially stack that my lockouts exponentially grow.

B

So this kind of gives us the same behaviors proof-of-work right we're not a pbft or VFT solution right, our Airedale abilities who are using kind of the threshold approach, but the threshold isn't based on electricity right. It's actually based on time and fundamentally, when you look at it, it's almost like a drink, but the trade-off between availability and consistency.

C

B

Cheat cap theorem right.

C

B

So literally you're shading here what you're saying is I will drop my availability for 8 units of time to break this consistency. Actually.

D

Show a couple questions votes. Is it the boating? Is this virtual right? So when I'm is a lender, you need.

B

Every validator, when they observe your data, they decide to vote on and they voting with a stake and.

D

That boat is going to be stored on the ledger so kind of my future in diffusion yeah. So let's say that social leaders so.

B

This mode, you know, maybe we've been recorded.

A

Here and actually submit submit information to the nice leaders. Our hair was in an hour so but.

D

I mean the probably need to store 40,000 votes. You.

B

Could compress them of BLS signatures in quality just.

A

A little care about data we have.

D

B

Data for all the data you want, because we we are, you know our goal is to you know: 200,000 transactions per second like steady-state right now. That's the other thousand.

D

Cool and again, and so for me to understand this law, it means that if this vote is recorded, yeah.

B

D

Some point within less so it.

C

D

C

Of time you know sometimes without most intelligence.

B

C

If someone else observes me having.

D

A competing vote within less than.

B

This is a vote for this particular branch and you voted here right. mmm Then you get sloshed, doesn't Rosen whatever you're effectively prevented from voting again Mason.

D

B

Slashing this, what is one thing we're looking at? Interestingly, all we need to do is actually just frees you from like freeze your account right from from getting rewards or foreboding mm-hmm.

C

B

Effective a slashing in terms of keeping the network steady, no.

A

You can move the money to different count.

B

No, it's like frozen. You can touch it for like six months out right so well,.

D

They might say: I moved 30 times in a row, yeah.

B

C

D

Up four billion for for all practical purposes, everything exactly so that.

B

D

Then let's say your work happened and actually like.

C

This branch was abandoned.

B

Was like, when was the last time Bitcoin reef, reorder, 32 I, don't.

D

Like these one that she's very honest, exactly attacked.

B

Right right, yeah, so pbft systems, just freeze, I, said for us to the 32 and time units. We can define that as meaning. You know in one week, financing.

D

C

Don't care one.

D

B

Right we can effectively make a practical decision to say that if a 32 blocks deep, free or needs to occur, the network should basically do nothing for a week. Yes,.

D

B

Should never happens, but if it needs to happen when I said, let's put a human constraint on it right and.

C

D

C

D

The majority of really understand that when they give them for about someone, yeah happen to.

C

Have this 20 blows then.

D

C

Say like we set your prices really neither for a week effect exactly.

B

Yeah, yes, it's about, we that's that's something we can effectively decide and even make it in that work, parameter that the network chooses and.

D

You have some extra knowledge of pros, let's say they're, looking on universe, so like 64 and.

B

D

C

D

B

Probably every season get up, it's all open source, even although all the wards are open source so feel free to like this.

A

B

Issues you know, look at the simulation and, like you.

A

Have some ovations every sinners, yeah.

B

Yeah, so that that's a really interesting part, so imagine like you guys- can observe this happening so and use like a stack yeah sure. So, when I vote here, these logons doubled mm-hmm right. So now this is 4. This is 8 and this is 16.

B

So that means that I can't really both underneath these branches- and let's say this- is a dead branch right. I have to wait until this point. It's actually both, but let's say a bunch of stuff happened and if I actually voted here.

B

What actually happens is you know, and let's say this lockup actually expired here right, so this phone is expired, okay, so Network, you know, you know he WS had like a sneeze and things stopped. Working and I only ended up floating here. They actually lockouts are like a stack and the stack is cleared and then your lockout is again- and this is still 16 question.

D

So this is e on the nice.

C

Thing the next time, oh, this becomes statically.

D

B

When this is a high C, so the way it works, yeah effectively you're building a tower like I, initially called a block tower, because if Bob goes, thank you thing yeah. So if this gets White's, this starts at 2 right now there needs to be two and four right and then, until it's fully rebuilt, you don't actually get to you know: 2 4, 8, 16 32. The idea here is that anytime, you submit a vote.

B

If this vote is at a point that is past any of the lock outs here, you clear the whole stack and you start over, and this allows you to progressively kind of do your work. So when you vote here- and you come into this branch or 32 lockouts you're, saying that I'm going to like effectively be unavailable for X amount of time to break this consistency.

C

B

So when that actually happens, do you remove this vote and you start rebuilding the stack again that make.

D

Sense, yeah, perfect sense and.

B

This allows like everybody to enroll. What else is interesting is that I can say that at 32, lock out, I want to observe 50% of the network to be at the same branch and I can make a greedy choice like this is not even Network enforce. So.

A

Your observed, like my into time, to observe 50% already voted on this branch alone or.

B

I will not I will not double my lockup past 32 on this branch, unless I've served, 50% plus I was in front right. Thanks, don't observe this expired and eventually it gets totally exponent yeah. So that is like effectively how the nodes in the network can force to be in the same branch. It's because they picked some branch behind right, like that's like at a expiration of you, know one hour, and they say that they will not commit to a branch and the network.

B

Unless an hour ago we see that 50 plus percent of the state pool has committed as well, and they.

D

Told us both what is the lockout in the total stroke, my.

B

Short, so one or two slots, one.

D

Of the slots but you're getting not can do the following getting a vote, wait for it to expire, but again wait for it. There's pirates will never.

B

Never so the reward actually comes in.

B

32 right so when you double again, this is DQ'd and that's your reward.

A

B

What this actually means is that we are kind of finalizing this ensure you belong right so 32 blocks ago. We said that this is the final branch, and this is the thing that matters right. This is a thing of the network. Will you know, take a big loss.

D

50% of people voted on this, but not all of them know that I'm having a large Tower try great it just they were so something have less to lose right, but.

B

Your threshold would be set somewhere here somewhere in the middle I'm.

D

Just saying 50% that I'm going to say 50% in this maybe are giving to this branch going back.

B

To this right, what I can do is I can actually, with both my boat and observe both l2 right, so I can observe both of these. mmm We can't we can't there's no way for us to stop those from doing this. I, don't.

D

Do is not the leader here right, no.

B

She was the leader.

A

D

And in the case where so.

B

Let's say so, let's say all tools to.

D

Live there and this slope yeah.

B

D

So this is the stand: how the three you had a faster.

C

A second over no.

D

So let's say this is time right right.

B

Yeah, okay, yeah, so.

D

They'll to be positive here and.

B

Nothing was published.

C

Here, yeah, but I as.

D

On for willing always choose the one that was later.

C

Because three is complete:.

D

B

The greedy choice, but we want the network to make this two. If they have multiple branches friend, it should be unlikely because we're using this forced away all right. So you actually have all three generate like some other data before they transmit, but thinking sheet they can use a faster ASIC or this one could just be slower right. No, so when you actually observe two branches, you should pick the one that maximizes the network reward so maximizes. The lockup I see.

C

So you, because you can examine both.

B

Branches, you can compute all the votes, and you can say that you know. Branch fund is the one that actually has the most lockup, but.

D

Isn't the point during treatment the higher the network reward is then we're inflated the salon? Token, the less.

B

Committed the more there is by net, you know you.

C

B

These lock outs, you can actually very easily compute economic finality, make sense right and the more economic finality there isn't a branch, that's the branch that you want to be a part of, because it's the one where you actually get a reward. Well, the islands die right, I, see well.

D

There's a jump if immediately obvious to me that it like.

B

I was multiplied.

D

Phenomenology, but how much these are the tokens right? It sort of sounds like.

A

A canary, presumably by dushan they're, so small, so comparators that report to me. Yes, it's.

B

Like it's an item or a brutal work right, every block dilutes Bitcoin, but you.

C

B

A

C

Of for ik, that is.

A

B

Be accepted into the main tree because you're wasting electricity right, so similarly, here you want to commit to before that has the most economic finality, because that's the one that's most likely to survive, because really for us lockout is economic finality right, mmm-hmm.

D

Just count those guys who just dropped milk beyond 2 to the 32nd, or you also like trying to predict how much people who get in the future. The.

B

Way it's written right now is it just basically adds up below lockouts from the brandeis of the penicillin, so that's and a rough approximation, but.

D

Then I might consider so.

B

Let's say everybody.

C

Has the student a.

A

6 is the same for everybody yeah right.

D

Now so this is time like the second zero on.

C

The second zero Y of the.

D

D

Second, neither of them would sail to get at your skin Gibbs.

C

Or something I, don't thinking is provided fragile G has only one block.

D

It observed and it could be pretty wealthy in something never right and that would be student right model to have been building for a while yeah and then at some point only publishes the transmitted message. Yeah.

B

Again so this one, the effect that we dropped by the network, because everyone that saw this T's have voted. Okay, the majority, the network that observe the data has voted when they observed this. That doesn't include these votes. They can't actually because.

D

They've been poaching.

B

Yeah then they blocked out right because now, if I observed three three.

A

Data transmissions.

B

My lock-on is somewhere here it, so this guy's affected me like make.

D

Sense, this is a.

C

Packet loss, yeah.

D

But this T is twice as important.

A

B

That means he's.

D

There and break.

B

So because this leader didn't observe this T, they didn't vote of it I'm, assuming that that's whatever the probability, but whatever the majority of the network has observed right if the next leader is also in that what you already, though, I'm.

A

Going to continue those guys like voyage, like for this and I'll, treat included holders phones in the first G and L to like.

B

Since intuitively think of it, this way right, I have some probability of success right and the next leader has some. You know other probability of success. If all of these are in the majority, then you continue kind of, but but as soon as somebody's in the you know, minority you know fail.

B

This node fails right in the network, has X amount of time to kind of get back to the majority state and another question so.

D

B

Make sense every time like every time we rotate there's some probability, that you've observed the data right.

A

B

And there's some probability that you're in the majority group and as long as that's true and continues to be true, we kind of keep running, keep going forward. Let's.

D

Say you make you know the chain of I'm sure the first to transmit the message. Yeah.

C

D

The second leader says: how does begin appear. The.

B

Second, the second leader is producing its own Poh from the last vote that they've taken from.

A

The last one patrons minute: they also prove that you see five disorders.

D

I would say, say so: chibi eventually be inside that isolettes analyzer I saw this transmission, which.

C

Is yeah so I'm gonna be, for example, from.

D

The other one yeah yeah.

B

D

People vote on this feat: no.

B

C

Accumulate when something is actually just me.

B

Yeah, these are virtual right, they have no data and they can be derived from the point in the state.

D

So you might have a technical example. This thing has like one vote, which was never done because nobody yeah it's going on it. While I hear this message, even though it is published later yeah, when this time has yeah so late, usually the chain which has much transmissions quickly chosen attach exactly yes,.

B

Yeah, so this this kind of saves us in the case, if somebody is like built an asynchronous circuit and they're like you know, based it laughs, thanks for that and they're. Like you know, generators.

A

Government right over and do.

B

80 here one because this will pop a bunch of boards, so it will have a lower economic finality and right, like it doesn't I think you know. Maybe, though yeah.

A

But it's possible to have like a like. If you have some portion of the steak you can like flip the finger I think what is the portion of the state doesn't need to this.

B

A

Is like I began to like this.

B

This is like an actual censorship attack, yeah right. So if you're all four and you're some portion of the steak, you can potentially just overrun the previous leader every time yeah. So you do a transmission hear that everybody yeah kills whatever the snow. Yes right so and if you're clever I mean you would include all the validator messages except the one, zero, the Sun sir, hmm and this should be the majority Network. So this is like the actual censorship attack. The only way we fight that is, the BBS will progressively approach whatever secret video.

B

If you can build and speed, so you know a difference between these zero and B. One might be to us, but if you wanted me to is gonna be fifty percent between b3 and before five percent. That's fine.

A

But so what's like right.

B

Now, for example, let's say.

C

If everybody has the same.

A

How genius hardware was a portion of the state to you need to use what so, let's say you have two leaders right now right.

B

So the rotation is steak weighted star rated yeah, so we, the active, set right, is computed from the network. So whatever it is, we compute this and then we decide. This is the new schedule mmm and you just rotate all the leaders and even.

C

A

You know a little snake known with.

B

Crappy hardware, the only thing that happens is a packet loss, so our goal is to make these slots as fast as possible. In theory could be 100 milliseconds, because this is not gonna, be computed. All the data right before their slot and started transmitting right.

A

Tomorrow, second, to China,.

B

You can you can pipeline.

A

B

So the someone who goes to shine all right- they might actually be this battle all that right. If they're in the minority state group they'll just get kind of censored out, potentially.

A

B

It's permissionless right: okay,.

A

Let's amend an example where you have a one one accountant that have twenty percent of state right so.

B

Then, if you don't, he later said once once.

A

Is the time and then they also have same signals votes so Ken's a pretty much do like transmits. You know you could go anymore, that's MIT and they add, like 2702, this branch with a candid sensor like the rest of the you.

B

Know because if they're the rest of the network is observing this, but she puts this and they're why they can't choose this unless I mean this is no can choose this right but effectively. Everyone.

A

Everybody else can't choose this branch before that they all on the same branch right right. So there will be folks to have voted for this as well. Right I can label that and your arsenal dealer has to skip right.

B

Yeah, those folks would have to actually actively participate on this. If they voted on this, they cannot move that that role is over yeah. They cannot move them over, so they're locked out right from for a better perspective. This transmissions is a failure because it doesn't include their previous boat, so it's a purpose it suffered for so they can't they can't switch over until this expires right.

B

B

They would have to effectively you can fry them. Tell them hey, you know, don't.

B

A

C

A

Way right possible.

B

Over but you can just you, don't even have to do just tell about it. If people agree, though it's the point they're, the main thing so what's interesting is in our simulation, if you have all of all the notes, instead, instead of picking the branch of the most your to pick, the branch that the least reward, but still not allow, like, let's say I think- was the 256 not allowing this lock up to be under 50% of the network. Right so knows, I just want to disrupt the network.

B

They can't actually vote on branches that die because then they just lock themselves out from the network right. So the the thing that they chooses to continue voting in a branch that becomes the main trunk, but always had kind of multiple branches, the worst one, the one has the least amount of Kannamma canal ax t. Thus network still goes forward just about like with everybody being parasitic and moves forward at about like ten percent of the rate, so that is a really cool.

B

That means that if the harbor is actually you know, you know uniform right, we can actually run this thing to be fairly censorship, resistant and fairly kind of attack resistant to parasitic votes. That's like if everybody's fair, stating financially yeah, it's still those four as well right, yeah versus like bitcoin. If everybody is voting, you know if everybody's student.

A

B

On-Brand right and just like it actually doesn't move forward.

D

So I have a couple. One of my short digression is Union feminist right. What.

B

Is that no we don't use? Rather, this also.

D

How they were selected, like how's, the wardrobe leaders.

B

Rather, this is harder, it's very hard to build a random source that is actually secure unless you.

D

B

So we do it again, like our BBF is biased, because every time.

C

We submit a message that.

B

Creates a bias, so you could in theory, like what would be easy for us to provide us. So right, you have like yeah, you know TT, but.

C

You're also everybody's computing.

B

These virtual branches from the last transmission, but you can do- is use this random value based on this. For this block, that's.

D

Not like a big year, it's.

B

A little week because anyone was just amazing, that's.

B

Yeah but that doesn't work because then everybody is throwing like 32 cores worth of PBS, and that is problematic. So for us we don't use randomness. What we actually use is just the number but number accounts the height as the seed to like see the random to just scramble this active stuff. Actually, so it's really just round-robin with a little bit of reordering as.

D

You, like the person, if I, have 10x more stake than you am I going to be.

A

Splitting splitting their account so every.

B

Time this is rotated over a long time. You will be right because this is random enough. It's just predictable. Hopefully it's totally predictable but unbiased, able I, see.

A

Wait a Centon number of accounts so when.

B

I had a number of the height, so this is like you know, 100 million. This is actually one enjoy it. That's the seed to like that's a some.

D

Residue this is this: you can emote like.

B

An advanced everybody that's going to be scheduled for the rest of time. You.

D

Know in advance everybody.

A

Was going to write and everybody's participating all regarding their descent, the next they don't decide if you want to participate. The.

B

When they stake right, there's an active, so you.

A

Can decide not to stay? If you know what yeah you can pretty much decide to stay or not just split your account and yeah.

D

So let's say: let's.

B

Say you don't know, I mean you don't have to stake, but everybody that's staked that wants to be in the rotation says up there.

D

Let them practice, let's see in practice. 99% of people are just the same.

D

If you already less for mining, if you're already my knee doesn't sense to start mining and you cannot mining yes like, presumably not that many people who can boarding, especially when they have efficiently operational the operational, then it feels.

C

Like what I can do is.

D

I can start I can create one yell any comments and then start using combinations of all those accounts to see how introducing those into the center validators yeah team comes to order right. I.

B

D

That good one I just move my assets in there. So.

B

Just to make this practical, this is computed over like two weeks right.

D

Okay, so I newbies I get three lines and find it for them to dig somebody else fresh.

B

And also their stinking requires like two weeks for about two weeks cool though you know, we just stay human timeouts. There.

D

A

That is pretty much.

D

For being shown for four hours, taking pictures full.

B

Bowls in everybody's, like you know, we're all circling with.

D

C

D

Is so the forty thousand notes you drew before energy? Each of them has a GPU right actually because he showed them used to validate now.

C

That's a point-of-sale.

D

Terminal, yes,.

C

What sort of both and.

B

Other noises yeah: this is the complicated part, so this is where Justin's magic PDF that has nice and tonic verification would be grateful to be really useful, but because that's not ready the way we're doing this is your validation message can point to the previous validation sup right.

B

So what use a thin client? In fact, this is what we're doing for kind of our like signing in play. You can observe that X amount of you know you start with some set of validators right as you trusted validators like you've observed in that. Where have you done that attestation right? This is my set of senator validators.

B

You can observe sets in the future of votes. That's simply point back so now you know that, like these guys are directing with trust us up, because they're subset of this as they voted on the branch that you care about. So if.

D

You know people thinking the hashes per state or the history.

C

You just just trust.

A

B

Of elevators general meaning this is weaker, definitely weaker than validating the time. The problem here twos we- what you need to do is like actual transaction validation right, like I, sent your transaction, and you want to quickly verify it, and this appears somewhere here in the middle of all these of all hashes and all the hashes. So what we need to do is actually pointers back up here. You're like Emeril cheat each.

D

Transaction transaction in the history isn't just a confirmation it was received. Well, that's a good information. This is valid at this point. This is valid. This is not.

C

Like a transaction.

A

Is only included.

D

Tickets, validated yeah.

A

Yeah so you're you have an accountant model or you kick so we.

B

Have an account level you connect, I mean you can in theory, have a chip. You can submit a transaction that pays out to two different accounts: yeah. Oh.

A

I mean account model of speaking yeah, but like for a camel, do you includes an infinite transaction, any new brokerage privilege after this transaction of life, we.

B

Just don't allow them to be in if they fail yeah.

A

B

If you, if, if this is just a payment, you can do that, but for smart contracts, you would need to do that. You would have to like pronoun pre write, write, something that's like a much more complicated model. But if I have joined a network.

A

So you aren't validating I need to like rerun it. The transaction suddenly came up to wait. These.

B

Validation messages include effectively the signature, the accounts right, so you can look at the network and see you know everybody is everybody. Snake is voting for these signatures. Go, give me a checkpoint and that's.

A

Not important yeah so voting is doing my daily ever. It.

B

C

Whether you're competing the.

A

Hashes, when did you.

D

Pee you right so here we have them so so, let's haven't come through. So let's say like I was the leader and I was like this? Is my your swamp yeah and then they try. This one is my historia. Another show is in industry yeah, but then perhaps for else you have sometimes action yeah right, and so this is my yeah hash. For so the.

B

Ledger of this is this is also writing.

D

In GPU, alright, so GPU has this like branch which says it has its transaction.

B

The data that you're like submitting to the GPU is just.

C

A

Hashes contractions yet, but.

B

C

Like a big blob of data and.

B

We hashed that honesty right and then we submit these.

D

No transaction that it's like a zero, these sort of heaven there's some sort of everything I see I, see, do.

A

Do it for each one, so you got like like h10 like every time for million times garbage so.

B

If I know that we don't store the 0.

B

So the zeros in shape the lunch of data structured, just contains the you know, effectively: number of counts, number of hashes number of theö ashes, they're secured in hash and then an optional passion. The blog.

A

B

Head love had to specific at this count, so when are isolated replaying this, you just simply stick it in there. But if you also have you know the transactions effective using.

A

A detector yep for this specific step, yeah and then how many do have like pretty.

B

Much your question.

A

Is so blob was to kick.

B

Sorry, 64 kilobytes depends on the transaction size because they're smart contracts very variable right now. Our maximum is 512 bytes, but like a simple move, the smallest one we can come up with this, like 176 bytes per transaction transaction, for just a 1 like 1, you know. Do you need a.

A

Minute, you need to give you two very file like.

B

Modern-Day GPUs of 16 gigabits per second I look. This is from yeah PCI 3 is one gigabit per Lane. Yes, 16 lines. You.

D

B

Some textual transactions just anything.

D

Else exactly supposed.

B

To fix you, actually we actually do send the whole transaction data, because you can't do that. Maybe signatures per second on CPUs, so we verify signatures right, yeah, yeah, that's our our defense against DDoS is just verifying signatures. You know many more signatures per second than that we can actually transmit well that won't stop like a router attack right, like everybody's splitting away yeah and what stream component you know at each node and verify way more signatures per second, but trust me yeah.

A

Yeah I mean presumably you're coming you're watching Network from making a link like children are back here. Is there any other bar to mention thanks for smart contracts, yeah yeah.

B

That is a totally fun from an excel. Yes, our contracts, you know what is the transaction for us.

B

Transaction is well, first of all running away the RIA, the extra like bottleneck and computation is in CPUs memory right, like the the problem with making things go fast is just memory pipelining through the system like if you look at just operating systems and how processes run like what you're dealing with this memory throughput like caches and all this other stuff. So how do we make this fast? Is every transactions specifies a bunch of public keys.

B

And these keys are pointers to accounts right.

B

And these accounts have data associated with them.

B

And you can think of it as user data right, that's the way I! Think of it is we have a operating system with a single address space. It just happens to be the size of public keys right, so 256 bit address space, but but also memory protection. So these accounts are assigned to process right. My process I mean you know, there's some account here whose data contains program code.

B

And this code is just code, it has no state. So this program code is a state transition function over these accounts and the accounts that are assigned to this. This is the only code they can modify the data. So whenever you doing assignment the status, bits are zero and this code is the only thing that can flip those bits right. So every time you submit a transaction, it does some bit slipping here. So what else you can.

C

B

And moves tokens around, you know power at once, and it can also pay out tokens to external to external accounts. This is outside of just code right. No, this code does this modification and then, when we, you know this, the sectional program can run on GPUs right. Our bytecode is that it's like very, very simple: it's culbreth a packet filter, it's designed for porting into any architecture you want, so we can actually report it to spear five. So we can execute all these programs on Jake and then on the pager who's beta.

A

First and DW yep.

B

Go fetch, it go run all the stuff in parallel, get them out of negations back and the only thing or validating is that if the data isn't modified, that it was done by this program right and if there's tokens that have been around that tokens going out where it moved. You know that this program didn't spend any external tokens and that that's a little somewhat soaked into the same.

B

A

There's a bucket here I mean maybe it's okay, yeah, the actual thing that pays for CP and the storage delivery. So if you're going to be few, for example, moving some other like yeah.

B

That's just data modification at state. That's a big! We don't really know what that is, but because actively, every transaction. In fact that we specifies all the memory regions it needs to access, we can schedule yeah.

D

And what might you just compile for that? Michael.

B

It's alluvium backends, so anything from a yeah. The backend is kind of weak. It's a little weak doesn't support like all the features yet, but that's just.

A

It's just work.

A

Wow, so if you have anybody, know the M character GPU, why we still write CUDA, no.

B

C

Use as LVM yeah.

B

Problem is amount, so we have a bytecode that is very stable right. This is a very stable byte gonna be Unchained forever yeah and we can pour it into whatever architecture happens, to be the fastest one. It may be for some use cases that spear 5 is the fastest thing, but for others it's not like x86. So we want to really keep it in this very simple to manage intermediate state. So.

A

Right now, if you get all this programs on GPU is oh, maybe sounds them may be way more efficient, don't seek on.

B

Ssds fashion, right now, it's all CP. This alright also give it. Yes, you guys want to some to spear five.

B

You can use the engine in your flock shape under chin yeah, so.

A

B

Awesome about like using a GPU for contract execution is, if you have a spike like cryptokeys yeah right.

B

The same program code can execute many many many different transactions at the same time on the GPU, because you have this like massive Carol massive amounts of planes for the same code. So if you had a cryptic Hades like spike, you actually don't use a different thread per transaction use. One thread that goes over this.

A

Enormous league-wide wait.

B

A

So that is like the really awesome thing: it's like I, don't you to write it just schedule space? Doesn't data access right, not.

B

What CPU threads? Because you have to use like vectorize code so like something like a bx mm-hmm.

C

B

The evenings by planets, like really really limited for a massive parallelization compared to a GPU one.

B

So in theory right, the cost of executed, like a more popular program, should actually be lower for the network. Pretty casual yeah, like you, don't need to page, then right and you're effectively. Just reusing of these lands will see. Is that these.

A

Other practice have to be measured.

B

Yeah, it looks like it'll work in theory, but.

A

This is done when, like honey in your case kind of no second increment check so like no.

B

Way so you have a bunch of transactions, they're all for crypto keys, right, yeah, there's a one contract that sticks up to get a scotch other than a meter behind. Okay, so I collect some transactions, execute them. You need a man like I. Don't need to do it anymore!

B

So, but if that's like the defense, the popular transaction, that's being thrilled Internet we're like they're all the time you can leave this program in the cash right, because you know that, like hey everybody's flooding in that work with these crypto kitties transactions right cool, those are like levels of optimization that I don't know, but there's fun fun to worry about. So.

D

Let's finish with a few non technical questions way, and when is the man that.

B

The hardest questions, so our goal is to PFC effectively can enter January and then made that will feature and I think by maybe more legal side than the US awesome.

D

Cool okay master I'm here yeah thanks a lot for questions, yeah cool.

A

Cool yeah feel free to ask questions on YouTube yeah yeah.