DSS 2026 / Day 1 / Amber Hall
What a libp2p Gossipsub Crash Taught Us About AI Security and DeFi Infrastructure
Manu Sheel Gupta
Source: YouTube automatic caption track · Exported 2026-11-16. Timestamps link to the source video. Cue end times were not supplied.
Transcript
0:01Good afternoon, everyone. I'm amazed to
0:03see the kind of work everyone is doing
0:06in security before the conference.
0:09It's one of the responsibilities when
0:10you are running an afternoon session to
0:12kind of figure out ways through which we
0:14can get everyone energetic. And I was
0:16actually wondering that over the course
0:18of last set of hours,
0:20I think the everyone who was speaking
0:22over here, they were like really
0:23energetic. The kind of
0:26thoughts and the kind of discussions we
0:27had, it actually gave me the idea that
0:30sure, Binance attack has happened, but
0:32that's not the end of the world. I think
0:34if you all come together, we can ensure
0:36that we can have a better DeFi security
0:39and a protocol system. So my talk is
0:41essentially about the same. I'm Manu
0:43Gupta from Li. P to P,
0:46one of the maintainers and the technical
0:47lead. And it is all focused on
0:50preventive security measure. Preventive
0:52security measure means a place where we
0:54have prevented some kind of Binance hack
0:57not to happen. So I am taking that into
0:59consideration.
1:01My talk is on gossip sub crash and the
1:03two very key developments which I'll try
1:05and cover over the course of talk.
1:07One is on DeFi AI infrastructure and
1:10second on AI security, which all of you,
1:11my wonderful friends, you have been
1:13working so well.
1:15Before we get into some of these
1:16glitches, I'll try and see if
1:18things work out well.
1:20Just one second.
1:24Yeah, okay, fantastic. So a very quick
1:26101 on the infrastructure level I wish
1:28to share. All of us have been discussing
1:31about how we can ensure safeguard of
1:33security measures at the smart contract
1:35level and at the application level.
1:37Fantastic. I mean, that's great. But I
1:39also wish to share with you, there is a
1:40very interesting surface beneath every
1:43application layer in a decentralized
1:45finance protocol. And that is
1:47essentially what we call as the
1:49infrastructure level. And infrastructure
1:51level
1:52is one of those things where you do not
1:54have to attack. You do not have to
1:56figure out keys to figure out where the
1:58problem is. You do not have to go for
2:00All you have to do basically is ensure
2:02that the network stops talking.
2:04And if the network stops talking in
2:06distributed systems, we have this very
2:08wonderful concept like there are three
2:09attacks which I wish to share with you.
2:11One is network partition, second is
2:13Cible, third is Eclipse. If any one of
2:16these attacks they happen, what we are
2:17basically saying is
2:19the entire assets that we have in DeFi,
2:22they will not be hacked, you'll not be
2:24moving across wallets. Basically the
2:26network will stop working. And that
2:28essentially is one of the things which
2:30I'm worried about when I look into some
2:32of these recent happenings at the
2:35Binance attack case as well as some of
2:38the previous ones. So what I'll be
2:39trying to unfold is probably a
2:41collective discussion and a
2:43collaborative method. I think that's
2:45what we should try and do during
2:46afternoon sessions. We'll try and kind
2:49of figure out what we can do better at
2:51the infra layer and also at the
2:53application layer using some of the
2:54existing innovations that have happened
2:57in AI and without AI, which means using
2:59fuzzing as was been mentioned previously
3:01by one of our friends. I would also kind
3:03of share some of the key things we need
3:06to understand at the
3:08network level where we are talking about
3:10very core methods of publish and
3:12subscribe and also of peer
3:14communication. So very quick 101 at the
3:17infra level what we'll be discussing in
3:19the course of talk one would be peer
3:21communication. Peer communication via
3:23gossip sub, flood sub and grid topology.
3:26Second we'll be discussing peer routing.
3:28So peer routing is via Academlia DHT.
3:31Third we'll discuss about some of the
3:32transport methods which we are using in
3:34Ethereum. The most notably is quick
3:36which is now now heading towards
3:38becoming an ITF standard. We'll also
3:40discuss about TCP, our connection
3:42oriented system, some of the connection
3:44less systems like UDP, web socket
3:46secure. We'll discuss on WebRTC direct
3:49which definitely Chromium deprecated,
3:51but we have got a WebRTC direct V2. And
3:54I would love to have some of you please
3:55contribute at the specs of lib P2B to
3:57try and see how we can ensure a better
4:00participation in that. And then we'll
4:01discuss how quick which is one of the
4:04key methods which is used by all the
4:06consensus clients of Ethereum. How we
4:09can ensure a better security using
4:11encrypted client hello and ECH ITF
4:14standard which is going to be discussed
4:16two weeks from now at ITS San Francisco
4:18and I would love to meet some of you
4:20over there and definitely looking
4:23forward to catching up on that.
4:25So first wish to share with you about
4:28one of the
4:29key learnings of D5 security management
4:31and infra level. What we see basically
4:33is
4:34one of the basic problems in the way we
4:37are writing code and the way a lot of AI
4:40work and agentic work is happening is
4:42that sometimes the training of data is
4:45happening from what I'll call it as not
4:47a sane data which means that agents
4:50sometimes they're trying to do under
4:52fitting and over fitting which is the
4:53same problem of
4:55like what we see in machine learning
4:57they're trying to have that same set of
4:59problems in infra level security
5:01management which means if I look in the
5:03last three months
5:05the number of security issues that have
5:07come to me at the infra level that is
5:09equivalent to the total number of
5:11security issues that I've received over
5:12my career working at the infra level. So
5:15that is the number that I'm seeing in
5:17terms of the participation of security
5:20issues. Now the point is what are these
5:22security issues? I'll first mention that
5:24and second I'll mention why many of
5:27those issues are not relevant and why
5:29many of those issues are like me too
5:31issues and how we a maintainer how a
5:33developer who is working at the infra
5:35level they can ensure that they look for
5:37those things which means like a in a
5:39funnel operation they look only for
5:41those issues where it's a make or break
5:43scenario. So we'll try and work out the
5:45optics for the same. So the first thing
5:47essentially is like if you look through
5:50this diagram what I have essentially
5:52explained is that if your nodes are out
5:54of sync and if there is a panic in one
5:56of the nodes. So we have a distributed
5:58network.
5:59Just a very quick question. You are all
6:02familiar with validator nodes and
6:04consensus nodes and how execution and
6:06consensus clients work. I mean, if there
6:08is any problem, we'll definitely discuss
6:09on that. But I'm
6:11assuming one particular part while I'm
6:13discussing this diagram. So we have a
6:15distributed system. We have the
6:16consensus nodes. They're all talking to
6:18each other and what happens is there is
6:21a panic that is getting triggered in one
6:23of the nodes. Now why this panic would
6:25happen? This panic would happen by one
6:27of the very basic things, integer
6:28overflow, which was explained in the
6:30last to last talk very nicely in the
6:32symbolic execution talk also. So that is
6:35one of the key things which causes a
6:37panic. And once that panic happens,
6:39people like as a long-lived network or
6:42as a fault-tolerant network, we would
6:43expect that Ethereum would kind of come
6:46back to its original state very quickly.
6:48Well, that is what we are trying to do
6:50and that's why I'm saying that my talk
6:52is a preventive security measure. But if
6:54that does not happen, if there is a
6:56quantum attack like what Justin very
6:58recently mentioned, he's a good friend,
7:00he shared about let's get into bunkers
7:02mode. So just to ensure that we don't
7:04get in even if you're in a bunkers mode,
7:06we do the security well at the at the
7:09bunker level. I do wish to share what
7:11really happens is p2p state sync stalls,
7:14the Oracle feeds halt, and liquidation
7:17bots fail. Essentially, what we are
7:18talking about is DeFi security coming
7:20under serious threat. Now it's not like
7:23I'm trying to make anyone of us. Like
7:25I'm so happy to see the energy over here
7:27and the kind of discussions and the kind
7:29of cases we have discussed that I'm not
7:31trying to share why this example that
7:34this would happen in Ethereum. What I'm
7:36trying to share is that these are the
7:37things that happens when you have an
7:40integer overflow and that integer
7:42overflow causes a trouble in the overall
7:44consensus client. So if a consensus
7:46client gets broken, then essentially
7:48what we are seeing is a DeFi problem at
7:51scale. Now how do you ensure that that
7:53does not happen is what we'll discuss.
7:56So we'll take a specific case which is a
7:58very popular one. It got a lot of
8:00attention at NIST at all the key places
8:04where we discuss about Ethereum security
8:08and it was
8:09being shared by everyone who was working
8:12in Ethereum and I was essentially the
8:14official spokesperson for doing all this
8:17discussion around CBE. So I wish to
8:20share first of all that this was one
8:22issue which we faced in Lighthouse.
8:23Lighthouse is like 27% of the Ethereum
8:26or 30% of the Ethereum consensus client
8:28market. It goes up to 43 now. So in the
8:31last year you can say that in consensus
8:32clients it's like 43%. It uses Rust
8:35libp2p. We also have Prism client which
8:37uses Go implementation of libp2p and
8:39then we have some of the others like
8:41Nimbus, Lodestar
8:43and some of the lean Ethereum consensus
8:45clients. So I'm taking one specific
8:47scenario and I'm sharing this because we
8:49were able to safeguard well. I mean
8:51generally where we are not able to
8:53safeguard well
8:54it's not easy for me to disclose because
8:57there's a lot of PR related discussions
8:59that happen but it's important to
9:01openly discuss some of those things.
9:03So I wish to share if you talk about
9:05this exploit what this essentially was
9:07that we have a swarm of autonomous
9:10agents and I'll also share how we can
9:12create our own swarm of autonomous
9:13agents. We basically had a very
9:16interesting integer overflow function
9:18and that was causing serious trouble.
9:20Now the point is if you would think like
9:22what
9:23would really happen if there is an
9:25integer overflow? So the funny part is
9:28it basically causes a node to panic and
9:31node to then go offline. So
9:33the trace is basically the following.
9:35You have an integer, you have the
9:37arithmetic operations. As was explained
9:39in the previous talk talk you have
9:40either the bitwise operations or you
9:42have the algebraic operations and you
9:44have a specific node which has not been
9:47defined or which has had a memory
9:50overflow or a memory specific issue. So
9:53if you have that kind of a specific
9:54problem what happens essentially is a
9:56fatal panic. A fatal panic means node
9:59out of sync and node out of sync
10:00basically means your DeFi protocol doing
10:03behaving or misbehaving for the lack of
10:06better word in a way that it should not
10:09behave.
10:10And
10:11sorry, and there is some trouble with
10:17yeah,
10:18just one second. I'll just try and see
10:20if we can move slides.
10:24Okay.
10:26Wow, sure. So
10:31yeah, thank you.
10:32Sorry, sorry for the hiccup. Yeah, wish
10:34to share that essentially what that
10:36basically means is that if a specific
10:38security operation has happened in a
10:40manner that it has caused the node to go
10:43out of sync, what it basically leads to
10:45is the entire DeFi protocol coming into
10:48serious problem. And this is what we
10:49have found with some of our friends
10:51working at Aave, Morpho, and some of the
10:54key DeFi organizations that some of the
10:56key mentioned points or the spotlight
10:58points that they do say is that if you
11:00have the nodes in sync and behaving
11:03correctly and you have to be right 100%
11:06of the time. I mean, you get one time
11:07you get incorrect and you have like some
11:09integers behaving inappropriately, you
11:12have the nodes going out of sync and you
11:14have the entire network going down. So
11:16here is a simple diagram I do wish to
11:17share and there are two parts to it. One
11:20is what is mentioned in this diagram.
11:23There's another thing which I also wish
11:24to explain. So this is essentially a
11:26simple workflow on how we ran a swarm of
11:29agents to do a code trace, to figure out
11:31the root cause analysis, to do
11:33coordination across different
11:35implementations. At Libp2p, we have 17
11:37implementations.
11:39Seven of those implementations are used
11:41in all the Ethereum consensus clients,
11:43which means if there is a problem at the
11:45Ethereum consensus client level, then we
11:47are to be blamed and we take full
11:48responsibility of the same. So as a
11:51person who is responsible for the
11:53consensus clients, I mean, we do follow
11:55some traces and this is one of that
11:57specific configuration.
11:59There is one thing which I've not
12:00mentioned over here which I would love
12:01to explain and I would love to have
12:03everyone work and that is how we have
12:05set up these autonomous agents.
12:08The configuration and the setup is very,
12:09very important and as we learned from
12:11the previous talk, it becomes very
12:12essential if we have to do the right
12:14kind of analysis. So first, you should
12:16be looking at agency directory service.
12:18You should be looking at how you can
12:20create a directory service of agents
12:23where they have a routing, discovery and
12:25a communication system. And I would
12:27strongly recommend going back to David
12:29Mazieres' time at Stanford University
12:31where he was working on Kademlia DHT.
12:33Use that as one of the key methods for
12:36intelligent routing. The second part
12:38which I do wish to share is very very
12:39significant and important is the
12:41messaging between agents. And there are
12:44two different ways of doing that. One is
12:46in many organizations we see that they
12:47have problems with non-HTTP protocols.
12:50So there I do wish to share if you're
12:52trying to do an A2A or MCP, try and do
12:54it over HTTP. And in many other cases
12:57you can create your own secure low
12:59latency messaging protocols. Sumanjit
13:01who just presented on one of the
13:03interesting talks at PII P2P four five
13:05talks earlier, he very nicely explained
13:08how we can have these nodes
13:09communicating and talking to each other
13:11and kind of sharing in real time. And
13:13the third thing which I do wish to share
13:15is identity interoperability of agents,
13:18which means if you want Ethereum nodes
13:20to not just function between your
13:21computer and one or two computers, but
13:23if you want to really do it as a
13:24collaborative action operation so that
13:27anyone who is trying to approach you
13:28from a nation state is not getting their
13:32kind of benefits. If we start operating
13:34together as a team, so the third part
13:36essentially is having identity
13:38interoperability of agent taxon. So
13:40there I would recommend try and use some
13:42of the IETF standards like OAuth which
13:44is OAuth 2.1 or you can use OpenID for
13:47doing identity interoperability. And the
13:49third fourth piece essentially is the
13:52observability part. Observability in
13:54agentic ecosystem is one of the key
13:56essence. I mean it's not just about
13:58tracing or OTel or telemetry logs, it's
14:00the entire action agent state capsule.
14:03And we would be presenting about it at
14:05IETF SF and would love to have your
14:08thoughts on the RFC that we have.
14:11So moving forward if I can, I think I
14:15Yes.
14:16Sorry, I just went one step back.
14:21Okay, fantastic. So we are moving
14:23forward. Why I'm saying that we should
14:26try and use AI security? Why I'm not
14:29talking just that we should look at
14:30traditional fuzzing? I think there are
14:31some very, excuse me,
14:33there are some very interesting benefits
14:35from the AI models, the way they have
14:38been approaching. I think one is that it
14:39kind of reduces the cognitive load on
14:42all of us as security researchers. I
14:44mean, if you are a researcher and we are
14:46working through massive loads of data,
14:48one of the key things we face is that
14:51the cognitive load is something we all
14:53wish to address. And triage, if you
14:55really believe my friend, is right now
14:57the product.
14:58I mean, the amount of as a maintainer,
15:00the number of security issues we are
15:03receiving in the last six months and I I
15:05I can definitely suspect in the last
15:07next six months it will be almost 10
15:09times more than what I have received in
15:10the last three previous six months. I do
15:13feel that
15:14one of the key things over here is how
15:17we set up the ecosystem, how we keep a
15:19balance between fuzzing that is
15:21traditional fuzzing and also the agentic
15:24forms and try and see where it works and
15:27where it does not work. And I do wish to
15:29share with you the inventor of fuzzy
15:30logic was interestingly my advisor
15:32friend Dr. Professor Zadeh at the
15:34University of California, Berkeley. And
15:36he said something very interesting and
15:38that is that
15:40fuzzy logic is very similar to is very
15:42similar to words with computing or
15:44computing with words.
15:46If you use the logic systems in a
15:49systematic manner, you do not need any
15:51AI model. And he used to openly say that
15:53and we have seen that it works very
15:54well. And in many cases where you have a
15:57logic task, I would still say go for
16:00traditional fuzzing.
16:02Agentic systems or agentic workflows are
16:04fantastic when you are working with
16:06differentiation, calculus, tensors,
16:08mathematics, isolated algebra. So there
16:11I do wish to share the AI systems are
16:14doing fantastically well. Why? Because
16:16they are been trained on certain things.
16:19And on those set of things they have the
16:22right kind of expected outcomes. In
16:24reference to logic paths I still feel
16:27the human mind and human brain works
16:29better and we have seen that across
16:32across number of exams.
16:36Yeah, just a second.
16:42Yeah, sorry for the hiccup. While this
16:45comes up, I hope it comes up. I'll
16:48definitely
16:50share
16:52it's just going the opposite.
16:59Yeah, I'll just move to the next set of
17:01slides very quickly.
17:03Sorry for the hiccup over here. There is
17:06something seriously. Okay. So here I'm
17:09discussing the security pipeline where
17:12we are doing coordination with the
17:14security team at Ethereum. So
17:16essentially what we have is we have
17:18Frederick and we have the consensus
17:19client team who are based out of London.
17:21And what we have essentially done is
17:23that since we work across different
17:25implementations and across different
17:28modules. So what we are trying to do is
17:30we are trying to ensure that it's a
17:32coordinated effort. And coordinated
17:33effort what it basically means that we
17:35use AI for doing verification, for doing
17:38auditing, for doing those set of things
17:41where we can reduce the overall
17:42cognitive load of our security
17:44researchers. But for many things we are
17:46still kind of embarking or following a
17:49very simple strategy that anyone who has
17:51a good interest or has a good experience
17:54in the deep tech understanding of infra,
17:56which basically encompasses your
17:58transport layer, your identity layer,
18:01your messaging layer, communication
18:03layer, routing system, secure
18:06communication channels, which is TLS
18:08noise and auto TLS. So someone who has
18:10deep interest on these things, we are
18:12basically sharing with them that okay,
18:14you kind of collaborate with us, work on
18:17a module by module level, try and see
18:19and understand from the existing CVEs
18:22which have already been issued for the
18:23last I would say seven to eight years.
18:25So we do have a good human knowledge of
18:27last seven to eight years and I would
18:29say the first four five years when there
18:30was no LLMs. I mean even knowledge at
18:33that particular time in terms of
18:34triaging, in terms of collaborating. So
18:37we have that eight to 10 years of
18:39knowledge very well developed in terms
18:41of how Ethereum security has been
18:43managed and how the DeFi security around
18:45Ethereum has been managed. So keeping
18:47that into consideration, keeping that
18:48into perspectives and building on the
18:50learning that we have so nicely
18:53I would say
18:54documented and also made it sure
18:58also ensure that it is available to
18:59everyone who wishes to join. We are
19:01saying to everyone and anyone who is
19:04interested in doing infra level
19:05security, join us, collaborate in some
19:08of these efforts, try and see and make
19:10it try and see on your system whether
19:13you can evolve your understanding of how
19:16things work at a distributed node level,
19:18how you can fix some of these issues and
19:20how you can ensure that peer-to-peer
19:22security really works. And this is one
19:25of the key learnings which I do wish to
19:27share that there were a lot of members
19:29who when I was having coffee outside,
19:32they mentioned that the work in AI
19:34security or AI research is now getting
19:38been done more by AI models and less by
19:40human beings. I personally believe it's
19:41quite the opposite.
19:43In
19:44in reference to the infra level
19:45security, we can do a much better job if
19:47we have some of these models. What I
19:50essentially wish to share in reference
19:51to this case was that we use AI agents
19:54to figure out a panicking of a node. Now
19:57node panicking is like one of the things
19:59which we can easily miss. Why? Because
20:02memory overflow and integer overflow are
20:03those cases where sometimes a human
20:06brain will say that okay, these are some
20:08of the calibrations that the original
20:10author would have done and I don't have
20:12to run those kind of code checks or code
20:14traces.
20:15And unless you are a symbolic
20:17researcher, I mean unless you have built
20:18a symbolic
20:20system where you are doing an if else
20:21construct and you're trying to really go
20:23at that level, doing an arithmetic level
20:25overflow research is harder. So, what we
20:27found is that if you are using an AI
20:30agent and you have a very basic integer
20:32overflow,
20:33you can have it fixed in one
20:34implementation that is rust and then you
20:37can have a very clear pathway through
20:39which you can share it across all the 17
20:42implementations. So, we have in total 17
20:44implementations and the good thing, the
20:47benefit of AI is that the security team
20:50remain as it is, which means if there
20:52are like 20 contributors and 20 security
20:55researchers in the overall libp2p
20:57project, the core ones I'm talking
20:58about, the ones who are being paid for
21:00doing
21:01the entire bug fixing, maintenance, code
21:04management. So, those members are
21:06basically coordinating faster
21:09and they are able to coordinate faster,
21:10they are able to ensure that the final
21:13code trace that is available, I mean
21:16they are able to kind of ensure that
21:17those exploits are not
21:20very specific to one implementation,
21:22they're also shared by other
21:23implementations. Why this becomes very
21:25important is that the like the attack
21:28will not come just from one specific
21:29consensus client that is a lighthouse
21:31consensus client. It can come from say
21:33like Teku which is an institutional
21:35level Ethereum client that uses Java
21:37libp2p. It can come from Nimbus which is
21:39using Nim libp2p, another very
21:41interesting language. It can come from
21:44Lodestar which is using Zig libp2p or JS
21:46libp2p. So, the point is that
21:49the first part is to fix that issue, but
21:51the second part is also ensuring that
21:53whether that fix has been propagated
21:55timely to the other implementations and
21:58to the other consensus clients because
22:00if one of those implementations have
22:02that integer overflow or has that
22:04specific issue, then this entire
22:06exercise of using an automat automated
22:09agentic workflow system, it does not
22:11reap those kind of benefits. If one of
22:13the consensus clients is in problem, so
22:15it's like one of your family members has
22:17a impact, then you definitely get
22:19impacted by that. So, it's the same set
22:21of things and I do wish to share the
22:24final uh piece
22:26over here.
22:27Sorry.
22:28We had a tough time today, but yeah,
22:31this is the last slide. I just wish to
22:34share that
22:35most important thing is that if we have
22:37the AI execution done properly, if we
22:40have the AI security workflow analysis
22:42done properly, then
22:43there couple of things to learn from
22:45this. One is build your triag pipeline
22:48so that it is not like let me accept AI
22:50or let me accept AI models, but it is
22:52like where I need AI.
22:54I mean, that is essentially the question
22:56I would like to share with everyone.
22:59It's not the question of I want AI to
23:01use security or to do better in
23:02security. It's like where I need it. And
23:05where I don't need it, I will use the
23:07regular human brain. I will use my
23:09friends and collaborators. I'll work
23:11with them to ensure that we get to a
23:14good conclusion to a specific issue. So
23:16these are like some of the learnings
23:18which I do wish to share. I do feel that
23:21many of the things which I shared in
23:22this discussion, you might be feeling
23:24that hey, this is very specific to how
23:27we work at infra level security for the
23:30DeFi
23:32for the DeFi project. Yes, definitely it
23:34is very much related to infra level
23:35security, but I also feel that it has a
23:38lot of implications to what we guys are
23:40doing at the wallet level and at the
23:42other different levels. That is
23:44essentially your application level, your
23:46DeFi protocols. They all are following
23:48the similar trend and some of the
23:49learnings we did from the infra layer is
23:51something which we have kind of
23:53propagated to the application layer. So
23:55in case if you are working and you are
23:57like a do it yourself person who likes
23:59to run nodes, Ethereum nodes, likes to
24:02run validators, I would definitely like
24:04to share two things with you. That's a
24:06friendly guest. Like that's one of my
24:08responsibilities. One is you please
24:10upgrade your gossip sub to 0.49.4 which
24:13is the latest version. And if you're
24:15using floods of fantastic, if you're
24:17using grid topology, I mean fantastic,
24:19you don't have to worry, but
24:21do the same. And second essentially is
24:24that you also prepare your triag
24:26pipeline for the AI security swarm. So
24:29essentially try and prepare your AI
24:32security swarm track pipeline. And yeah,
24:35if you have any questions,
24:36please let me know. And also one last
24:40request, if anyone of you is interested
24:41in applying for an AI grant or for a
24:44infra level grant, please do let me
24:46know. We do have some opportunities.
24:47Would love to discuss. Thank you.
24:50[applause]
24:57You can speak.
24:58>> I just want to understand that you
25:00mentioned that there are multiple
25:02like
25:04let me do in Java, let me do in
25:07>> Different implementation languages.
25:09Yeah, yeah.
25:10>> So
25:10I wanted to understand does it make
25:12sense to have a
25:14central document
25:15>> Mhm.
25:15>> which has the specs written out and that
25:18is useful for me to verify or
25:21symbolically
25:22figure out if the
25:24>> Absolutely.
25:25Beautiful question. I do wish to share
25:27we do have specs. It's called as let me
25:29to be specs. More than happy to share
25:30the link. And I do
25:32so that is that gives a very so that's a
25:34very interesting question. I mean we are
25:36working in a decentralized environment,
25:38but at the same time there are points of
25:39convergence. Let me to be spec is the
25:41point of convergence. That's where all
25:43of us we kind of collaborate. We discuss
25:46for every specific function how it has
25:48to be written or for a specific
25:49architecture how it has to be worked on.
25:51For any spec like for any spec proposal
25:54which is shared as an RFP or even if it
25:56is shared as a PR in our specs repo. So
25:59we have a very clear path of review
26:01cycle. So it's generally reviewed by all
26:04the maintainers of different
26:05implementations. We have a weekly spec
26:07meeting. Would love to have you join
26:09that spec meeting also. We are now
26:11having it more often than usual because
26:13of a lot of security issues coming our
26:16way. Many of them are minor issues, many
26:18of them are me too issues, but some of
26:19them are definitely high severity
26:21issues. And I do wish to share that yes,
26:24having a common spec is one of the ways
26:26through which we can align all these
26:28different implementations.
26:29Thank you.
26:31All good?
26:32>> Thank you.
26:33>> We are just running short of time. So
26:35we'll definitely catch up outside for
26:38coffee if anyone would love to catch up.
26:40Thank you.
Source and accuracy details
- Names, organizations, technical terms, code and numbers may be misrecognized.
- Source wording and disfluencies are preserved; no editorial corrections were applied.
- Export contains whole-second cue starts only; cue ends and subsecond timing are unavailable.
- A full caption-track export does not establish that every spoken word is represented.
- Automatic YouTube captions; not human-verified against the audio.
Source export SHA256: bdcfddc89f8a144ef7043e0f7a70370d2859cef91f4d73fcfd885feee782f735