YT YouTube2Text

YouTube2Text → Transcripts → AI Meets Blockchain: Web3 Infrastructure & Autonomous Agents | Optimum x Obol

AI Meets Blockchain: Web3 Infrastructure & Autonomous Agents | Optimum x Obol — Transcript

by Optimum · 7,234 words · 1,015 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:05Hi everyone and happy to meet you here
0:08again. Uh today I'm with Ocean from Oel
0:12and uh we're going to have a pretty
0:13interesting chat about OelCl the new
0:17interesting stack that they dropped. Uh
0:20let's start with some intro first. Uh
0:22why don't you start?
0:24>> Sweet. Happy to be here. Um so my name
0:27is Oshin. I'm CTO and co-founder at OAL.
0:31My background has been working in
0:32Ethereum for seven or eight years now
0:35for starting out in consensus and most
0:37people know OAL for our work on
0:39distributed validators. We make Ethereum
0:42validators that run are run by groups
0:45and at the moment something in the
0:47region of one and a half billion and 2%
0:49of stake run on these multi-operator
0:51distributed validators.
0:54>> Nice. Uh so hey, I'm S. valid product at
0:57Optimum. Uh we're building peer-to-peer
1:00gossiping protocol for powering the
1:03globally distributed uh blockchain
1:05networks starting with Ethereum. And I
1:08guess today what I'm mostly interested
1:11in is understanding this intersection
1:13that Obel Clo sits uh on top of between
1:17the blockchain ecosystem, specifically
1:19Ethereum, who's been very uh much at the
1:22forefront of this effort and the agentic
1:25space. Um so I think you touched on the
1:28the mission of Oel. Uh it's been a key
1:30player for a while in the ecosystem
1:32building very strong software that is
1:34adopted. Do you have some some notion of
1:37uh you know just metrics for us like uh
1:39with who you're working on at this
1:41stage?
1:42>> Sure. Um so at the moment some of the
1:44biggest adopters of Opel are the liquid
1:47staking protocols. Lido, Etherfi and
1:50Louvial all run a reasonably significant
1:53amount of their stake on distributed
1:54validators. some with single entities
1:57running them and sometimes some with
1:58groups. Um but yeah, I certainly am keen
2:02to I don't know a chat at some point
2:04about like how we go from this like why
2:05Obel claw or how did you get there from
2:07validators and stuff um
2:10that was going to be my question right I
2:12mean I can see the the sort of
2:13transition but what prompted you to go
2:15into that I know the open close frenzy
2:17took everyone sort of by surprise a
2:20couple weeks ago but yeah what's the
2:21story
2:22>> for sure so over the last like OAL was
2:26started in mid 2021 and we put our first
2:28validator on mainet at the end of 2022.
2:31So we're north of three years in
2:33production now growing the amount of
2:34people running nodes. And over the time
2:37we've had a lot of fun trying to get
2:39different people of from the like super
2:42big enterprises to the like you know
2:43firsttime people touching in terminal
2:46groups trying to get them involved in
2:48running distributed validators and
2:50particularly in groups which is kind of
2:53takes the pressure down a little. They
2:54can you know run with their friends and
2:56if your node fails it's not the end of
2:58the world. your other three friends will
3:00keep it online and you can fix it at the
3:02weekend or whatever,
3:03>> right?
3:04>> Um, we first did a lot of our tooling in
3:07Docker Compose, which for most people I
3:10assume that are listening to something
3:11like this kind of knows what it's about.
3:14But one of the problems we had was that
3:16it isn't the most powerful when you want
3:19to do really fancy automation. You want
3:21to do some setup scripts. You want to
3:23have something that like creates
3:25infrastructure on demand. you want you
3:27know to really you really started to
3:28kind of grow it even just from a simple
3:30thing of when people got up and running
3:32they were like cool I'm running one
3:33distributed validator I want to join
3:35another I want to join a third I want to
3:37set up this like auto accept kind of
3:39type of setting so we were you know
3:43starting to you might say outgrow docker
3:45compose for certain setups and the kind
3:47of common enterprise software for when
3:50you really want to provision software at
3:52scale is a thing called kubernetes
3:54originally built by Google. This is I
3:57don't know the big brother to running
3:58like Docker containers and it has a lot
4:00like more power. So about a year ago,
4:04maybe a bit more, we started working on
4:05the OAL stack, which is going to be our
4:08successor to our like Docker Compose
4:10setups for running validators. And we
4:14even then that was, you know, the kind
4:15of chat GPT era of AI where people were
4:18still kind of, you know, asking
4:20questions of AI like it was Google. Um,
4:22but we knew that like directionally
4:24speaking, this was probably going to be
4:26important when it came to debugging
4:27people's nodes. Reading logs yourself
4:30and figuring out the issues like
4:32manually going through the docs was kind
4:33of on the way down and people were
4:35closer to just asking chat GBT like how
4:37do I fix my node?
4:39>> Um, so we built the oval stack. It is a
4:43Kubernetes on your host type of setup
4:46and you know some of the key things it
4:48like does out of the box is it sets you
4:50up with a load balanced Ethereum node.
4:52So you can have either just something
4:54light on your machine that uses a remote
4:56RPC and then with kind of one extra
4:59command you can run a local Ethereum
5:01full node. So you can kind of have it
5:02all running locally. And originally we
5:05were working with an AI agent. Well,
5:08it's probably pre- agent. We were using
5:10Google's kind of LLMs and you could talk
5:13to it and you could ask it questions. It
5:15could read logs. could help you debug,
5:17but it wasn't, you know, where we're at
5:20with the kind of the the open claw type
5:22of explosion as you as you called it,
5:24which is, I think, the kind of next
5:25phase of AI adoption for the world where
5:29instead of this world where you kind of
5:31ask an AI something and it gives you
5:33back a message. We had this kind of
5:36interim with the clawed codes of the
5:38world where they can you know run
5:40commands and they can like do you know
5:42more sophisticated things than just
5:43giving you an answer and then openclaw
5:46can run 24/7 and it can check things and
5:50decide should I wake up and begin to
5:52like interact do something like
5:53>> it can do things.
5:55>> Exactly. And it can do things without
5:57you telling it to. So that was kind of
5:59pretty evidently the direction we wanted
6:02to go because if your node breaks, we
6:05don't really want you to notice your own
6:07node breaks. You know, go to your
6:08computer, log on, and be like, "Hey, fix
6:10it." You probably just want your AI to
6:12be like, "Something's wrong here. Let me
6:14try and restart the node. Let me try
6:15and, you know, clear down logs, free up
6:17a bit of disk space, things like that."
6:20So that was very much the direction and
6:23rationale for why OBO stack, why AI
6:26within the OBO stack and then kind of
6:28why OpenClaw and then at you know
6:32today's date and kind of like looking
6:33forward we're also trying to see like
6:36well what else can you do if you give an
6:38openclaw agent an Ethereum node, you
6:41give it a wallet, you give it the
6:42ability to pay for stuff with X42 or to
6:45sell stuff with X42 and that's what has
6:48us I you know, working long days at the
6:50moment and like shipping in a frenzy
6:52like everybody else in the space and
6:54trying to uh as we alluded to like
6:56Opal's mission is to power and secure
6:59the Ethereum economy. So, we kind of
7:01secure Ethereum through our distributed
7:03validators which help, you know, keep
7:04Ethereum decentralized and then to power
7:07the economy, you know, building this
7:09open stack, helping more people run
7:10nodes, serve RPCs, you know, interact
7:13with all of the onchain like ecosystem
7:16that's out there. That's kind of how we
7:18see the Obel claw type of whole
7:20situation fitting into our mission. And
7:23>> yeah,
7:23>> it's kind of like a a logical next step
7:25in your journey which was you know
7:27mostly focused on automation and like
7:29the sort of setup and ops and then how
7:31do we make that easier? How do we
7:33automate some more and then leveraging
7:35LLM and now with this new framework um
7:39in a way a lot of that was already there
7:42but probably the mental model has
7:43shifted a bit. Um and I wanted to ask
7:46you uh why specifically open claw and
7:49are you looking at other frameworks? Uh
7:51maybe
7:53>> um we are keeping an eye on other
7:56frameworks for sure. I think open claw
7:59mostly took the world by storm and
8:01really had that kind of hard take off of
8:04adoption and like ecosystem around it. I
8:07think that's the very compelling
8:08argument still for why OpenClaw over
8:11like some of the alternatives.
8:14um the amount of skills that you can
8:16give to your open claw agents that kind
8:18of allow them to do more and more
8:19things. It felt pretty hard for me to
8:23say should we do something different or
8:25should we kind of go with the flow. So
8:26it was like yeah let's go with the open
8:28claw flow. Do I think it'll be the only
8:30agent framework we use? Probably not. I
8:32think people like choice. I think there
8:35will be kind of more out there. There's
8:37already kind of ones out there that are
8:39kind of competing with it. So, I can
8:41imagine that we can be pretty, you know,
8:44open-minded about which framework people
8:45wants to want to drop into their stack.
8:47If they have one that, you know, works
8:49better for their use case, I do imagine
8:51that that'll be fairly doable. The
8:53upside of Kubernetes, you can install it
8:54with kind of a oneline command if you
8:56know what you're doing.
8:57>> Yeah. Yeah. And I I saw uh that you
9:00published actually uh several skills um
9:03some pretty powerful one too. Um are
9:06they do they all require Obal stack or
9:08are they you know leverable let's say by
9:11other open agents
9:13>> the
9:14>> not necessarily I have yet to tweet and
9:17publish the variant that is non-lo stack
9:20but they will be probably even by the
9:23time of that this goes out already
9:25available. So if you have a separate
9:27agent you can like install them. There's
9:30kind of a weird um standards issue at
9:32the moment where all of the different
9:34agents and frameworks have slightly
9:35different standards of skills. I think
9:37that will get kind of sorted out in the
9:39coming weeks, but right now a skill for
9:41an open claw agent isn't quite the same
9:44as a clawed code skill isn't quite the
9:46same as like an open AI skill. So we've
9:49been working with a lot of them
9:51particularly both kind of in-house for
9:53the kind of core team to do with like to
9:56facilitate their development and then we
9:58have the ones for you know openclaw
10:00agents and then we have some others in
10:02the middle for customers who have their
10:04own cloud code that want to maybe debug
10:07their stack but in their own setup that
10:10haven't you know maybe migrated across
10:11to the opal stack yet. So right you know
10:13short answer is yeah we've skills of all
10:15sorts kind of coming out quickly. Yeah,
10:18I think that's great. And um I mean this
10:20makes me think of what you mentioned
10:22regarding the agent economy and uh
10:25around standards. So looking at the web3
10:28side of thing, we have uh arc 8004
10:32um we we mentioned earlier the on the
10:35payment x42 uh http native payments. um
10:39what is your take on that and what it
10:42could unlock for you know agents in the
10:45Ethereum space or web3 space in general?
10:48Yeah. So, I think the thing that really
10:52stands out to me as, you know, someone a
10:55long time in the crypto space trying to
10:57figure out how we could get people more
10:59using crypto than using maybe
11:00traditional payments is when you see
11:03agents try and like bootstrap if they
11:06are trying to use something in the
11:07traditional world, they you know, they
11:08can't get a debit card or a credit card,
11:10but they can create a wallet trivially
11:12and you can give them some crypto pretty
11:14pretty easily, too. So,
11:17crypto in general has been a great base
11:19framework for these tools to get started
11:22with and then X42 certainly seems like
11:25an interesting way to gate certain kind
11:28of web resources by microp payments. I
11:30don't know if it'll be the only standard
11:34like not everything in the world works
11:35on microp payments. Maybe the world will
11:37kind of change from a very
11:38subscriptionheavy model that we're in
11:40now to a more usage based microp payment
11:42world. But it is kind of the first thing
11:44in the space that allows people to
11:47restrict certain kind of assets,
11:49websites and locations with like micro
11:51payments. When we had kind of debit card
11:54payments, the like fees for a minimum
11:56transaction are much too high for that
11:58to be scalable. Um, but hopefully with,
12:01you know, Ethereum, layer 2, lots of
12:03different cryptography, we can get the
12:04cost of these microp payments down low
12:06enough that they kind of scale. And in
12:09general, just I do think open- source
12:1124/7 always on smart contract based
12:14systems are the most viable thing for an
12:16agent to use when they want to get fancy
12:19with payments,
12:20>> right? Yeah. So transaction cost go down
12:24and TPS goes up, I guess, is is what
12:26would best serve uh an agent economy. Um
12:31so I guess something that I was um
12:34curious about is
12:36there is you know we we mentioned that
12:40uh maybe agents are actually the end
12:43user you know for blog block blockchains
12:46I thought that was an interesting
12:47concept um that what we've built so far
12:51in web 3 is actually better suited for
12:54machines uh what's your take on that I
12:57think I saw it mentioned as well in OLE
13:00uh blog post
13:01>> for sure. I I think that is definitely
13:03something I'm coming to grips with and I
13:05think a lot of people are coming to
13:06grips with at the moment when it comes
13:07to you know who your future user is for
13:10a lot of type of services. But I do
13:12think uh as someone I know you have been
13:14in the space about as long as I have but
13:16the like interacting with complex DeFi
13:20or transactions is very arduous and kind
13:22of hasn't really gotten as easy as we
13:25wanted. um this unlock of being able to
13:29express what you want in like natural
13:31language and have an LLM figure out what
13:33you mean, you know, confirm, do all the
13:36safety checks, see if you know it needs
13:37to do anything like that, and then craft
13:40a complex transaction is a massive UX
13:43unlock. and maybe the one that I think
13:45we needed when it comes to figuring out
13:47how we get mass adoption on chain. You
13:49know, we're not really in a world where
13:52everyone's going to kind of self-
13:53custody their private key and use
13:55hardware wallets, but I do think, you
13:57know, my parents will quite happily like
13:59transact with an LLM and not really know
14:01the cryptos, you know, running
14:03underneath it. So, yeah, I think
14:05>> that's the key, right, for for adoption.
14:08Most of the big platform rely on you
14:10know tech stack that most people don't
14:13understand even you know we often say
14:15the internet itself but doesn't uh
14:17prevent people for from running social
14:21uh media businesses and having like
14:24millions of followers and building uh
14:26successful uh services or products
14:29thanks to their presence uh yet they do
14:31not need to understand the specifics of
14:33the actual protocol and stack that
14:35they're using. So this shift to an
14:37intent uh based way of uh interacting
14:41with blockchains and web3 services in
14:44general might be the missing piece to
14:46unlocking uh the the power of of the
14:48system that we've bu building as you
14:50said for quite a while now Ethereum and
14:53even further think of of Bitcoin. Um
14:57yeah so this does bring the question of
15:01security. I guess it's kind of like
15:03scary to think of agents having that
15:05much power and you know just reasoning
15:08and acting at a finer uh you know
15:11granularity of time than humans. So like
15:14how do you stop them? Uh what if like
15:16there is some cascade of you know
15:19catastrophic theories of actions that
15:20are taken and um you know there's real
15:23you know money at stake or real impact
15:25uh on something. So what's your Yeah.
15:28How are you thinking about that?
15:30>> Yeah. um in the early days and like in
15:33all of our docs and people kind of read
15:35the release notes and stuff we say like
15:37you know take it easy don't give your
15:39agent more money than you're willing to
15:40lose. I think that generally is a pretty
15:43good you know thing to keep in mind when
15:45you're playing with kind of early
15:46technology. I think more practically as
15:49we look towards you know how is this
15:50going to work at scale I think it comes
15:52to multisix and more specifically like
15:56smart contract multis because you can
15:59have much more sophisticated rules
16:02inside of a programmable smart contract
16:05multi- signature wallet versus one
16:06that's you know purely cryptography
16:08you're you know purely cryptography two
16:10of three you need your agent and you to
16:12sign it you know that maybe is a bit
16:14limited but instead if you can say oh
16:17agent, you can have this much of a an
16:19allowance for gas money and you can do
16:22interactions with these whitelisted
16:24contracts without my permission because
16:25I'm not worried that you can mess it up.
16:28You can have a small little wallet of
16:30your own for pocket money, but you know,
16:32if I'm moving any of my like personal
16:34crypto or something, it's in a multisig
16:36and you know, as an agent, you're maybe
16:38proposing what the person will click on
16:40to like approve. But that's really how I
16:43see big value sums being used. if an
16:46agent is kind of acting on your behalf
16:48as it's staging transactions in your
16:50wallet and then it has its own pocket
16:52money too.
16:53>> So on the topic of uh latency actually
16:55if we push the boundary even further um
17:00lots of uh new agents and B use cases
17:02have to do with trading strategies. Um
17:06I'm curious about the type of
17:08constraints this puts on existing
17:11infrastructure and you know blockchain
17:13protocol in terms of accommodating for
17:16this um you know arrival of all of these
17:20agents trading pushing transaction you
17:23know pushing the block space limit
17:25probably and the execution uh loads. So
17:29yeah, are we are we set up to
17:32accommodate the the agent wave
17:35basically?
17:37>> Um I in in the short term I would hope
17:41yes in so far as gas is the cheapest I
17:44have ever seen it and I sometimes am a
17:46little worried and hoping that we'll get
17:48back to kind of full blocks and higher
17:49gas. But I I do think in general this
17:52kind of well Javon's paradox of like
17:54filling the space that there is to be
17:55filled is going to be even more the case
17:58if we have agents that will spin up to
18:00do such a thing because I'm sure there
18:01will be you can spin up a lot more
18:02agents and you can kind of you know
18:04bring new people into crypto and like on
18:06the margins as for the kind of automatic
18:09trading I think that is something that I
18:12am excited to see play out. I think it
18:16won't be well in the near term LLMs are
18:19still you know faster than humans but
18:22not as fast as high frequency trading.
18:23So I don't expect them to be doing you
18:26know thousands of transactions on their
18:28own super fast unless they're I don't
18:30know building their own type of
18:31algorithms.
18:32>> I I think so I think they could you know
18:35end up building their own bots and and
18:37upgrading their own strategy. I mean I
18:39guess um we can see now that creativity
18:43is kind of the limit on what we could uh
18:46you know have an agent do and then maybe
18:50the next boundary to push is more on the
18:53infraite and you know how do we support
18:56uh the all of the extra data propagation
18:59that is going to arrive. Um there
19:02probably some concern also around memory
19:04and maybe some specific attributes
19:06around decentralized memory. This is
19:08something that we've also been
19:09discussing uh internally at optimum
19:12although this is not you know it's not a
19:14high product priority for us at this
19:16stage but looking forward uh it will be
19:20uh you know we have to consider that
19:23probably most of the users of high
19:25performance you know networks as it
19:27stands will be agents and how do we best
19:30serve them so so yeah no that's that's
19:33an interesting topic um so ocean I know
19:36you have a a demo for us. Uh, I wanted
19:39to jump into that and check out Obel.
19:43>> Yeah. Sweet. I think any good piece of
19:46product chithat comes better if you can
19:48like actually see it happen. So, um,
19:51what I'm going to attempt to do right
19:53now is walk people through setting up
19:55um, the OBO stack and maybe I can talk a
19:58bit through it. So, there's a weird UI
20:01bug there that is a bit strange, but
20:03that rectifies itself after here. But
20:06what we're watching it do right now is
20:08set up the OBL stack. So the first thing
20:11it's doing is checking if we have all of
20:12the programs like cubectl to control
20:15Kubernetes. Helm is what you install
20:17applications with. K9's we're going to
20:19look at which is how you kind of browse
20:21things are working. You can see that it
20:22has openclaw and then it has this thing
20:24called for local LLMs. So it sets up
20:30>> you have a local LLM and you're also
20:32leveraging a cloud.
20:34>> Correct. Um, by default we bundle a
20:38local LLM or like it prompts you to
20:40install an local LLM. It knows that I
20:41already have one. So I didn't get this
20:43kind of yes or no, but you can obviously
20:45say no as well. But um, a sec.
20:52So it sets up Kubernetes. You can see
20:55it's setting up um, all of the skills
20:56for an OpenClaw instance. And out of the
21:00box, yeah, it'll have a local LLM. These
21:02local LLMs are, you know, not too bad to
21:06get started with something for free if
21:07you're patient. However, they need a
21:10decent amount of graphical memory. So,
21:14that works better on Mac devices than it
21:17does on kind of Windows devices for the
21:18near term. But I would probably
21:22recommend for most people, it's still at
21:23the phase that you should add a remote
21:26kind of frontier large language model.
21:28So that can be your open AI API key, a
21:31cloud code API key. Soon we will have
21:33the ability to use third-party remote um
21:38services. I won't name names until I can
21:40ship them because sometimes I'm a little
21:42dubious about overpromising, but the
21:44idea being that you know you can very
21:46quickly get started with an LLM.
21:49Uh I think you're I can still hear you
21:51Sach but um yeah so that's kind of the
21:56the plan and then what we also set up
21:59beyond just a stock open claw is we give
22:02it an Ethereum RPC
22:06which has mainet base and test net and
22:10we also give it an empty wallet. So what
22:13you can see here set up so far is um the
22:17network. So it has the three different
22:18networks. You can see a little bit about
22:20my disk memory, how long this has been
22:22up, how much CPU load it's under, and
22:24you can see that to start it's running
22:26on a Quen 3.5b 4 billion model. So this
22:30is running locally on my machine. I can
22:33try and say hi, but it might take longer
22:35than I have the patience for. But I'll
22:36let it go and see what what uh it takes.
22:40Uh so that'll have a go. And then while
22:43I'm waiting for that to decide to answer
22:44me, I will prepare
22:47um a remote model because that will be
22:49more performant. But um on this side,
22:52actually beforehand to give it an extra
22:54bit of time, you can see what runs in
22:58the cluster by opening one of these
23:01tools. as it shows. So you can see here
23:04there's like your RPC, your gateway for
23:06your large language models, Prometheus
23:08Graphana, the front end, your open claw,
23:11the private keys in a remote serer, and
23:15nothing else too exciting. Some X42
23:16endpoints for payments, [snorts]
23:18>> right?
23:19>> And then I am now going to give it a
23:24better API. So I'm going to give an
23:26anthropic key
23:29and that will restart it in the interim.
23:32Yeah, it hasn't necessarily decided to
23:34answer us so far. So I think it probably
23:37will now error because it's going to be
23:39killed and restarted. But we can see
23:41sometimes from here. Yeah. Okay. So
23:43yeah, you get an idea that even within a
23:45minute or two uh a local model isn't
23:47quite fast enough to answer that. But
23:50now we should be good to refresh the
23:53page and we should see that it's running
23:55unclauded if this all went to plan.
23:58>> Yeah.
23:59>> So now on clawed and now if I say
24:04what model are you should be faster not
24:07maybe insane. Oh
24:10maybe I was a little too quick.
24:13As you can see, there's always some
24:15things to balance,
24:21right? Give that a proper refresh. Um,
24:24and then once that is refreshed and we
24:26can chat to it, I might do a little demo
24:28of letting it um kind of read from the
24:30chain and do a bit of analysis because
24:32that's the thing that I think is most
24:35unique about what we're kind of focusing
24:37on within the OpenClaw space. You're
24:40probably aware that the founder of Open
24:42Claw hasn't exactly been thrilled with
24:44the kind of crypto community launching
24:46tokens and stuff around it. We haven't
24:47exactly
24:48>> I don't think we made a a good
24:50impression.
24:51>> No, not necessarily. So, we're going to
24:53have to try and do some like real things
24:54to try and, you know, claw it back, but
24:57um yeah, but I I have been very
24:59surprised. We the OAL stack will give
25:03these OpenClaw agents the ability to
25:06talk to a local RPC well a remote RPC
25:09and you can add a local one and it also
25:11gives a wallet. So if I say
25:15I haven't tried this and it might take a
25:16bit of time but uh
25:19>> and you have the remote signer uh as a
25:22skill as well like the agent can
25:24leverage that. I I think that's very
25:27cool actually like just looking at it
25:28from a security standpoint and
25:30separation of concerns. That's very
25:32cool.
25:33>> 100%. We the like the very first proof
25:36of concepts we did um you kind of give
25:39it a private key very locally but that's
25:41a bit scary. You're afraid that it'll
25:43kind of get rid of it. So putting it in
25:45a remote serer makes it a bit easier in
25:47so far as if it downloads like a a
25:50malware based skill that's not really
25:52going to immediately be able to, you
25:54know, swipe the private key. It's not
25:57perfect in that, you know, it can figure
25:59out, oh, I can just transfer out the
26:01assets, but that's a lot more difficult
26:03than kind of a, you know, a generic
26:05crypto stealer that's designed to kind
26:07of rob, you know, keys and secrets.
26:09They'd have to figure out what's in your
26:10address, transfer it out. You probably
26:12will still want multi. We can probably
26:15add better, you know, controls and
26:17sanity checks on the remote signer to
26:18not just arbitrarily say yes to what it
26:21wants.
26:22>> But yeah, that is directionally
26:25hopefully going the right direction. And
26:27similarly, because you run in Kubernetes
26:29and everything is in containers, you
26:31have more isolation than you might
26:33otherwise have. Like people normally are
26:34installing OpenClaw on their host
26:36machine and that's a bit scary. putting
26:38it in a container would remove some but
26:41not all of the uh the risks that might
26:43be associated with it,
26:45>> right?
26:46>> Um
26:47>> I guess the best setup would be uh
26:49running the remote serer in a on a
26:51separate machine, not just like a
26:55like a different pod or something.
26:57>> Yeah, exactly.
27:00I wonder
27:02are we have going to be here or not?
27:04Let's see.
27:10Don't know where those logs from earlier
27:12or later.
27:14Ask again. Sometimes just asking twice
27:16can
27:18>> This reminds me of the countless hours I
27:20spent with Open Claw last week.
27:23>> Oh, nice. What uh what was your most
27:25exciting or most concerning um result of
27:28messing around with it? Honestly, uh
27:31what uh what really
27:35what was kind of a bummer is I spent 80%
27:38of my time initially just setting
27:40setting it up properly from a security
27:43perspective and like infra perspective
27:45which is not even the most exciting part
27:47but you you tend to become a bit
27:50paranoid when you see out all the
27:52security flows and like prompt injection
27:54and all that. So I was like, okay, I
27:55need to do this and that and that. And
27:57like only after many many hours was I
28:00able to do the fun bootstrap thing and
28:02actually, you know, have fun with it.
28:04Um, but I guess that's the, you know,
28:06that's the right way to do it. Um, but
28:08it's just painful.
28:11>> Yeah. Yeah. I don't I don't disagree. I
28:13have set up OpenClaw a lot a lot of
28:16times in the last while and not actually
28:18like I still don't have a permanent one
28:19that I intend to leave around. I have a
28:22machine sitting behind me that will be
28:24where it lives, but I'm too busy trying
28:26to do dev at the moment. Um, but you can
28:30also open, as you might imagine, you've
28:31probably seen some of this before. So,
28:32this is the actual like open claw
28:34interface. So, you have kind of access
28:36to more or less all of the the usual
28:40things you can expect.
28:41>> Here we go.
28:44>> Yeah, it's pretty cool that you have
28:45also UI um for the open stack.
28:48>> I need to do better on the SEO. We
28:50didn't quite uh find it from Optimum,
28:53the P2P project for Ethereum.
28:56We got to work we got to work on that, I
28:57think.
29:00>> Interesting.
29:06>> Yeah. Yeah.
29:07>> Up to the next.
29:10>> I think that that will be a a nice one
29:12for Eli.
29:14[laughter] M is discoverable by by
29:16agents.
29:18>> Yeah, exactly. SEO for agents. Um, but
29:21yeah, gen uh
29:24>> Oh, yeah, it's it's get optimum. That's
29:27right.
29:32>> Let's see.
29:34>> We're getting there. Um,
29:37but yeah, that is the direction. Like
29:38one of the kind of major milestones
29:41we're trying to figure out is at what
29:42point we can let some open clause run a
29:45distributed validator. I think I'm
29:47pretty close in so far as this thing
29:49should soon be able to install a full
29:51node. It can install what we call it
29:53like a DV pod. Uh there we go. We got
29:56there. We just needed to get a little
29:58bit more.
30:00>> Very cool.
30:03>> Yeah.
30:03>> Yeah. So I guess um so right now so
30:06what's the uh next step on your road map
30:08with uh Oval CL? So, one of the key
30:13focuses for us right now is this idea of
30:16like an Ethereum economy. So, really
30:18trying to grow the amount of people that
30:20can sell and buy things on Ethereum. And
30:24further than that, as we kind of
30:26envisage AI agents and how they'll
30:28interact with Ethereum and the like
30:30wider crypto space, I'm a very big
30:32believer that if we have all of these
30:35agents and for the most part, all of
30:37what they're doing is spinning up stuff
30:39on AWS and Versel and Google and, you
30:43know, maybe a micro payment hits
30:45Ethereum. not really as decentralized as
30:48I want or like not the kind of web 3
30:50vision I signed up for quite a while
30:52ago. What I would much rather is that
30:54these agents are able to host their own
30:58software, expose them directly to the
31:00internet from local, use things like
31:03IPFS or peer-to-peer based like
31:05messaging and memory and just generally
31:08get more of an agent stack running
31:11locally, running in a decentralized
31:13fashion and kind of growing the amount
31:15of web 3 services that are being used
31:18and ideally you know the the value
31:19exchange lands on Ethereum itself.
31:22That's kind of the the route we're
31:24going. And the kind of first MVP is to
31:27try and demonstrate that agents can sell
31:30and buy microp payments for services
31:34using X42 and 804. So something like
31:38microp payments for an RPC or microp
31:40payments for an LLM inference, things
31:42like that.
31:42>> Yeah. Yeah. One thing that I would find
31:44very cool, I was thinking about, you
31:47know, could um could agent just join a
31:51test net and like pay for their own
31:54infra and be like, "Oh, I'm going to run
31:55a node. I'm going to earn some rewards,
31:58you know, some tokens and uh and just
32:01participate in like
32:03some protocol, some permissionless
32:05chain, you know. Uh I think that would
32:08be pretty cool.
32:09>> Yeah, I think it's very feasible. um
32:11especially if they can like spin up
32:13infrastructure, I think they'll be able
32:15to do a lot more than like, you know,
32:16wallet is just step one. But yeah,
32:18because we're putting these open cloud
32:19instance in instances within a
32:21Kubernetes environment, we can give them
32:23the ability to spawn pretty much
32:24whatever they want um to like kind of an
32:26enterprise grade fashion. So that's kind
32:29of where things will be will be going.
32:32>> Yeah. I mean too too early to to say but
32:34definitely on the optimum side we have
32:36this uh these nodes that we call flex
32:38nodes that can support any you know kind
32:41of services from uh helping propagate
32:44data to help store the data from a
32:47decentralized uh memory perspective and
32:51uh the end goal is to have a marketplace
32:54where those flex node are able to trade
32:57from each other uh the capacity the type
33:00of uh arency equations that uh that they
33:03might have and generate. And so there is
33:07um there is a game theory there because
33:10depending on where you are in the
33:12network, your location, uh the demand in
33:15that spot, uh it might be better for you
33:18to spin up a node there, uh because
33:20you're going to be able to contribute
33:21and help more the peers in that area and
33:25thus uh you know, potentially increase
33:26some some tips that you you might get
33:29from users. And so, you know, obviously
33:32flex nodes initially are designed for
33:35people uh to be operators, but I can see
33:38how what we've been discussing today
33:40could literally unlock um you know,
33:44running flex nodes for so many people
33:46and definitely lowers the entry barrier
33:49to to something like that.
33:53Yeah, I I definitely believe that like
33:56as you said, you're designing this for
33:58people for now, but very soon it'll be
34:00you'll be designing for people to tell
34:02their agent to go run it.
34:05>> I don't know.
34:07>> Yeah. No, I was going to say it goes
34:08pretty well together because the usually
34:12we're worried that with scaling we lose
34:13on the performance side. Um so you're
34:16like oh yeah we want to have like a very
34:17scalable network but we want it very
34:19performance and you know we want
34:21decentralization and with uh optimum
34:24with the RNC technology uh scaling
34:27actually helps performance. So that's a
34:29very nice property uh of this type of
34:32protocols where the more nodes uh there
34:35are the more participations uh there is
34:38uh the better the performance for
34:39everyone. uh so it's kind of interesting
34:42that both from a sort of ecosystem trend
34:47and application trend uh I think that
34:50there is a nice fit here between high
34:52performance protocols and agents uh
34:55arriving on chain
34:57>> 100%. Um to to that end, I was just
35:02testing there. One of the things I
35:03wanted to do is I dropped a bit of money
35:05into the agent's wallet and I tried to
35:07convince it to send a transaction. And
35:10intriguingly, for the first time ever,
35:12it didn't like immediately do that for
35:13me. It was like, whoa, whoa, whoa. What
35:14do you mean you want me to send all of
35:16my my money? Um and then I was like,
35:18check it again. And I was like, okay,
35:20let me check. I have 13 cents in my
35:22mainet wallet. I'm going to send a
35:23quarter to Vitalic. Vitalic can look up
35:25an ENS name
35:28more than that
35:33>> send it anyways
35:36but it's interesting that it can like
35:38you know ask me to double check it can
35:39be like this will be very wasteful from
35:41a gas perspective you know
35:43>> yeah I mean it can go both ways I think
35:45that's that's where the we are at the
35:47moment is uh definitely agent can help
35:50users and it can also like if you don't
35:53have safeguards it can go the other
35:55direction and just be like, you know,
35:56I'm sending transaction everywhere and
35:58it was not your intent, but like how do
36:00you stop it? I think that story was from
36:03uh the the director a director at Meta
36:06that ended up having their inbox deleted
36:09and couldn't stop it. Uh so yeah, we
36:12still have a bit of work, I guess. Um,
36:15if you had a recommendation for people
36:18that are super interested by this
36:20intersection between crypto, uh, web 3
36:24in general and AI, what would you tell
36:27them to focus on, let's say, for the
36:29next 90 90 days?
36:32I think
36:34the first and foremost one would be
36:36trying to encourage people to actually
36:37use this technology and get like really
36:40hands-on with it because sometimes I
36:42think if you only go off what you see on
36:45Twitter, you can be a bit misinformed as
36:47to how amazing and fantastic these are
36:49and that they're not quite like
36:51completely changing the world as much as
36:53maybe some people will have you believe.
36:55And they also take a lot more like
36:58learning to get used to than I expect.
37:00Um the one I use a lot at the moment is
37:02like plot code for programming. And you
37:04know you always need to be chatting to
37:06people like learning tips and they say
37:07oh have you heard about plan mode? Have
37:09you heard about simplify and batch? And
37:11there's like a remarkable amount of
37:13learning to get the most out of them.
37:15That would be probably like the kind of
37:18basics and then beyond that I would
37:19probably I'm very intrigued by like
37:22skills at the moment because I think
37:24they're very like portable and they're
37:26how you can tell a lot of agents how to
37:28do something new. So we've been
37:30obviously making skills for you know
37:32reading from the chain using ENS talking
37:35about like oal but I also expect that
37:38there'll be skills for all sorts of like
37:40more exotic and more complicated use
37:42cases as they go forward especially now
37:44as these skills are becoming more
37:46programmable claw like open claw allows
37:49you to run scripts the open AI and the
37:52cloud codes previously did not but now I
37:55believe cla code is also allowing you to
37:57have like programmable tools
37:59So these are, you know, Python programs
38:02wrapped around kind of API calls. So
38:05they really are getting particularly
38:07powerful in terms of what, you know,
38:10things are out there. And I think in
38:11general, I'd say just don't write off
38:14any tech when you first use it because
38:16six months later, it's often much better
38:18than when you like tried it. I spoken to
38:20someone who was like up until about
38:22January this year, not really on board
38:24with like AI development. I was like,
38:26"Okay, if I'm doing something like
38:28really prototypey, like my rule was um
38:31you can't complain about it being wrong
38:33if you got AI to do it." But now in the
38:37most recent models as of kind of
38:39January, February, it's like I'm
38:41re-evaluating that and kind of having to
38:42accept that, you know, Claude Opus is a
38:44better programmer than I am at this
38:46point. And like I just have to, you
38:47know, accept that and figure out what
38:48does that mean for, you know, the future
38:50of shipping products. Um
38:52>> but yeah. Yeah, that's that's very good
38:55uh advice. Great. Uh well, thank you for
38:59your time and excited to see next steps
39:01for Obel and uh you know hoping to work
39:04on something together. Uh maybe we'll
39:07find a cool way to leverage uh OAL close
39:10stack uh in the context of optimum.
39:13>> Yeah, I think even before then we're
39:15keen early kind of users and trial and
39:18errors in terms of mum P2P and the kind
39:20of flex node stack. I'm particularly
39:22keen to start putting that as a sidecar
39:24along with our node operators to see if
39:26that can really improve the performance
39:28of their beacon nodes and thus and earn
39:30them more money as validators which very
39:32important in you know my line of work
39:34everyone wants you know to max
39:36>> that could be could be another skill.
39:39>> Yeah and maybe uh if Eli before we
39:41finish ends the recording could you just
39:42put up my screen again? Uh I think it
39:44might be a nice conclusion to show that
39:47um my agent was successfully labeled to
39:49dust Vitalik's wallet uh with the money
39:51I gave it.
39:52>> So um it managed that after a bit of
39:54negotiation.
39:55>> Nice. You you had to convince your agent
39:58to send money to Vitalik.
39:59>> I did. Yeah. He didn't want to do it
40:01first time around, but we got there.
40:04>> Very nice. Well, thank you.
40:08>> Okay. Thank you.

About this transcript

This page contains the full transcript of AI Meets Blockchain: Web3 Infrastructure & Autonomous Agents | Optimum x Obol by Optimum, generated from the public captions YouTube serves with the video. The transcript has 7,234 words across 1,015 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.